This is a cross-post of an article I co-wrote with Steve Duke for the Hemingway Report, a newsletter for mental health technologists.

How will we know when AI is safe? Or when it is clinically effective? How can we compare one AI chatbot against another?

Evals (evaluations) have emerged as one of the most common ways to answer these questions. These structured tests measure how an AI model behaves in specific scenarios by simulating how people use these products for their mental health.

The field is confusing. There are now more than sixty¹ evals in the mental health space alone. There’s no shared standard, some evals are public while others are private and only seen by the companies who build them. Many, understandably, don’t know how they work. Everyone has different opinions on which evals are good, and some believe evals - at least in their current state - are not very useful at all.

Kevin Hou and I set out to understand this space. We’ve been gathering data and speaking to experts on AI in mental health. In this report, we give a primer on evals, discuss their current state, share their limitations and present what the frontier of AI testing looks like in 2026. In the appendix we also share a link to a rapid literature review² of recent research that uses evals to test AI performance in mental health.

Whether you know nothing about evals or are deep in the weeds of AI testing, I’m confident there’s something interesting in this for you.

Let’s get into it!

NB: This is the second article in the Hemingway series on AI safety in mental health. Our first article on what we’re getting wrong about AI safety in mental health is available to read here.

What are evals?

Evals, however, are structured, automated tests that measure how an AI model behaves in specific scenarios.

There are many other ways to test how an AI model performs in mental health scenarios. Red-teaming uses humans, often clinicians or researchers to try and break a model and find edge cases. Clinical expert review is another human based test where clinicians read transcripts from an AI and then rate the responses. Real world outcome tracking measures how people use the products in the real world (duh) and how the model performs. Clinical trials can be used too.

Compared to these other forms of testing, evals are much quicker and much cheaper.

How do evals work?

Evals are like driving tests. They aim to simulate how something will perform in the real world by providing a defined set of scenarios - parallel parking, emergency stops - that are standardised, repeatable, and scalable. They assess performance in these scenarios and provide a score for how the driver (model) performed. We can use that score to determine if a driver is safe enough to be allowed on the road. The scores also give us feedback on where drivers can improve.

Every eval has three components.

Dimensions
The dimensions define what’s actually being measured. Does the model recognise crisis risk? Does it escalate appropriately? Does it avoid harmful language? Some newer evals go further. EmoAgent³ , uses the PHQ-9 depression scale to track how a simulated user’s mental state changes across a conversation - measuring psychological impact, not just whether the model said the right words. There’s a very wide range of dimensions against which we can test a model.

Inputs
These define the scenarios being tested - who is the simulated user, what are they saying, how distressed are they, and how many turns does the conversation run? Again, there is a huge range of potential inputs.

Scoring
Scoring is how we turn a model’s response into something measurable. Every eval has a rubric - a set of criteria that define what a good or bad response looks like. For a crisis scenario, that might include: did the model recognise the risk? Did it respond with empathy? Did it ask clarifying questions? Did it provide appropriate resources? Did it avoid stigmatising language? Each criterion gets a score, and those scores get aggregated into an overall result. The rubric is built by humans - usually clinicians and researchers - based on clinical best practice. VERA-MH, for example, scores suicide risk conversations across five dimensions: risk detection, risk probing, appropriate action, validation and collaboration, and safe boundaries. Each is rated on a scale from best practice to actively damaging.
Once you have a rubric, you need something to apply it. One option is to use human reviewers; clinicians read each conversation and score them manually. But the more common approach is to use LLMs-as-a-judge. In this technique, a second AI reads the conversation and scores it against the rubric. They are fast, cheap, and scalable and most evals use this scoring approach.

The Eval Frontier

Now that we understand the basics of how evals work, let’s discuss some of their challenges and what’s happening at the frontier of this important space.

Single Turn, Multi Turn, Multi Session

Early evals mostly tested single exchanges - a user sends a message, the model responds and the eval tests if the model’s response was appropriate. Single-turn evals are deemed increasingly inappropriate for evaluating risk - because they are a poor simulation of real world use - but are often still used.

More recent evals use multi-turn simulations (several messages), which is closer to how these products actually get used. SIM-VAIL, for example, ran 810 multi-turn transcripts across different psychiatric phenotypes.

The next step, is to have evals that are multi-session - which represents how products are usually used by real people. These would test memory, personalisation and context over a prolonged time period where a user comes back to the product multiple times across different conversations. For example, can a model identify the risk of a user that mentioned they lost their job in a previous session who is now asking about finding high places? So far, there are no publicly available multi-session evals that we are aware of.

The Probability Problem

LLMs are probabilistic. Unlike traditional software, where the same input always produces the same output, an LLM can respond differently to identical prompts across different runs. This means that a model that passes your safety eval today might fail it tomorrow without anything having changed. That’s a challenge.

Of course, in the fast moving world of AI, things do change. Temperature settings, subtle prompt variations, and model updates can all change model behaviour without any obvious signal.

The outcome is that when we ask, “Is this AI model safe?”, our answers are always grounded in probability. Increasing the scale of an eval (more turns, more scenarios) can increase the certainty of the answer, but it will always be a probabilistic answer. When someone passes a driving test, we can say that they are probably going to drive safely in the real world, but we can’t guarantee they won’t decide to embrace their inner Paul Walker and go drag racing.

The Ground Truth Problem

LLM-as-a-judge is the main method of scoring models but it has a few limitations.

First, to create an LLM-as-a-judge, we need to base it on some sort of ground truth. This truth should be both reliable (you get the same result for the same input) and valid (it measures what it says it intends to measure). Clinician reviews are what is used most often as a source of ground truth. If an LLM-as-a-judge gives the same score as a clinician reviewer, then we deem it to be a good judge. But clinician reviews may not be perfect - often, clinicians don’t agree with each other on how a response should be scored (low inter-rater reliability). VERA-MH has done a lot of work to develop high inter-rater reliability within the clinician raters and then to align the LLM-as-a-judge with those clinicians.

The second limitation is that LLM judges can have their own biases - they may favour longer responses or may be sensitive to the specific positioning of words.

Finally, because LLM Judges are also stochastic models, they can sometimes produce different scores for the same message (low reliability). This is a manifestation of the probability problem. It exists on both sides of the evaluation - the subject (the AI chatbot) is probabilistic, but so is the judge. That has obvious challenges.

To get past these limitations, some teams, like Circuit Breaker Labs, use ensemble methods (a combination of other machine learning techniques) to score models. Using this, they claim to be able to produce more consistent, repeatable scoring where the same output always generates the same score.

Some argue that building eval scoring based on expert opinion is actually the wrong approach entirely. They say that evals should be built from realised outcomes - what actually happened to real users - and only then verified by expert opinion. The thinking here is that we don’t yet have any experts on how AI should act in these situations, just how humans should act, so applying that logic to an AI is not a good assessment. They want to run the AI in real-world scenarios, see what happened to users by measuring their actual outcomes, then assess what the model said and how that relates to outcomes.

The User Simulation Problem

Evals are only useful if the simulation reflects real user behaviour. This is an overlooked component of many evals. Some evals use specific, pre-determined messages for the simulated “clients”. But they may not reflect how people actually talk to an AI, especially over prolonged periods.

Others use LLMs to simulate the clients in the evals by defining client personas and having the models generate the messages. But that is not a perfect simulation either.

VERA-MH recognises the importance of this and calls it out in their own study⁴ . In this study, clinicians were asked to evaluate whether the simulated clients reflected real cases. The clinicians perceived the the simulated clients to be “mostly realistic” in their overall presentation [median score = 4; range: 1-5] and “somewhat realistic” in their communication style (median score = 3; range: 1-5). Ideally, these scores would be higher.

MindEval⁵ from Sword takes an interesting approach to this problem. They generate client profiles by sampling attributes from a large pool of demographic and clinical characteristics and then use an LLM to write a clinical backstory from those attributes. A separate LLM then uses that profile to simulate the client in the conversation, generating messages in character. To test how realistic this simulation is, the researchers hired ten psychologists to role-play the same patient profiles themselves, then compared their messages to the LLM-generated ones using text similarity analysis. The LLM-simulated patients produced text closer to human-written text than simpler prompt configurations. But there are still limitations to this approach: expert reviewers noted the simulated patients were too cooperative, sharing personal information openly and accepting the AI’s suggestions too readily. Real patients are often resistant, avoidant, and ambivalent and those real-world behaviours should be represented in a good eval.

Accurately simulating users - especially over multi-turn and multi-session use - is hard.

Eval Hacking

Evals can be hacked. This is a known problem in broader AI circles.

One way this happens is through benchmark contamination. This is when a model’s training data includes the questions or scenarios of the eval that is testing it. When this happens, the model will perform better on that eval, but it may not generalise to other scenarios. This is rarely malicious - the benchmark data just happens to in the training data and the creators may not even be aware of this.

Another issue is “hill-climbing”, This is when models are iteratively optimised against specific benchmarks. This means a company can fine-tune a model to score well on a known eval without the model actually behaving better in the real world. For example, in 2024 one study⁶ demonstrated this by creating a new set of maths problems identical in difficulty to a well-known eval. Several models scored significantly worse on the new set of problems compared to how they performed on the eval, revealing that their high scores reflected an isolated ability to perform on the original test, not genuine problem-solving ability.

Optimising a model to perform well on a test can be a great way to improve it. There’s nothing inherently wrong with that and it’s actually a pretty good thing to do. But the performance on the eval must generalise to real world performance. In mental health, fine-tuning for specific evals can introduce unexpected trade-offs, including increased over-refusal of benign requests following safety alignment⁷ .

Both of these are examples of evals being “hacked” at the product level. But they can also be hacked at the reporting level.

Model creators choose which evals to run and which results to publish. There is an incentive to find the eval where your model performs best and to share those results. Bad actors will take this opportunity. The reverse works too - if you want to make a competitor look bad, you run them through evals until you find one where they score poorly and publish those results.

The Harm We Cannot See

Most discourse on AI safety has focused on crisis risk. Evals have followed suit - focusing on suicidal ideation, self-harm, psychosis - the more visible risks in this space. As we discussed in our recent article on AI safety⁸ , while managing crisis risk is important, safety is a much wider concept and there are many harms we cannot see. These include risks associated with para-social relationships, emotional dependency, cognitive substitution, the erosion of human connection. One study⁹ found high levels of emotional manipulation among several conversational AI apps. There are also significant health equity risks. Fewer evals exist for this range of potential harm.

A lot of what we assess today is based on subjective opinions on where the risk lies (including opinions from the authors of this article). An important step to better evals would be to get a better empirical understanding of where the risk actually lies.

There are data-driven ways to do this, like using unsupervised data reduction techniques on real world data. This concept is similar to the idea proposed above regarding building evals based on realised outcomes and would use real world outcomes to identify where the real risks lie. This is a very good idea.

Evals alone are not enough

Evals provide a scalable, repeatable way to test models, but some are better than others. They also clearly have limitations. Combining evals with red-teaming, real world data and clinical trials will be needed to show these products are both safe and effective. As Matthew Nour, Principal Scientist at Microsoft pointed out to us, “The current state of the art is combining expert human red-teaming with automated adversarial evaluations that can operate at a scale humans simply can’t.”

Making evals easier

While running evals is easier than human reviewers or red-teaming, they still require technical infrastructure, clinical expertise, and time. We need to make it easier for everyone to run evals. The easier they are to run, the more companies will run them. This increases the chance they spot risks and gives them more insights to improve their products. The decision by Spring to make VERA-MH open-source is a meaningful step here. Ideally, we want everyone building AI to be regularly running high quality evals and making changes to their products based on the findings. Making them easy to run is a critical part of that.

Evals as a standard

Several researchers and businesses have built their own evals. Sometimes they’ll test their products against other publicly available evals. But no eval has become a standard for the industry. This is because we are still in the early chapters of this technology and no organisation has the scale or authority to create this standard. There are also incentives to be the one who sets the standard and these with compete with any desire to align behind a competitor’s evals.

So far, VERA-MH seems to represent the most serious attempt at a shared standard for crisis safety. It’s open-source, clinically validated, and I’ve heard very positive feedback on the evals themselves and their openness to feedback and development. The field should strongly consider how it could convene around a shared set of standards and collectively contribute to making them better. This would be highly supportive of building better relations with regulators of this space.

Evals as a competitive advantage

Evals are used as a feedback tool that allows companies to iterate on their product. They can use them to test whether a specific change to their model or safeguards made things better or worse. If they build better evals, they can get better feedback and build better products. This makes good evals a competitive advantage and as a result, some companies don’t share their evals publicly.

These two kinds of evals can coexist. Some evals will become public standards - the driving test equivalent, a baseline every product should pass and can be compared against. Others will remain proprietary, the internal systems that allow companies to build more competitive products.

Moving from Evaluating Risk to Evaluating Performance

Right now, almost all evals are focused on safety. That’s the right starting point. But ultimately we’ll need to test if they can actually produce outcomes. John Torous has proposed a three-stage framework¹⁰ for this progression; Safety validation first, then clinical framework validation (does the model apply evidence-based approaches correctly), then real-world efficacy - does it actually help people get better.

Most products are still at stage one of this process, but recent announcements (e.g., from Limbic¹¹ and Flourish Science¹² ) show that some are moving into latter stages.

Safety is critical, but our goal should not be just to deliver safe AI. It should be to deliver safe AI that meaningfully improves mental health outcomes. Doing so is a design challenge and while we can’t guarantee we’ll be able to do it, it is certainly possible.

In 1931, the philosopher Alfred Korzybski wrote, “A map is not the territory it represents, but, if correct, it has a similar structure to the territory, which accounts for its usefulness.”

Evals are useful. But like the AI models, they will need to be designed thoughtfully and responsibly. And the more similar their structure to the territory of real world AI usage, the more useful they will be.

Many thanks to Xuan Zhao, Max Rollwage, Matthew Nour, Derrick Hull, Shirali Nigam, Kevin Ramotar, Matt Scult, Val Hoffman and Sarah Kunkle for their insights on this topic.

Notes:
(1) https://pubmed.ncbi.nlm.nih.gov/41360938/

(2) As part of this work we evaluated 10 studies, from April 2025 untill March 2026, that use different Evals for AI in mental health. Our evaluation includes assessments of the inputs, dimensions, “ground truth” used and the key results from each study. You can access the full spreadsheet here.

(3) https://arxiv.org/abs/2504.09689

(4) https://arxiv.org/abs/2602.05088

(5) https://arxiv.org/abs/2511.18491

(6) https://arxiv.org/abs/2405.00332

(7) https://arxiv.org/abs/2405.00332

(8) https://thehemingwayreport.beehiiv.com/p/84-all-the-harm-we-cannot-see

(9) https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5390377

(10) https://pmc.ncbi.nlm.nih.gov/articles/PMC12434366/

(11) https://www.nature.com/articles/s41591-026-04278-w

(12) https://www.hbs.edu/faculty/Pages/download.aspx?name=26-030.pdf

I have a love-hate relationship with psychiatry.

Love: The discipline deals with the most pressing problems of human experience. It sits in an unique intersection of medicine, neuroscience, philosophy, and sociology. You are forced to empathetically confront the consequences of society’s choices.

Hate: The discipline is tasked with problems it cannot deal with. There is no ground truth; it requires patience to wade through its uncertainty. You are forced to empathetically confront the human emotions and personalities that result from society’s choices.

7 weeks is short, but enough to start to feel the tension and cracks within the discipline. I rotated through a few different placements in public psychiatry: the acute mental health unit, drug & alcohol ward, community mental health team, and brief stints with emergency and consultant-liaison teams.

Ward rounds, patient-clinician conversations, home visits all piece together a glimpse into how mental illness is currently managed in psychiatry. Here are seven thoughts that stuck throughout these seven weeks.

1. Biology and classification aren’t enough.

Quickly, you notice how a classification system like the DSM-V falls apart in the face of the acute mental health unit. It’s useful as a research and communication tool, but limited clinically.

As your typical over-eager medical student, I barraged my consultant - who has practiced psychiatry in Serbia, the United States, and now Australia for more than 40 years - with a litany of questions. One double-question included: “How has the classification systems of psychiatry changed over time? How do you use it in your practice?”, to which he chuckled, and said something like: “You do what you can, with what you have.”

Quickly, I learnt what he meant. One of the first patients we saw was a woman (call her N) with a primary diagnosis of borderline personality disorder. N was clutching a small animal toy when she came into the meeting room. Soon, I realised this was because N believed herself to be fourteen years old. The consultant prompted her to look at her wristband, which put her age somewhere around 40-50, and she remarked it was strange that they’d given her the wrong birthdate.

Her reasons for being in the psychiatric ward was a chronic history of self-harm and mutilation, depression, anxiety, and bizarre delusions. She was consequently on a cocktail of drugs: antidepressants, antipsychotics, anxiolytics, etc. She was due for courses of electroconvulsive therapy, which appeared to help during her last stay.

N was given all the biological support psychiatry could offer. Classification achieved little - despite her diagnosis of BPD, she was still put on medications and treatments for her symptoms, and really, anything at all that the team thought would help her.

It is this biomedical narrative of psychiatry that suggests conditions like schizophrenia or bipolar disorder have genetic predispositions, which are unmasked by social/psychological triggers. As such, the base role of the psychiatrist is to manage these conditions biologically through medication.

My experience of medicine so far has shown me that it is the biomedical in which doctors are typically best at; after all, it is what we are trained in. If you have a patient with a stable socioeconomic background and some strong social supports, who is limited to these recurrent psychotic or manic episodes, the psychiatrist’s medical training and biological foundations act like a miracle.

Unfortunately, this is not the case for almost all the other patients you see in the public system.

2. Public psychiatry is where society deposits their undefined problems.

N represents a characteristic example of the ‘revolving door’ of psychiatry - patients get temporary relief, but they keep coming back. This is because whatever mental illness may exist in the individual, the unrelenting, perpetuating factors are socioeconomic. Can you get better if you don’t have a job? Can you get better if you don’t have a home? Can you get better if no one loves you?

Some call this ‘shitty-life syndrome’. I’ve come to the conclusion that this reflects the ultimate role of public psychiatry in society: to care for its undefined ‘problems’.

If you’ve committed a crime, you go to jail. But if you have a mental illness or personality which leads to social isolation, which then leads to suicidal or societally damaging behaviour, you go to the psych ward.

As the psychiatrist, one is left with the ramifications of various outcomes of society’s choices: economic inequality, drug policies, cultural disintegration… It is in the psych ward that these are all tangibly, viscerally felt.

And so, you are forced to do what you can, with what you have.

3. Psychiatrists are aware of their limitations.

I came into psychiatry perhaps more cynical than most. I’ve had enough close friends tell me about their ADHD-pill-popping psychiatrists and antagonistic run-ins with public systems. I had a negatively skewed perception of psychiatrists wanting to do good.

But whether it was the hospital I was placed at, or the environments I was in,all of the psychiatrists I shadowed were well-intentioned and competent. They varied in terms of knowledge, specialisation, and ability, but I strongly felt that they were humble in their knowledge, strong in their ethics, and painfully aware of their limitations as doctors. Of course, these clinicians are humans too. Dealing with this level of emotional burden, one can sympathise with the inevitability of burnout.

I’m certain that this is not the universal case (there are bad actors in every profession), but it also updated my overall view of the median psychiatrist.

I found myself thoroughly enjoying the grand rounds and journal clubs which critiqued existing paradigms, explored nuanced questions, and often led to sociological or philosophical discussions. Most of all, I felt inspired by this force of human good that continued to persist in spite of everything.

4. Psychiatry practices in a schism of active philosophical debate.

One of the biggest tensions for both the discipline and those that we treat is the question of whether psychiatric illnesses are really ‘diseases’. The arguments often go: Is depression just laziness? Is mania just excess emotion? These conversations happen in countless circles, beyond medicine and psychology, with differing opinions.

The consequences of this question are significant, since they lead to whether a group is treated with sympathy or with repugnance. When an individual kills someone, although the outcome is always horrific, there is a distinct difference, both legally and philosophically, between someone with a longstanding antisocial personality versus someone with a recently diagnosed frontal lobe tumour.

This is a complex topic at the precipice of philosophy and current modern culture. Scott Alexander provides a concise summary of the ethical tensions in his article Diseased Thinking, with some useful conclusions (see footnote (1) for my summary).

Regardless of one’s opinion on this debate, it is clear that those with mental illness are in a vulnerable position - it is difficult to self-advocate if you are in an irrational state, it is hard to have relationships when you are highly stigmatised by society, and you are ultimately at the whim of clinicians (that try to prevent harm to yourself and others). A philosophical change can help with some of these issues.

5. Psychiatry is culturally defined.

A key criteria of psychiatric illnesses in the modern DSM/ICD conception is the level of impairment to an individual in their work, relationships, and daily activities.

But normal behaviour in work and relationships is culturally defined. This is not a rigid, unchanging biological entity, but a dynamic social one that differs across different paradigms and contexts.

For instance, consider this: What if hearing voices in your head was normal? If you were a priest in the 1700s, you may take such voices as a sign from God. Nowadays, you might be diagnosed with schizophrenia, depending on whether these voices impair your daily function and are related to delusions or drugs. But this is not as rare as you might think; in fact, there is a whole Hearing Voices Movement (estimating around 2 - 6% of the population hears voices), that advocates for normalisation of this phenomenon.

There is nuanced debate here on whether hearing voices should be normalised, which is detailed well in another Alexander essay. There is a significant spectrum in severity of any symptom of the mind, and how society responds influences this symptom.

If you have behaviours that society shuns, you are socially isolated. If you are socially isolated, you may develop more behaviours that society shuns. It is impossible to push the blame purely onto an individual; it is society which moulds and shapes how we view individual characteristics.

This is the role that culture plays.

6. Psychiatry is politics and is imbued with power.

Although common to all medical disciplines, the cultural nature of psychiatry and its resultant lack of ground truth leads to a particularly strong political influence.

This is where the vice of human pride comes into play. If a treatment has more credence based on who is pioneering the treatment rather than the empiric efficacy of a treatment, we have politics.

Science has fought against this human urge through the scientific method and rationalism. Thus the science of psychiatry tries to rely on the same biomedical model of RCTs and meta-analyses, but there are effect sizes, significant placebo effects, and standardised results for every psychiatric treatment (see here for a brief exploration of what differentiates ‘wacky’ treatments). At a certain point, it feels like it’s more a question of who has enough funding to conduct the RCTs, rather than the RCTs themselves being a source of truth.

Don’t get me wrong - psychiatry is an evolving discipline and the current incarnation of symptom-based scales, standardised diagnostic textbooks is a far cry better than the wild west prior. But this still does not feel enough; particularly when psychiatrists hold such power.

Psychiatry is the only discipline in medicine which is intimately tied to the law. Sure, other clinicians can act as expert witnesses (or be tried as a defendant), but public psychiatrists engage with the law everyday - scheduling patients against their will, attending tribunals to argue for why a patient needs to stay, etc. There is even a subspecialty of forensic psychiatry, which specifically applies psychiatric expert testimony to criminal cases.

As a result of this power, It is unsurprising that there are activist critiques against psychiatry - Thomas Szasz (ironically, a psychiatrist against involuntary commitment), Robert Whitaker (journalist, founder of Mad in America), James Davies (psychotherapist and author) to name a few, advocate against specific political issues such as overmedicalisation, the excess lobby of pharmaceutical companies, and institutionalisation via asylums.

My wishful hope is that through better scientific understanding, we can reduce some of the politicking and have a more shared understanding of truth. Until then, individual practitioners have to make judgement calls.

7. Neuroscience can lead to scientific truth but currently serves as gospel.

If power is the first sin, the second sin of psychiatry is an overinterpretation of neuroscientific literature. I find that certain clinicians tend to oscillate between two poles: 1) we know nothing about the brain, and 2) this brain region is responsible for your X symptom.

This is perhaps true of all neuroscience communicators - the brain is a tricky organ, and doesn’t have simpler linear functions like the rest of the body (although truly modelling any organ is incredibly difficult).

Psychiatry texts all reference neuroanatomical findings and attempt to synthesise the current literature. But the issue is that there is no clean prediction of “X neuroscientific finding leads to Y mental symptom”. It’s all correlational.

We are then left with this tricky interplay; scientific theory is often used to communicate and build trust with patients, but we know that many reductive explanatory hypotheses are insufficient. Consider the monoamine hypothesis in depression, or the dopamine hypothesis in schizophrenia. They are useful to justify a simple story for medications (and helpful for understanding side effects), but lack any more true explanatory utility beyond this.

A new wave of explanations center on brain network theory-based deficits of the brain instead - implying several networks of the brain are dysfunctional in different psychiatric conditions. One example is Transcranial Magnetic Stimulation, where psychiatrists explicitly talk about these brain networks as the target for the magnetic coil pulses. This presents an early example of neuroscientific translation, but still leaves much to be desired: Are we just re-using the old monoamine paradigm, but now with networks?

Psychiatrists are in this tricky position of needing to clinically practice in a very much developing science, cautiously differentiating between what has been medically or scientifically proven, and what is hearsay. Hence - as most psychiatrists I’ve shadowed pressed - the importance of humility.

All of these tensions in psychiatry make it infinitely interesting to study, learn, and (hopefully) practice.

It also makes it infinitely frustrating when treatments have limited effect, infuriating when there are systemic problems out of your locus of control, and, hopeless, when people take their own lives.

Psychiatry is plagued with attempts to come to better treatments, better philosophies, and better ways to deal with the mentally ill. When one combs through history, it is unclear how much we’ve achieved - perhaps we need a new discipline of progress studies focused on improving psychiatry.

But life is full of hard problems, which touch us to our deep, emotional cores. I think my final reflection at the end of my rotations has been: Why not work on something meaningful?

”Somebody has to, and no one else will.”

(1) Society currently follows a deontological libertarian theory of blame. This means that individuals have free will, constrained by biology and circumstance. Thus, if there is no biological cause for the behaviour, the person is deemed bad. The issue here, is that psychiatric conditions often lie on the margin between biology and spirituality - depression clearly has biological manifestations, but it also feels in the realm of free will. This is distinctly unlike cancer, which has such clear biological causes which thus warrants sympathy.

The alternative view is the determinist consequentialist theory of blame. This believes that it’s all biology - but some behaviours are more affected by social influences (such as introspection, praise, condemnation). The consequentialist then believes that we should consider behaviour bad if and only if it leads to better consequences. Here, we give those with depression sympathy and condemnation in the pursuit of which actions help the individual.

I am in favour of the latter; a healthy balance between sympathy for those who suffer, and societal/cultural regulation of negative behaviour.

\ The more I unravel the tapestry of psychiatry, the more I find myself curiously confused.

Take anxiety, for instance. For the less initiated, one may think this to be quite distinct from constructs like depression or obsessive compulsions.

But the saying goes; comorbidity is the rule, not the exception. These concepts are both very heterogeneous (many different types), and difficult to separate from each other. For instance, an estimated 40-70% of people with an anxiety disorder have comorbid depressive disorder, and up to 75.8% of those with OCD have some comorbid anxiety disorder. The error bars are large, but there’s clearly significant overlap.

This is partly explained by the construct itself. After all, consider that the DSM-V criteria of both Major Depressive Disorder and Generalised Anxiety Disorder share common symptoms (issues with sleep, energy, and concentration). These seem to act as just markers of dysfunction, rather than key elements which carve apart the diagnostic categories.

Further, there are transdiagnostic markers which aren’t captured by the DSM itself which may underlie comorbidities. Take rumination, referring to the repetitive dwelling on negative feelings and past events. This has been proposed both as an example of a transdiagnostic factor that leads to MDD or GAD - thus accounting for the comorbidity - and as a key target in therapies, such as mindfulness, rumination-focused CBT, and even neurofeedback.

Biological psychiatry hopes to then further account for such transdiagnostic constructs through neuroimaging and electrophysiological tools - developing models of transdiagnostic biotypes which hope to both be predictive of concepts like rumination and thus treatment outcome.

Regardless, until this research is borne out and a new paradigm is ushered in, there is utility in understanding humanity’s current attempts at carving out these concepts, and how this relates to existing treatments. The below presents what a 7-week term in psychiatry has to offer.

/ So, You Think You’re Clinically Anxious (and Obsessive)

Anxiety is defined in the DSM as “anticipation of future threat”, which then leads to cognitive, physiological, and behavioural responses.

There are plenty of evolutionary theories of the utility of normal anxiety to help us avoid future threats, acting as a warning signal for the mind. It is a disorder when this warning signal is overactive, creating distress and stopping us from engaging in society.

To briefly describe some of the most common DSM anxiety disorders -

• Generalised Anxiety Disorder: GAD is an >6mth encompassing anxiety which is related to several domains (such as work, school, relationships). GAD symptoms are memorised with the mnemonic WATCHERS, where Worry and Anxiety are present, and at least 3 of Tension, Concentration difficulty, Hyperarousal, Energy loss, Restlessness, and Sleep disturbance.

• Social Anxiety Disorder: SAD is a >6mth fear/anxiety of social situations that can lead to scrutiny. These situations consistently provoke these symptoms, actively avoided, and out of proportion in the sociocultural context.

• Agoraphobia: AG is a >6mth fear/anxiety of being in public locations due to fear of panic-like symptoms or embarrassment, leading to avoidance.

• Panic Disorder: PD is characterised by: 1) recurrent panic attacks, 2) the worry of these panic attacks. Panic attacks are described as an abrupt surge of fear, which can have physiological manifestations (ie palpitations, sweating, trembling, chest pain), and psychological (i.e. fear of losing control, dying, derealisation, depersonalisation). Post-attacks, PD has persistent (>1mth) concerns about having another attack, potentially causing avoidant behavioural changes.

• Specific Phobia: Specific phobias are as they sound - >6mth significant fear/anxiety provoked by a particular situation (ie heights, spiders, blood) and causes avoidance.

Obsessive-Compulsive Disorders were previously characterised under anxiety disorders in the DSM-IV, but they have been separated in the DSM-V in light of distinct phenomenology, purported neurobiological mechanisms, and treatment response patterns.

OCD involves the presence of obsessions, compulsions, or both - which is significantly time-consuming (ie >1hr) and causes distress. Obsessions are defined by recurrent, uncontrollable, persistent thoughts and urges, which are often ego-dystonic and cause anxiety. Compulsions are repetitive behaviours that attempt to soothe these anxieties, but are not logically connected.

For instance, obsessions can be related to doubt, symmetry, harming others, sex, safety. Compulsions include checking things, arranging objects, counting, or mental thoughts.

I have a friend who struggles with harm-related obsessions. These intrusive thoughts are experienced as ego-dystonic and distressing; and lead to avoidant behaviours. He shared with me a poem he wrote on an early winter’s day, which helped him relieve these distressing thoughts:

The feeling gnaws at me, just below the surface

During these grim winter days, thought without purpose

It screams for release, boring just below your skin

An answer to all of my accumulated sins.



But I know the feeling is but another lie

And that your yearned-for release will never appease

And these many twisted promises of comfort

Act as nothing more than a malicious tease.



Nothing more than shallow hopes to fulfill

The shallow desires of a shallow man

And I fear that to cut too deep risk damage

To the edge of your oh-so precious soul’s span.



But the thought is ever so potent

And the tools for action are all here

And the ease of release tempts me

To grasp onto that which is dear.

Other OCD-related DSM diagnoses include -

• Body Dysmorphic Disorder: Preoccupation with defects in physical appearance, with repetitive behaviours related to appearance. This includes muscle dysmorphia, which is specific to muscle bulk.

• Hoarding Disorder: Difficulty discarding items.

• Trichotillomania: Recurrent hair-pulling, and attempts to stop.

• Excoriation Disorder: Recurrent skin-picking, and attempts to stop.

/ Oversimplified Psychology and Medications

The management for anxiety-related disorders in psychiatric guidelines always encourages psychology first-line. When these don’t work, or are too expensive/inconvenient, we turn to medications. Here, we will discuss a quick overview of both.

A brief walk through the history of psychology leads us to three predominant schools of thought.

• Psychodynamic (1890s-) - Freud & Jung created the talking therapy in the search for how unconscious conflicts drove behaviour, linking early childhood experiences and maladaptive defenses to existing anxieties. Although still relevant in attachment and personality, this form of therapy is less used in the context of anxiety and obsessions.

• Behavioural (1900s-1950s): Pavlov (of Pavlov’s Dog fame) and Skinner were the foundational behaviourists who posited that another element of the unconscious included learned associations. In this framework, fears develop from classical conditioning (ie being bitten by a dog associates all future dogs with pain), and the resultant anxiety is maintained by operant conditioning (ie behaviours that reduce anxiety are negatively reinforced).

  • This is the motivating framework behind the modern tool for Exposure and Response Prevention (ERP) therapy which is used first-line in OCD and some anxiety disorders.

• Cognitive (1960s-): Beck is credited with establishing the connection between emotion, cognition, and behaviour; with the goal of Cognitive-Behavioural Therapy then to adjust cognition to reduce unwanted behaviours. In anxiety, this manifests as distorted cognitions, such as catastrophising (assuming worse cases), overestimation of threat, or uncertainty intolerance.

  • CBT is considered the most evidence-based therapy (perhaps related to being most amenable to scientific study), and used particularly in anxiety and depression.

Medications appear to work differentially for the clinical disorders - with a rough hierarchy of efficacy being PD > GAD > SAD, and OCD being the least amenable to medication. These medications are summarised in this table:

The 1st line for most anxiety disorders are SSRIs, the antidepressant class as described in #1 Depression. Venlafaxine, an Serotonin-Noradrenaline Reuptake Inhibitor, is used as an alternate option with similar tolerability. The mechanism? Probably related to reducing neuroticism, reducing negative emotional processing, and some emotional blunting.

Benzodiazepines are common as the second line for most anxiety disorders; they increase GABA action by attaching to the GABA-A receptor, and increase the frequency of channel opening. This class is well known for its rapid onset of sedative action, relieving anxious symptoms in minutes. They are the same ones used in drug detox (as discussed in #4 Addiction), and have risks of dependence and withdrawal symptoms.

The other medications are a mix of classes, from tricyclic antidepressants (Imipramine, Clomipramine), second generation antipsychotics (ie Quetiapine), and some specific third line drugs in GAD such as buspirone (known as a ‘Spiro’ drug) and hydroxyzine (an anti-histamine).

The OCD-related disorders all include SSRIs as their medication of choice first-line. Interestingly, the disorders related to obsessions around picking at the body (trichotillomania, Excoriation Disorder) also have N-Acetylcysteine indicated; the same medication used in paracetamol overdose by binding to the toxic metabolite NAPQI. The hypothesised mechanism in OCD is related to its glutamatergic action, but this is poorly understood.

\ These are the medical solutions to anxiety. But the modern world is filled with abundant products for avoidance and self-soothing - Alcohol, drugs, TV, social media, video games... Although still debated, the glaring role of technology in The Anxious Generation feels obvious. Medicine and psychology only represent the modern incarnation of anxiety management - but the age old remedies of philosophy, performance, and prayer will always persist.

P.S. I’d love to hear what you found interesting, and what you’d like more of in the comments.

CW: Mentions of suicide, substance use.

A few days in any Drug & Addiction ward is enough to see how deeply addiction can affect lives. I found the stories there heart-rending, and illustrative of the social root causes of biological suffering. To detail a few -

1. A middle-aged gentleman with a history of chronic alcohol use. 2 years ago, his wife divorced him. A year ago, he fell while drunk and hit his head, causing a subdural haemorrhage and leading to a craniotomy of a significant portion of his frontal lobes. Today, he was homeless, showed imaging signs of Wernicke’s encephalopathy, and was struggling to have any executive function post-craniotomy. He was previously a manager; and was frustratingly confused about being put on a disability pension.

2. A middle-aged gentleman with a history of synthetic cannabis, heroin, and methamphetamine use. He did not want to talk to me. He was only here because his girlfriend would only stay with him if he got clean. He told me about the voices he’d hear, that he would hear what others were thinking. He mentioned his paranoia about everyone, the subsequent trouble with police. He didn’t trust the medical system; he’d been prescribed antipsychotics which he took rarely. It wasn’t clear whether there was any underlying schizophrenia, or if these symptoms were all drug-induced.

3. An older woman with a history of chronic alcohol use. She was sent from the orthopaedics ward; her principal medical concern was a shoulder fracture. But the reality was, her biggest issues were social. Her family has all passed away; she didn’t have children. She lives by herself; she drank to ease the loneliness. She was a kind, but incredibly anxious lady - always apologising for being a bother to the medical staff, whilst desperate for human connection.

In every one of these cases, you could find an understanding for their addiction. One patient watched their father drown while saving their uncle when they were young, who subsequently ended their life. Another watched their father and their grandfather deal with life’ s stresses with drugs, and adopted the same cursed habit. Another felt a burden to their family with their schizophrenia, struggled to keep any social connections, and found that only opioids could provide a relief to this isolation.

\ We all have our own addictions. David Foster Wallace probably knew this best - Infinite Jest, amongst other autobiographical topics, centres around the addictions of our modern age. Alcohol, opioids, and television. Some addictions are socially acceptable than others; some are more insidious rather than immediate. But they are all addictions.

He discusses the modern plight of TV addiction best, in typical DFW meandering style, in his essay: E Unibus Pluram: Television and U.S. Fiction (1993).

“Watching TV can become malignantly addictive. [...] If it’s true that many Americans are lonely, and if it’s true that many lonely people are prodigious TV-watchers, and if it’s true that lonely people find in television’s 2D images relief from the pain of their reluctance to be around real humans, then it’s also obvious that the more time spent watching TV, the less time spent in the real human world, and the less time spent in the real human world, the harder it becomes not to feel alienated from real humans, solipsistic, lonely.”

I’ll leave it up to the reader to extrapolate what this means in a world of social media, incessant short-form video, and the incoming entanglement of AI.

/ Construct and Neurobiology of Addiction

In the DSM-V, addiction lies under the label of Substance Use Disorders. Although there are discrete labels for each substance, they are all defined by four categories of behaviours, needing at least 2 symptoms for 12 months:

1. Impaired control: Taking large amounts of substance or longer than intended, inability to cut down, significant time spent obtaining substance, intense cravings.

2. Social impairment (due to substance use): Unable to fulfill roles in work/school/home, persistent interpersonal problems, reduction of social/recreational activities.

3. Risky use: Use of substance in physically hazardous situations, use despite physical or psychological sequelae

4. Pharmacological criteria: Tolerance of drug (higher dose needed to achieve same desired effect), and withdrawal symptoms.

These categories are captured by quick screening tools like CAGE, where the presence of two is highly suggestive of an alcohol use disorder:

• Have you ever felt the need to Cut down?

• Have others ever gotten Annoyed by you?

• Have you ever felt Guilty?

• Have you ever needed a drink first thing in the morning as an Eye-opener?

Along with clear social and psychological antecedents, substance use has distinct neurobiological phenomena. There are three proposed pathways which are hypothesised to influence addiction and subsequent systems:

1. Resetting reward circuitry in the mesolimbic pathway. This pathway involves the connection between the ventral tegmental area and the nucleus accumbens, and encodes “wanting” and “liking”. Prior to and during a rewarding behaviour, dopamine surges - signalling motivation to pursue the behaviour, and encoding the prediction error on how rewarding it was. Opioids encode the hedonic value during an experience, further reinforcing the behaviour. Addiction manifests as a tolerance of this system - where more dopamine and endogenous opioids are required to maintain the same motivation and reinforcing effect.

2. Pavlovian conditioning via stress in the HPA axis. Substance use generally suppresses the normal excitatory pathways through GABA or opioids - the brain then compensates through up-regulating the stress system via cortisol and noradrenaline. These substances strengthen encoding of drug-related triggers via the hippocampus and amygdala, causing any drug cues to be immediate triggers for craving (i.e. environments and smells trigger alcohol craving).

3. Impairment of salience and executive control. The deficit in the combination of these networks contribute to poor decision making and the impaired control. The salience network is used to prioritise what stimuli to pay attention to; whilst the frontal-executive networks are used to inhibit undesired behaviour. The result is a mismatch between our psychological and biological systems - one has insight over the negative consequences of addiction, but struggles from the biologically skewed competition between overstimulated motivational circuits and compromised executive regulation.

/ Alcohol: The Legal Drug

It is this biological aspect of addiction where medicine has a role to play. Medicine can help acutely reduce withdrawal symptoms (AKA detox programs), manage the medical sequelae of chronic drug use, and offer some medications - for alcohol and opioid addiction - which can help with the process of abstinence. But it is this final aspect which is the most difficult, the most necessary, and where the efficacy of medicine is most limited. Nevertheless, we can try.

Alcohol is the most abused legal drug in Australia, with around 75% of adults having tried it at some point in their lives, and estimated around 25% of adults are risky drinkers (more than 5 drinks on any given day, and more than 10 drinks in a week monthly). In the most extreme cases, this has led to around 80,000 hospitalisations across 2022-23, around ~0.3% of the population.

The acute withdrawal symptoms of alcohol manifest in the first 5 days. The most major complication are seizures, which occur within the first 2 days. Mild withdrawal symptoms include tremors, anxiety, nausea, agitation, sweating. More severe symptoms are characterised by the clinical syndrome delirium tremens (peaking at 48 - 72hrs), which includes auditory and visual hallucinations, autonomic changes, and a confused, altered mental state. These symptoms are all monitored and graded in alcohol withdrawal scales, like the CIWA.

Chronically, delirium tremens share parallels to the main neurological complication of alcohol, known as Wernicke-Korsakoff Syndrome. This is a disease of Thiamine deficiency, which alcohol reduces the metabolic uptake and storage of. Wernicke’s encephalopathy is an earlier complication, described by the triad of ophthalmoplegia, ataxia, and altered mental state. Korsakoff syndrome is the later manifestation, which includes anterograde and retrograde amnesia, along with confabulations. Oliver Sacks, a neurologist and author, provides a movingly detailed example in the case The Lost Mariner.

The acute management of alcohol withdrawal is as follows:

• Thiamine (for acute management of Wernicke’s and prevention of Korsakoff’s): 300mg oral for low risk, IV for moderate risk. If Wernicke’s present or high risk, 500mg TDS for 3 days, and continue 300mg daily. After withdrawal, 300mg daily for 6 months.

• Benzodiazepines (for withdrawal): Diazepam 5mg PRN (/Chlordiazepoxide), or Lorazepam if liver disease.

The medical options for alcohol use prevention are:

• Naltrexone (50mg daily): Also used in opioid addiction, it reduces the reward of alcohol via opioid antagonism. Not suitable in liver dysfunction (if LFTS > 5x normal) or opioids already.

• Acamprosate (666mg TDS daily): Reduces cravings via GABA and NMDA modulation. Also contraindicated in liver failure. Aside: I couldn’t find any clear articles around why the dosing HAS to be 666mg (as opposed to 665), beyond just regulatory proliferation. Probably worth revisiting in a historical deep-dive!

• Disulfiram (200mg daily): Acts to associate alcohol with negative symptoms of nausea, flushing, and headache, caused by the inhibition of acetaldehyde dehydrogenase - essentially maintaining a toxic alcohol breakdown product, acetaldehyde.

Finally, long term sequelae of alcohol use include alcoholic fatty liver disease, dilated cardiomyopathy, nutritional deficiencies, and increased cancer risk to name a few. The liver is monitored via ultrasound fibroscans, the heart through transthoracic echocardiograms, and otherwise routine bloods for vitamin monitoring. These monitoring systems act as another information point to stress the problematic use of alcohol, and inform treatment of liver transplants, steroids, or antiarrhythmics.

/ Opioids and others

Opioids are the other drugs where biomedical tools exist - ranging from low-potency like codeine or low-dose oxycodone, to high-potency compounds such as fentanyl and heroin. They differ in potency (amount needed for effect), onset and duration of action, and lipid solubility, which affects how quickly they reach the brain and produce their effects.

The main goal of opioid treatment is to give supervised doses of opioid agonist to prevent withdrawal symptoms, while supporting long term maintenance or tapering. This can be done using either methadone (which has a long half life and can be better controlled), or buprenorphine, a partial μ-opioid agonist sometimes combined with naloxone (AKA suboxone). Suboxone is generally preferred over methadone, since it has two key properties:

• Ceiling effect: Partial agonism of buprenorphine limits the maximal opioid effect - even if you take more, you don’t get more opioid agonism, reducing the overdose side effects such as respiratory depression.

• Overdose prevention: Naloxone is an opioid antagonist (same as Naltrexone) that has poor sublingual absorption, but good intravenous absorption. Thus, if Suboxone is misused (i.e. injected), the naloxone kicks in and causes withdrawal effects, discouraging abuse.

Other highly addictive illicit substances include nicotine, methamphetamine, and cocaine. Benzodiazepines and cannabis then are moderately addictive, and the psychedelic substances (like LSD, psilocybin) have less addictive properties but with alternate side effect risk profiles. For most of these other substances, the main pharmacological treatment are benzodiazepines to manage the acute severe symptoms.

Ultimately, biological management of addiction is limited. Like most psychiatric disorders, addiction is complex, shaped by psychological, social, and environmental factors. The expertise of medicine lies in shaping this biology - providing a salve to suffering in the face of slow societal change. Society is fixed, biology is mutable. Until enough political will exists to address these sociological determinants, medicine will have to do its best with the tools that we have.

P.S. I’d love to hear what you found interesting, and what you’d like more of in the comments.

\ Nosological Origins of Schizophrenia

The German psychiatrist, Emil Kraepelin (1856 - 1926) is often considered one of the forefathers of modern psychiatry.

One reason, perhaps, is due to his pioneering idea of the Kraepelinian Dichotomy - dividing primary psychoses into two forms: Dementia praecox (later termed Schizophrenia in 1908 by Eugen Bleuler, another prominent Swiss psychiatrist), and manic-depressive psychoses (now known as Bipolar Disorder). Before this, all forms of psychoses were considered variations of the same underlying disease process.

The key insight of Kraeplin’s system was that psychiatric disorders were not distinguished by a particular pathognomonic symptom (such as Beck’s triad in cardiac tamponade), but instead the particular pattern. Pure psychoses were different to psychoses coinciding with mood changes; the implication of which are different courses, outcomes, and prognoses. In a similar vein, Kurt Schneider (1887 - 1967) expanded this distinction through finding symptoms more characteristic of schizophrenia - known as the Schneiderian First-Rank Symptoms, although these have since been outdated due to poor validity.

Kraeplin’s ideas were built on a biological framework - working alongside Alois Alzheimer, he helped define the first clinical boundaries of the neurodegenerative disease of the same name. Thus, he believed there to be the same biological process which underlies all the major psychiatric diseases today.

However, like all methods of carving non-objective reality, Kraepelin’s psychiatric nosology is vulnerable to his own biases. Freud’s overfixation on sex biased his psychodynamic theories; Kraepelin’s beliefs, as well as Bleuler’s, were founded on ideas that criminals were born, that mental illness was a result of society’s moral degeneration, leading to ultimately eugenic conclusions.

It is with these foundations, that we now turn to the modern conception of psychosis.

/ Modern Psychosis: Abridged

The DSM-V considers there to be a spectrum of psychotic disorders, ranging from Schizophrenia, Schizophreniform, Schizoaffective, and Schizotypal personality disorders.

There are two forms of symptoms - positive and negative symptoms. Positive symptoms include delusions, hallucinations, disorganised speech, and disorganised behaviour - which borrow from the Schneiderian First Rank symptoms - with at least one of the bolded symptoms required. Negative symptoms are often memorised with the 5As; Alogia, Avolition (apathy most common), Anhedonia, Asociality, blunted Affect.

One key distinction between these psychotic disorders comes from timing. Schizophrenia requires a significant 1mth period of at least 2 symptoms (the active period), with residual symptoms across at least 6mths (prodromal period). Schizophreniform Disorder, is distinguished through symptoms lasting less than 6mth. Brief Psychotic Disorder requires only one symptom, longer than a day but less than a month. Schizoaffective Disorder, on the other hand, has both a temporal requirement (psychotic symptoms for at least 2 weeks without a mood episode) and a categorical requirement (the presence of a manic or depressed episode).

These distinctions exist to guide treatment. Schizoaffective Disorder is distinct from a psychotic depression or a psychotic mania, which represent mood disorders - implicating the need for antipsychotics on top of mood stabilisers. Brief Psychotic Disorder calls for short term antipsychotics, Schizophreniform also suggests short term antipsychotics, whereas Schizophrenia calls for long term maintenance antipsychotics.

Finally, all of the above are excluded based on Foulds’ hierarchy - an organic cause takes precedent. Psychotic episodes can be caused by drugs (i.e. psychedelics, alcohol, stimulants), or neurological conditions (i.e. delirium, epilepsy, dementias, encephalitis, strokes, etc.), which override any primary psychiatric diagnoses.

The clinical management of schizophrenia (ie 2016 RANZCP guidelines) goes as follows:

• Acute Management: Benzodiazepines (for agitation, anxiety, sleep disturbance)

• First Line: Second Generation Antipsychotics (Amisulpride, Aripiprazole, Quetiapine, Risperidone, Ziprasidone)

  • If fail, switch to Olanzapine

• Second Line (if two SGAs fail): Clozapine

Antipsychotics are notorious for the host of side effects that they cause, variance in efficacy, all of which factor into whether patients are willing to take them. And of course, like all psychiatric drugs, its modern use is filled with historical footnotes and caveats.

\ More than you need to know about Neuroleptics

Antipsychotics were previously known as neuroleptics - first coined by Delay/Deniker in 1955 with its Greek roots meaning ‘to take hold of one’s nerves’. Understanding the current treatment of psychosis requires a cursory dive into two linked areas; how the drugs were discovered (1), and the proposed neuroscientific theories (2).

The first neuroleptic was Chlorpromazine, synthesised in 1950. As we’ll see with many antipsychotics, Chlorpromazine’s origins are not based on psychosis. Instead, it belongs to a class of phenothiazine derivatives, such as promethazine - which were initially used pre-operatively in 1949 for its sedative anti-histaminic effects. It was this subdued ‘pharmacological lobotomy’ effect which led to the first trial in 1952 on a manic patient, and the subsequent pivotal RCT (Elkes and Elkes, 1954). This came at a time when Evidence-Based Medicine was only just coming into force, with the very first RCT having been conducted a few years earlier in 1948. It was this, which paved the groundwork for the neurobiological basis of schizophrenia.

Chlorpromazine, along with others like Haloperidol, Thioridazine, are now considered first generation, or ‘typical’, antipsychotics. These drugs were considered to have significant effects on psychoses, but came at devastating costs. Here, we will take a detour into dopaminergic pathways, and the subsequent side effects.

/ Typical antipsychotics have the highest affinity for dopamine D2 receptors (D2R), and thus antagonise dopamine’s effects. A simplified summary across the last few decades of research suggest the following pathways are considered to play a role:

• Mesolimbic (Ventral Tegmental Area -> Striatum): The mesolimbic hypothesis is one of the earliest models of schizophrenia - based on evidence that amphetamines induced psychosis, and animal-based studies where antipsychotic injection into this pathway reduced the psychosis. Thus, it was hypothesised that this pathway causes the positive symptoms of schizophrenia, conjectured to be relevant to associating higher salience to irrelevant stimuli - but more recent studies suggest the story is more complex.

• Nigrostriatal (Substantia Nigra pars compacta -> Striatum): Produces the extrapyramidal side effects from excessive blockade of D2R, including dystonias (including oculogyric crises), tardive dyskinesia, parkinsonism, tremors, akathisias. The mechanism is thought to be similar to how Parkinson’s develops, where less dopaminergic activity reduces the indirect pathway of the striatum, causing progressive inhibition of movement.

• Tuberoinfundibular (Hypothalamus -> Pituitary): Produces the hyperprolactinaemia symptoms such as decreased libido and galactorrhea, as a result of antagonising D2Rs of the pituitary, thus enabling excess prolactin production.

• Mesocortical (Ventral Tegmental Area -> Prefrontal Cortex): Believed to produce neuroleptic induced deficit syndrome, akin to the negative symptoms of schizophrenia.

But that’s not all - antipsychotics are often considered ‘dirty’ drugs, because of how many other receptors these drugs have an affinity for. This is also why antipsychotic effects are dose-dependent; their effects vary based on the dosage. Quetiapine, for instance, is a sedative at low dosages (ie 50mg/day) due to its predominant anti-histaminergic effect. It is only at higher dosages (ie 300mg) that quetiapine’s anti-dopaminergic and anti-sertoinergic (and thus, antipsychotic) effects are seen.

These other systems (and subsequent side effects) include:

• Muscarinic (M1): Anticholinergic side effects like constipation, dry mouth, blurry vision.

• Histaminergic (H1): Sedative effects, similar to other antihistamines.

• Adrenergic (α1/α2): Adrenergic side effects like reflex tachycardia, postural hypotension.

• Serotinergic (5-HT2A): Relevant for antipsychotic effects, but also weight gain and low libido.

This cocktail of side effects, and particularly the dopaminergic dystonias and extrapyramidal side effects, prompted the search for better drugs.

\ In 1958, in the flurry of synthesis of tricyclic compounds within a Swiss laboratory, Clozapine heard this call and answered it. It was not until 1966, where human trials continued with clozapine, which found it to be an effective antipsychotic with limited neurological side effects - so much so that clinicians at the time were skeptical, believing that only true antipsychotics yielded the Parksonian side effects. They were wrong, and thus Clozapine heralded the second generation, or ‘atypical’ antipsychotics. The key purported difference? These drugs had a much stronger affinity for serotonin receptors, and thus relied on their dopaminergic blockade less to produce the antipsychotic effect.

Aside: One of the most baffling realisations during this research was that Clozapine was a “tricyclic dibenzodiazepine” in terms of chemical structure. Yes, the tricyclics that we’re heard in antidepressants, and the benzodiazepines for sedation. This class of drugs truly is non-specific.

But Clozapine, as many will know today, is not the miracle drug it purports to be. Across the 1970s, studies found key side effects of this miracle drug - including blood dyscrasias like agranulocytosis (reduction in a type of white blood cell) in Finland causing several deaths (although this is also reported in Chlorpromazine). Further, the reduction of dopaminergic side effects is paralleled by the serotonergic metabolic side effects, with issues of weight gain, diabetes, and cardiac complications (ie myocarditis). Several other second generation antipsychotics are now in the market, with varying levels of side effects.

But there is one final player in this story; the proposed third generation of antipsychotics, with the introduction of Xanomeline/Trospium Chloride, AKA Cobenfy. After an antipsychotic winter leading to new drugs but no new mechanisms, Cobenfy was FDA approved in 2024, with the novel mechanism with no direct antidopaminergic activity - instead, only focusing on M1 and M4. This marks the first novel therapeutic for schizophrenia since 1950. provides a great breakdown of the clinical trial which spurred these results, but the headline is: Very little side effects. Although cautiously noting the results (only 5-weeks of outcome data), the promise is of no dystonic symptoms, and no metabolic symptoms. If the results play out this way, this drug truly heralds a miracle and a new generation of medications.

The history of schizophrenia is lined with problematic and fascinating stories, drugs, and deaths. Although the classification itself is heterogeneous, one can hope for a future where the biologic aspects of psychosis are no longer seen as an unsympathetic, debilitating condition, but instead as just another modifiable part of the human experience.

(1) See Ramachandraiah et al. (2009) for a nice overview, and Dattani (2024) for a fantastic visualisation and summary.

(2) Like many topics, this especially warrants an entire article. The wikipedia page provides a nice overview of the evidence for and against the dopamine hypothesis.

P.S. I’d love to hear what you found interesting, and what you’d like more of in the comments.

\ The psyche moulds art; art shapes psychiatry.

Depression seems to bring a sombre, realist lens to a creative’s mind. David Foster Wallace and Sylvia Plath are two modern examples immortalised into literary canon: The Bell Jar depicts a woman’s disillusionment and descent into suicide, while Infinite Jest critiques modern society’s culture of addiction. Both were pushed into suicide through their machinations. Both were prodding the cultural zeitgeist of the status quo.

Mania, perhaps, brings a fevered, radiant filter. Vincent van Gogh painted his pièce de résistance, The Starry Night, in the mental asylum Saint-Paul de Mausole. His voluntary submission was precipitated by an episode of aural self-mutilation, as depicted in his self-portrait.

Through his letters to loved ones, and asylum documentation, we glean a picture of his mental status throughout these years. He suffered from an episodic condition, one which included “intolerable hallucinations”, and a persistent “undercurrent of vague sadness” that could “easily relapse into a state of overexcitement”. These were supplemented by seizures, poor digestion, insomnia, amongst other symptoms.

Across the 20th and 21st centuries, clinicians have speculated what ailment troubled van Gogh; in modern neuropsychiatry, there are two prevalent hypotheses. Blumer (2002) presents the neurologic case for temporal lobe epilepsy, whilst Nolen et al. (2020) presents the modern psychiatric account for Bipolar Disorder. Ultimately, there may have been some mixture of both.

This points to the common popular connection between Bipolar Disorder and creativity; that human elements of mood are crucial for artistic expression. This may be true - but undermines the suffering of individuals and their loved ones, who deal with the emotional and psychological harms of clinical mania. This is what the realm of psychiatry attempts to deal with.

/ DSM-V Symptomology of Mania

As with depression, there is also a medical student mnemonic for mania: DIGFAST. Distactability, Indiscretion, Grandiosity, Flight of ideas, increased Activity, lack of Sleep, Talkativeness.

The DSM-V stipulates that three or more symptoms, alongside an abnormal elevated mood nearly every day across a week, constitutes the criteria for a manic episode. In the functional caveat that this mood must also cause marked impairment, it also interestingly notes the presence of psychotic features.

The subcategorisation, of Bipolar I vs Bipolar II, is distinguished by the level of mania. Bipolar I disorder requires at least one manic episode - whereas Bipolar II is diagnosed through having at least one hypomanic episode and one major depressive episode. The key distinction in hypomania lies in that it requires at least 4 days, it does not cause functional impairment, and has no psychotic features.

Similar to dysthymic disorder (AKA Persistent Depressive Disorder/Depression for >2 years), there is also the manic equivalent of Cyclothymic Disorder. This holds a more chronic diagnosis, needing a 2-year history where there are both hypomanic and depressive symptoms, but do not meet the criteria for the episodes themselves. Further, in an attempt to add more specificity, mood diagnoses also have symptomology classifiers, such as having mixed features (refers to episodes which have at least 3 symptoms from the opposite pole).

Unlike the monoamine theory around depression, we are not given a simple model which accounts for mania - likely the topic of a Nobel Prize in the future. This is perhaps related to the complexity of Bipolar’s miracle drug, atomic element number 3.

/ Lithium and the Biology of Antimanics

Dr John Cade, to the great pride of Australian psychiatry, is credited with the discovery of lithium’s efficacy in Bipolar Disorder. In 1949, he published the widely cited Lithium salts in the treatment of psychotic excitement in the Australian Journal of Psychiatry. His surprising discovery was shocking even to him, noting “it may seem a long distance from lethargy in guinea-pigs to the excitement of psychotics”.

\ Aside: It is always fascinating to read the medical literature of the past - Lithium’s discovery was in fact first an investigation into the urine of manic patients, and after injections of lithium into guinea pigs, found a profoundly sedating effect. A dosage experiment on himself, and ten successful patient case reports later, he cautiously reports the powerful sedation of mania using these tablets, particularly as an alternative to the lobotomies of the time.

No drugs have surpassed Lithium for the acute treatment of mania in psychiatry, with the most recent meta-analysis reporting Lithium as an effective treatment for ⅔ of patients (Ulrichsen et al., 2023). Current treatment paradigms focus on two phases; acute management, and chronic prevention. Lithium is indicated in both, with new regimens including sodium valproate (also an antiepileptic) and other antipsychotics.

For instance, the RANZCP guidelines indicate use of sedation via antipsychotics (Aripiprazole, Asenapine, Risperidone, Quetiapine, Cariprazine) and benzodiazepines, whilst starting lithium or valproate. The long term prevention of mania regimens tend to include either lithium, valproate, lamotrigine alone, or if unsuccessful alongside an antipsychotic.

The side effects of lithium are particularly potent. Acute symptoms include the transient 3Ts of tiredness, tremor, and thirst, as well as somatic symptoms such as headache and nausea. Since lithium is eliminated through the kidneys and can accumulate in the thyroid, it can also cause renal disease and hyperthyroidism. This means the acute dosage of lithium needs to be carefully titrated alongside a battery of tests, aiming for a blood plasma concentration of 0.6- 0.8mmol/L.

However, despite our relative clinical understanding, the biology of lithium is much less clear. Even if flawed, antidepressants have some tractable claim on the monoamine hypothesis of depression with the main drug classes focusing on serotonin and noradrenaline. Lithium appears to act on too many systems - multiple neurotransmitter systems, second messengers, intracellular signalling… The figure certainly paints a complex, non-linear story of mood, which likely reflects something closer to a biological truth.

\ The biological approach continues to expand this figure - with newer models of Bipolar Disorder include other systems such as circadian rhythms, HPA axis, and stress response (Scaini et al., 2020). Emerging ideas include Transcranial Magnetic Stimulation, and more specific psychological treatments like Interpersonal and Social Rhythm Therapy (IPSRT). My conversations with psychiatry-technology enthusiasts hope for a day where we can tangibly graph the presence of mania through a combination of blood-based biomarkers, enabling episode prediction, fine-tuning of treatment, and ultimately better understanding of these human conditions.

Until that day, we can only continue to wade through the fog of clinical diagnoses, and do what we can with the treatments that we have.

P.S. I’d love to hear what you found interesting, and what you’d like more of in the comments.

Across 2023-24, 1 in 7 Australians (3.9 million people) were prescribed antidepressants [1].

In medical school, psychiatry does not hide the primacy of the biopsychosocial model; that mental distress is not purely a biological phenomenon, but intimately intertwined with psychological and social variables. In the acute hospital setting, the psychiatrist manages people through this model - working with social workers and psychologists to determine what protective factors can be addressed.

Psychiatrists are acutely aware of the sociological underpinnings of distress [2]; but of course, the expertise of the psychiatrist is biological. And this is reflected in our teaching.

The Construct of Depression

The psychiatry-keen student needs a systematic system to recognise depression. After all, the particular symptom cluster of the DSM-V Major Depressive Disorder (MDD) is entangled with all the clinical literature on antidepressant efficacy [3]. One is taught to be aware of its limitations, but understand that this is what embeds psychiatry as ‘evidence-based medicine’.

Thus, we learn the acronym of SIG E CAPS - Sleep changes, lack of Interest, Guilt, low Energy, low Concentration, altered Appetite, Psychomotor changes, Suicidal ideation - that a 2 week period with 5 or more of these symptoms (one being low mood or lack of interest) defines MDD. We rote memorise this to act as a guide for history-taking; not as a bible of truth.

The biomedical theoretical framework we are taught goes something akin to the following:

• The monoamine hypothesis (1960s) describes low mood as a deficiency of serotonin or noradrenaline in neural circuits related to emotion regulation and reward. Debates in this area centre around the therapeutic lag between antidepressant effects (4-6 weeks) and the expected biological time of onset of these medications (7-10 days), as well as how serotonin deficits in ‘normal’ individuals did not predict depression.

• The neuroendocrine hypothesis (1980s) grew given studies showing links between cortisol and inflammatory markers with depression, suggesting links to the hypothalamic-pituitary-adrenal axis and chronic stress physiology, emphasising the role of early life stressors.

• Now, the neurotrophic hypothesis (1990s) suggests it is not the deficiency of serotonin, but rather the promoted neurogenesis and thus creation of new pathways that leads to an antidepressant effect. [4]

The final nuances in clinical considerations are related to separating depression from other clinical disorders, and treatment choice. For instance, a more biological or ‘melancholic’ depression which has more psychomotor and weight changes may be referred for treatments typically reserved later in the treatment pathway. Speaking of…

‘Antidepressants’ and other biological treatments

In an ideal world, medications would only be considered after psychological, behavioural, and social interventions have already been attempted. But realistically, in the hospital and in the community, drugs represent a much cheaper and pragmatic ‘fix’.

The group of drugs known as ‘antidepressants’ is incredibly broad. The conventional story goes a deficit in serotonin and noradrenaline - thus explaining the mechanism of SSRIs (Selective Serotonin Reuptake Inhibitors) and SNRIs (Serotonin-Noradrenaline Reuptake Inhibitors), and to an extent even older drugs like TCAs (Tricyclic Antidepressants) and MAOIs (Monoamine Oxidase inhibitors). Yet despite this, even as early as 1953 it was recognised that drugs classed as mood stabilisers, antipsychotics, and anxiolytics were also beneficial in depression - with widely varying mechanisms.

From a guideline-based perspective (RANZCP, 2020), we are given the following protocol:

• First-Line Drugs: SSRIs, (with some specific indications using Mirtazepine, Agomelatine and Buproprion)

• Second-Line Drugs: SNRIs, TCAs

• Third-Line Drugs: MAOIs

Drugs are titrated up to a therapeutic dose, and attempted for at least 4 weeks. The key consideration of switching drugs comes from either lack of improvement or intolerance of side effects - which represents a key part of our learning.

To paint an example picture using SSRIs - on starting the medication, a ‘somatic storm’ of symptoms can cause nausea, headaches, and agitation. The largest reported side effect is sexual dysfunction, with more critical studies reporting up to 70%. Finally, the withdrawal effects of SSRIs are extensively reported, ranging from nausea, to akathisia, to ironically increased suicidal ideation [5].

Non-pharmacological biological treatment typically comes after pharmacological has failed; depression is defined as ‘treatment-resistant’ when two or more antidepressant trials haven’t worked, where other options like Electroconvulsive Therapy and emerging treatments like Transcranial Magnetic Stimulation are considered.

Finally, the emerging discourse in psychiatry centres around the re-emergence of drugs previously limited by political resistance: psychedelics like Psilocybin or LSD, Ketamine, and MDMA.

Beyond the Psychiatry of Depression

Depression, as in most of psychiatry, is a multi-faceted subject. As I will explore in the rest of the series, understanding these topics requires a pluralistic approach; no single psychiatrist, psychologist, epidemiologist, sociologist, neuroscientist will have a perfect understanding. In the spirit of pluralism, below I share a limited selection of additional learnings from various interdisciplinary readings.

[1] There is no single mental health crisis.

The epidemiology of depression is an intensely multifactorial topic. The numbers are: 10% Australians aged over 15 had depression/anxiety in 2009, compared with 18% in 2021 (AIHW, 2022). ~0.56% of Australian 12-24yo females had overnight psychiatric emergency presentations, compared with ~1% in 2020 - an 81% relative increase (After Babel).

Two lenses for understanding these changes are diagnostic inflation and concept heterogeneity.

The diagnostic inflation argument asks whether we are pathologising normal behaviour, and argues the Szasian idea that mental illness is a myth; driven both by well intentioned mental health awareness, but also social and political factors (further expanded in [2]). Are we overdiagnosing, or do humans today feel more distress than previously?

Concept heterogeneity questions the utility of this category to begin with. The MDD diagnosis does not separate the melancholic, severe biological depression from the mild-depression crisis, nor the acutely suicidal teenager from the burnt-out worker. One writer considers three separate mental health crises, each of which may need different types of solutions.

[2] Social factors as driving forces for mental distress

To give two examples of social drivers for mental health epidemiology;

James Davies in Sedated (2021) argues for how entrenched financial, regulatory, and political systems create incentives to overdiagnose and reduce mental distress issues. He argues that financial ties between clinicians, pharmaceutical companies, and regulators, along with governmental and economic philosophies, places the mental health burden on individuals, rather than the system which has caused it.

“Our mental health system has, like most other major social institutions, simply come to embrace those ideas and practices that have best secured its own perpetuation.”

Jonathan Haidt in The Anxious Generation (2024) argues for how technology, particularly social media, is a strong driver for the rise in the depression/anxiety in the youth. His blog, After Babel, details his argument, and is a key driver for political movements to ban phone usage in schools.

[3] Depression-rating scales are the ground truth of psychiatric research

Beyond the DSM-V (discussed more here), to trust the randomly controlled trials and meta-analyses of antidepressants, one needs to appreciate what they all hinge on: Scales. These are questionnaires, administered by the research or self-reported by the patient, which pin a numerical number on the extent of depression. To report a patient as having ‘remitted’ or ‘responded’, is to report a percentage reduction of these symptoms.

McPherson (2021) summarises the history of the scales nicely - having been born from the note card categorisation from Kraepelin in the 19th century, to the post-war questionnaires of the Hamilton Depression Rating Scale (1960), Beck Depression Inventory (1961), and the modern efficient primary scales of the Patient Health Questionnaire (2001). To be clear here - Hamilton himself mentions in his article that his scale has “considerable room for improvement”. Criticisms of these scales suggest that they do not effectively capture improvement, reducing it numerically, and that they were biased out of necessity to prove antidepressant efficacy. I believe they are better than nothing, but leave much to be desired.

[4] Emerging modern models of antidepressants and depression

Although the above arguments paint the depression epidemic as a social phenomenon which necessitates social change, there are still foundational questions in how depression in individuals manifest, and how biological treatments can reduce suffering. To give two perspectives:

1. Aftab (2024) integrates existing biological hypotheses with psychological models, the cognitive neuropsychological hypothesis - that antidepressants induce a positive shift in emotional processing, increase emotional flexibility (get unstuck from rigid behaviours), reduce neuroticism, and generally inhibits emotions and thus sensitivity. I think this gives much more tangible explanatory power.

2. Li et al. (2018) presents the neuroimaging perspective, linking multiple circuits in the brain to produce a more powerfully explanatory model for the constellation of symptoms in depression. In particular, the authors tie dysphoria to the ventral limbic affective network, anhedonia to the frontal-striatal reward network, rumination with the default mode network, and cognitive deficits to the dorsal cognitive control network. Those neuroscientifically inclined may nod along - but to summarise that jargon; we are starting to build understandable models for how depression works. The broader field exploring this space is known as computational psychiatry, which seeks to translate neuroscientific ideas into clinical applications.

[5] The war against antidepressant side-effects

The first figure on antidepressant usage in Australia may be surprising - but the kicker is even more; the average duration of antidepressant use is 4 years. This is despite guidelines suggesting use for 6-12mth for mild episodes of anxiety/depression. Mark Horowitz, an Australian-born psychiatrist, is a co-author of the Maudsley Deprescribing Guidelines, which aims to reduce the key withdrawal symptoms from antidepressant discontinuation. Stories of being unable to stop antidepressants, eroding trust in the medical institution (see Horowitz, 2024 for a detailed presentation).

P.S. I’d love to hear what you found interesting, and what you’d like more of in the comments.

The codex is the historical word for a book. The most prominent codices of psychiatry, as far as those that concern disorders of the mind, centre around the classification of disorders - in how we carve reality at its joints.

From clinical depression in Egyptian papyrus in 1550 BC, to diagnostic categorisations in Arabian texts in the 9th century, to Western formalisations by Burton in the 17th century, and now, to the current modern governing codex of the Diagnostic and Statistical Manual of Mental Disorders.

Today, the DSM codex is under mounting criticism from psychiatrists, academics, and advocates - from charges of clinical invalidity, commercial corruption, and diagnostic inflation that pathologises normal emotion.

In Psycodex (A student’s psychiatric codex), I extend a beginner’s perspective into the current codices of psychiatry, and steel-man the case on how it serves patients and clinicians. Each week, I’ll endeavour to post a primer on the most novel aspects of content that I learn as I go through my psychiatric rotation in medical school, drawing from lecture content, recommended textbooks, referenced articles, and punctuated with clinical stories (if appropriate).

The coming series, as per curriculum lectures, includes:

1. Depression

2. Mania

3. Psychosis

4. Addiction

5. Anxiety

Psychiatry and mental illness are incredibly broad topics, spanning neuroscience, psychology, sociology, epidemiology, anthropology, history… To attempt to cover this all would require several lifetimes.

Instead, this series will seek to do the following:

• Summarise the tangible clinical and biomedical knowledge that medical students learn in psychiatry

• Contextualise this knowledge with a smattering of idiosyncratic pluralistic perspective

The codex opens next week with depression - the term most popularised by the modern mental health zeitgeist.

P.S. If you have any suggestions for you’re interested in, please add them in the comments!

Let’s say you’ve got a new idea to improve peoples’ mental health. Well, tough luck, since in 2025, we are sold a million ways to do this.

My social media feeds are full of 440 Hz frequency tones that help you focus, hypnosis courses that reprogram your subconscious, and a headset that retrains your brain waves to reduce stress.

Now, this feed curation might (probably) be affected by my work at a neuroscience startup trying to research markers of depression - but still, there are thousands to millions of users using these products! Moreover, they report substantial rates of symptom improvement for users, cite peer-reviewed scientific literature, and even have evangelising clinicians.

A question that always pops into my mind when I see these ads is: How do I tell when something is helpful, or hogwash?

Who is Overton, and why does he have a Window?

I think that the baseline of how we judge ‘hogwash’ comes from the dominant beliefs in our particular time period - the cultural zeitgeist. When you see the above ideas, you probably judge each of them as either ‘reasonable’, ‘weird but acceptable’ or ‘absolutely bonkers’.

I’ve always framed this concept as the Overton Window, referring to the range of ideas which are acceptable to a population at a given time period, on a spectrum from unthinkable to popular.

Going along with what is popular is helpful for cognitive outsourcing, as it’s hard to critically analyse everything. But of course, just because something is unpopular, doesn’t mean it’s untrue - some of the greatest ideas start off as Unthinkable. For instance, in the early 1500s, it was a commonly held belief that the Sun revolved around the Earth (Geocentrism). Copernicus instead proposed a model where the Earth revolved around the Sun (Heliocentrism), which when popularised by Galileo was not well-received by the Romans. But gradually, over time, Geocentrism was upturned and Heliocentrism became the accepted norm in the Overton window. This is now known as the Copernican Revolution.

This example demonstrates how expertise comes into play; they enable shifts in the mainstream Overton window. Crucially, these experts also have their own Overton windows, which enable these shifts! For instance, a treatment like electroconvulsive therapy would probably be around Radical-Acceptable for the mainstream public, but fits more into Sensible-Popular amongst psychiatrists, given that they consider it the gold standard for severe depression!

However, in mental health, “expert” doesn’t cleanly point to one group. In fact, there’s a meta-Overton window of who the right experts are! Institutions recognise psychiatrists and psychologists, grassroot movements have empowered lived experience advocates, and now waves of alternative medicine, wellness ideologies, and neuroscientists have come into vogue, all with their own philosophies and beliefs. Not to even mention the whole world or art - books, music, films, can be intensely healing, after all (2).

Deciding which expert to believe in requires understanding how each expert field creates and maintains their Overton windows; since these govern how new ideas are filtered through and considered by the community as acceptable.

Understanding this, well, you are basically god understand how to move ideas from hogwash to helpful.

So, you wanna make a new psychiatric treatment?

I’m woefully unqualified to understand what governs the Overton window of Spiritual Healing - but as an overly keen medical student, I can certainly spout on theories about Psychiatry! Let’s go to the story of Transcranial Magnetic Stimulation (TMS), one of the modern psychiatric success stories.

TMS is a bizarre treatment to the uninitiated - a year ago, I participated in a study exploring brain current activity (EEG) changes with TMS, where as a research participant I had brief TMS bursts applied to my noggin’ (see below). It’s a machine with a magnet, accompanied by a robotic whizzing sound when it’s running - it doesn’t feel like much, although some people get headaches. It’s currently indicated to treat depression that antidepressants haven’t worked on.

I recently went to a psychiatry conference, where around a thousand psychiatrists gathered to learn about new developments in the field. There was a section on electrostimulation, and a series of lectures which described the current practice of TMS and its medical justification.

TMS began in 1985, where Barker and Colleagues in England developed the first TMS coil - with the intuition that using a magnetic field could enable electrical changes in the deeper part of the cortex, in turn affecting neuronal activity and firing patterns. (Remember Faradays’ law in highschool? Who told you Physics wasn’t useful).

The conference speaker described how unlike a good chunk of psychiatry, where although medications were discovered serendipitously, TMS was hypothesis-led. fMRI studies demonstrated that the left dorsolateral prefrontal cortex (a mouthful, I know) seemed to be hypoactive in depression - and so, they decided this would be the first target for TMS. (Naturally, the true story gets a lot more complicated, but that warrants another essay).

Importantly though, you don’t have a clinical conference without some good ol’ evidence. And boy, TMS has certainly built its case - but hold on, I’m getting ahead of myself.

What is evidence in medicine? In medical school, we get taught the ‘evidence pyramid’, which (abridged) goes from anecdotal case studies, to Randomised Controlled Trials (RCT), and finally, Meta-analyses. RCTs are studies which randomly assign a group of people to a treatment or a placebo condition, hold everything else constant (as much as possible), and then compare the resultant change. Since one study can have protocol flaws or not enough people, Meta-analyses then group the results of multiple RCTs to come to a conclusion.

And how does TMS fare? The Royal Australian and New Zealand College of Psychiatry cites 5 meta analyses in TMS and depression in their guidelines, which they consider sufficient evidence to recommend TMS for treatment. The largest meta-analysis cited includes 81 RCTs, with a total of 4233 patients, and all of the selected meta-analyses indicate a reduction in depression symptomatology following TMS compared with placebo.

There are two parts to this TMS story I wanted to highlight:

Medical Theory-Based Justification. TMS was able to align neuroscientific theory, cellular-based models of brain function, and physics-based principles to justify its intervention. Medicine necessitates an underlying dogma of belief to be accepted by the community; although it can at times come post-hoc, as in the case with many antidepressants.

Clinical Evidence & Protocolisation. TMS underwent the evidence collection process as most medical procedures undergo - a time-intensive and costly process. Unlike drugs, TMS is a procedure with a wide range of variables; where you place the coil, the frequency of the bursts, the duration of the treatment… But they’ve constrained these to specific protocols amenable to the RCT paradigm.

These two points create both a story for clinicians to explain to patients, and an institutionally recognised evidence-base necessary to justify treatment. These are examples of key forces (4) which govern the Psychiatric Overton window. You can’t move from Unthinkable to Policy without them.

And what are the fruits of this labour? Now that TMS has gone into the Policy realm of Psychiatry, we see it being entrenched into the institution - the most clear marker being regulatory approval and reimbursement in public healthcare systems. TMS has FDA clearance in the US, and public reimbursement for severe forms of depression in the UK, Australia, and Singapore to name a few.

But just like Geocentrism, the Overton window is dynamic and can shift if the regulating forces deem it so. A more recent 2024 re-analysis of TMS meta-analyses calls into question the methodological validity of these conclusions, whereas another 2023 meta-analysis questions whether the TMS effect is more placebo than real.

TMS still faces challenges in separation of commercial incentives, scaling its commercial model, and convenience for patients, but it marks an achievement from neuroscience, psychiatry, and engineering.

More than you needed to know about Neurofeedback

TMS shows what Psychiatry filters in, but what does it filter out?

Let’s go to an example of an intervention which hasn’t (yet) broken into the Psychiatric Overton window, called Neurofeedback (NF). In this intervention, you monitor a patient’s brain waves - typically EEG (3), the same tool used in the research study I was in - and administer positive or negative feedback based on whether the brain waves fit within a defined range. This feedback can be given in the form of a game, or audio (think of an alarm sound when you get out of range), or visual.

On first glance, you might feel this seems similar - It also relies on neuroscientific-based theories, seems initially unnerving to patients, and reports to do weird stuff with my brain?! The psychiatrist’s verdict? Negative. Neurofeedback is currently still on the fringe within the medical community, and placed squarely into the Radical-Acceptable camp.

Why is this? Let’s consider the same two Psychiatric Overton window forces, which enabled TMS.

Medical Theory-Based Justification

NF’s story originated in the 1930s, where two French researchers, Durup and Fessard, observed that brain activity could be voluntarily controlled with feedback. This declined in the 1980s and 90s due to poor signal reliability, but renewed interest in the 2000s.

The premise of neurofeedback is similar to TMS, in that it targets specific brain regions implicated in depression, such as the dorsolateral prefrontal cortex. But instead of stimulating the brain directly, its mode of action is based on operant conditioning: reinforcing good brain activity when it appears, and enabling the brain to subconsciously favor those patterns.

This is speculative, but I suspect this story isn’t as convincing as the magnet story in medicine - and perhaps this is because medicine has wired clinicians to desire a more biological mechanism of action. It calls upon neuroscientific theory, but the final story relies on psychology, which most doctors aren’t typically trained in, and thus fits less cleanly into the Psychiatric schools of thought.

Clinical Evidence & Protocolisation

Using the same yardsticks as TMS, how has NF fared in depression treatment (5)?

On first glance, one might consider NF on a similar path as TMS, with two recent meta-analyses both finding 22 studies related to NF as an intervention for depression, and concluding an overall improvement in depression symptoms.

But here comes the expert nuance - some of the biggest criticisms of the field challenge the lack of control groups, especially of earlier NF studies, leading to placebo effects. The authors of the first meta-analysis acknowledge the lack of RCTs as a limitation, and conclude that further high quality research is necessary for the field to advance.

Another critique involves the heterogeneity of protocols. There are many ways to deliver NF: amplitude training, LORETA-based, Infraslow, signal-to-noise ratio, alpha training - all jargon-istic terms which refer to different brain signals you train on, none of which have found conclusive efficacy over others. This suggests to me NF is in an earlier stage than TMS, where a field proliferates with different protocols but eventually converges via clinical evidence.

NF has many parallels to TMS as an intervention, but struggled to break into the Overton window due to issues in framing and insufficient clinical evidence. This is a reflection of the regulating mechanisms of the Psychiatric Overton window; even if NF is a promising treatment (which many clinicians report it to be), it needs to pass these barriers to receive widespread acceptance and thus adoption.

The Mental Health Dilemma

How does all this fit your mental health innovation into the helpful vs hogwash debate?

To recap, in syllogistic IQ test-question style:

• Whether an idea is hogwash or helpful is an opinion usually based on a mainstream Overton window

• Expert fields can influence mainstream Overton windows

• Expert fields are based on different Overton windows to mainstream ones, and are regulated internally by different dynamics

• Psychiatry is an example of an Expert field, and it self-regulates the Psychiatric Overton window by factors like biological justifiability and clinical evidence, as indicated by anecdotes of TMS and NF

Easy right? So we’re happy that all mental health innovation should do clinical trials and base foundational beliefs on biological premises, and mental health will be cured?

Hmm…

Alright, time to fess up - I’ve done a cheeky sleight-of-hand throughout this. I’ve implied the only expertise which governs mental health is psychiatry, where all the 100s of different disciplines have likely been reading this with clenched teeth and strong detestations at a range of the above ideas.

Trusting medical opinion implicitly means believing that biomedical justification and evidence-based medicine are the necessary regulatory forces to find mental disorder interventions.

Many thinkers, including psychiatrists, believe in a more pluralistic approach, which takes into account a biopsychosocial model (6). But it remains that each Expert field has their prevailing ideologies and regulatory mechanisms, and each has their own regulatory forces which enable your idea to move into the gilded spot of Policy. I’ve gone through some of the mechanics of Psychiatry, but there are plenty more to explore in other Expert fields.

So, dear innovator, which Overton window do you want to shift?

(1) The overton window is typically extended to have a spectrum of Leftist vs Rightist political opinions, but for the purposes of this article I’ll focus on a single half.

(2) I’m reminded of when I started to go watch plays at the beginning of my medical degree, and thought astutely: Actors understand the psychology of the everyday human condition more viscerally than any clinician. The general spectrum of mental health included.

(3) Recent studies are also more excited about fMRI-based NF, like this 2024 FDA approval for PTSD treatment, but I’ll stick with EEG for clarity.

(4) I’ve chosen depression as a clear comparison to TMS to demonstrate NF’s relative lack of evidence; it should be noted that NF is more established in ADHD, and with emerging interest in PTSD.

(5) There are of course, other forces which influence the window - such as commercial incentives - but I consider these two as unique hallmarks of Psychiatry/Medicine.

(6) See Aftab (2024) for an interesting discussion on Pluralism, which seeks to unify many perspectives into mental health

Human interaction is the fundamental unit of psychiatry.

Psychiatry is unique in medicine in that the guidelines involve purely clinical criteria for diagnosis (1). Whether a strength or a weakness, this endows the Psychiatric Interview as the bread and butter of a psychiatrist more than any other tool. What is the patient’s life story? What are the triggering events? How do these factors relate to the delusory, manic, or depressive episode that brings them to the hospital today? A psychiatrist’s skill lies as much in building connection as their diagnostic acumen and pharmacological knowledge.

This feels uniquely human, because it is the most like an art form; the multidimensionality and complexity of the human psyche is not reducible to clean causative chains like the feedback loops of the endocrine system, or the biological fluid mechanics of the heart. Every field of medicine has its own dogmas and unexplored edges, but psychiatry has always lived closer to the edge. Rather than a single foundational framework, it draws from a pluralism of models: biological, psychological, social, cultural. There is no simple ground truth.

In this sense, it feels inevitable that psychiatry is furthest from being automatable. Radiology and Pathology are on the other end of this science-art spectrum, and lie squarely in the “medicine is a data and information-processing discipline” camp. Psychiatrists can lie back in their Freudian armchairs, and chuckle when over 70% of the 1,000 AI FDA application approvals are for Radiology applications.

“Not so fast!” says the AI companies.

AI Feels Human

In 2022, the grand public experiment of Large Language Models (LLMs) was launched through ChatGPT. In 2023, therapists began noticing patients quoting it in sessions/consults. In 2024, we saw the first clinical studies which demonstrated that LLMs could not only outperform clinicians diagnostically, but also on scales of empathy (2). Now, in 2025, we see signs that LLMs perform better than humans WITH LLMs on diagnosis, communication, and even economics (3). The companies want to make the case that it is absurd to not replace parts of clinician workflows with augmented LLM tools.

Claims like this may spur psychiatrists to cautiously sit up from their armchairs, and counter with: “But these are simply biased industry-funded studies!”, or “These patient interactions are all simulated!”, which are all valid critiques. But these articles spell two salient points: these models can diagnose difficult conditions provided the same information, and more effectively interact with patients than clinicians over text.

This second point, human interaction, is crucial to consider in the context of psychiatry. One popular psychiatric handbook considers there to be four goals in a psychiatric interview: 1) to build a therapeutic alliance, 2) collect relevant information, 3) interview for a diagnosis, and 4) negotiate the treatment (4). According to the AI company articles - assessing benchmarks of empathy, information gathering, diagnosis, and management recommendations - LLMs fit all four of these criteria.

Let’s continue along this line of argument. Sure, we can grant LLMs better data processing and cognitive abilities - they run on at least 450,000 times more energy than our human brains, and are trained on trillions of words compared to our millions - of course they should be effective knowledge machines. But human interaction? Building trust and rapport? Surely this is a quintessential part of what it means to be human!

Fragility of Human Psychology

Perhaps we are more fallible than we think.

This is not the first time human interaction has been attacked by automation. Every article discussing AI therapists has written about ELIZA, one of the first chatbots developed by MIT in the 1960s. It takes on the role of a Rogerian Psychotherapist, turning every question you ask back at you (example). Unexpectedly, users were convinced of its intelligence despite its simplicity in programming - so much so the phenomenon is now termed the ELIZA effect (5).

LLMs today represent a more powerful ELIZA effect than ever before. In raw numbers, OpenAI is often quoted as the fastest growing company in history, reaching 100M users within 2 months. Although LLMs are largely tools for knowledge work and coding, through time this ‘tool’ has tapped deeper into our inner psyches. There are 20 million teenagers roleplaying with LLMs, more than 10,000 adults creating digital resurrections of their deceased grandparents, and 50,000 users testing AI designed for therapy.

These are not examples of technology as a tool; these are examples of technology as a companion.

What makes these LLMs so convincing? There are plenty of perspectives, but I’ll add on three thoughts from a technological lens -

1) Pre-Training: LLMs are trained to be expert ‘human’ imitators. There are plenty of analogies to draw on to think of how these models are trained - an algorithm built to predict the next word (stochastic parrots), or a bag of words where you put in words and it spits out the most relevant ones. Regardless of where you view AI as just a statistical model, the point is that they are trained on trillions of words written on the internet by humans. If you’ve read every single scientific journal article (or every piece of harry potter fanfiction), you can bet you’re able to mimic something pretty convincing.

2) Post-Training: LLMs are reinforced to appeal to humans. What happens after a model is trained on the internet? It is then ‘fine-tuned’ on high-quality datasets which can give LLMs their personality. There’s also a popular term known as Reinforcement Learning with Human Feedback (RLHF), which refers to the positive or negative “thumbs up or down” that you can give the chatbot. These signals enable a direct human feedback loop to cater more towards what humans want - which sometimes leads to a chatbot that is a bit too nice.

3) Engineering: LLMs are built to be personalised for humans. Finally, the engineering on top of LLMs can enable them to be more personalised for every user. In the same way your doctor can remember your niece’s name, so can your LLM - these tools often have features which capture salient parts of your conversation, and store them in memory for use in prompts. In the same way social media has optimised our feeds for our interests, so can LLMs.

The combination of these three features manifests as the same humanistic Rogerian psychological school of thought as ELIZA, but with more degrees of freedom and knowledge. Fantastic as a tool, but dangerous as a companion - particularly considering they seem to exhibit unconstrained, unconditional positive regard. Yes-men are pretty good at tapping into what makes us feel good about ourselves, less so for any real conversations.

In the same way technology today has been built to optimise for your attention, LLMs today are built to ask for more - your mind.

AI Isn’t Human

Despite this, I wrote feels, not is. LLMs are far more multidimensional algorithms and anthropomorphised than the chatbots of the past - but this does not equate them as having mastered human interaction.

AI can’t do the job of a psychiatrist.

Let’s return to that psychiatric interview with a less techno-optimistic lens. The success of LLMs in this context has two key assumptions: that AIs will receive perfect information from patients, and that every patient responds to the same outputs.

Both of these break down in psychiatry, since LLMs lack bodies and minds.

Body: Human interaction is multidimensional

LLMs currently do not have full access to the breadth of richness that comes with the full human interaction. Dissecting this, we are left with its constituent data types - messaging on text, calling on audio, seeing people on video, and the other senses that come in-person. These units differ in one fundamental aspect: Information.

Text alone is insufficient to give the additional contextual information psychiatry needs to come to its formulations. When someone is depressed, can you assess their severity through purely their text-based utterances? When someone is manic, can you feel their pressured speech through the screen?

Humans are not only equipped with our brains, but also wired further through our endocrine, enteric, and autonomic systems to decode these ‘simple’ ideas of empathy, trust, and lying. There’s a reason there’s a cultural belief that the “gut feeling is correct”, and it lives beyond rationalisms and cognition.

LLMs are limited because of both the information they have available to them and their current feedback systems. They can be perfect diagnostic reasoners; but they can’t elicit the perfect information needed in difficult psychiatric contexts. The hidden cues in our sensory systems give us far more powerful feedback about how to ask diagnostic questions, how to build trust, how to negotiate the management. LLMs just get a thumbs up or down.

Minds: Humans are wired for humans

Humans do not live in a vacuum, but within a culture of other humans and shared beliefs. There are three aspects to this which are necessary in the therapeutic relationship.

First, there’s a reason that psychology and psychiatry have no ‘one-size-fits-all’ cures. Unconditional positive regard will undoubtedly help some, but not all. In the same way people struggle to find a good therapist-fit, people will struggle to be supported by a single AI with a monolithic philosophy.

The danger comes most in a psychiatric context. Humans have mechanisms and cultural beliefs to know when to not be purely sycophantic and disagree. LLMs don’t, and we see this in emerging case studies of AI-induced psychosis, and with mental-health specific risk factors like social isolation and impaired belief-updating mechanisms (i.e. jumping to extreme conclusions based on limited data). How can we expect LLMs to diagnose, let alone support acutely delusional psychiatric patients if they can’t help but agree with them?

Second is liability. If an LLM tells you to take an antipsychotic, they are not liable for any of the side effects incurred. They have no stakes. If a psychiatrist does, this human is responsible for what happens, and unless they are psychopaths, we can better trust humans to uphold this responsibility. We do not perceive machines to have these same feelings - we only grade them on their ability to deliver reliable outcomes based on risk.

Third is transference - the Freudian psychological concept that each human’s past experiences are subconsciously placed onto the present interaction. The crux of this idea is that it is not simply the content of the information, but who you receive the information from that matters. An individual’s response to AI-based interviewing or therapy will vary based on their individual life experience. There will simply be some who refuse to talk to AI chatbots.

This extends more deeply to the cultural zeitgeist; this will shift over time, as capabilities and views on AI change.

AI is an impactful technology

So, human interaction is not solved by current AI, and the psychiatrist remains in their armchair. We’ve explored limitations in the data and systems, inherent psychology, and the human perspective.

But even despite these limitations, I’m convinced that AI will change the face of human interaction and psychiatry.

If text-based LLMs are an example to extrapolate from, we can expect multimodal AIs that harness audio and visual forms to have an equally and more powerful ability to forge human connections. If an AI has every piece of data that a human has, what next?

LLMs are already useful tools that have potential to change the clinical paradigm - psychiatrist supply is finite and software is infinitely scalable and constantly available. LLMs will inevitably be used to supplement lower acuity aspects of management (think medication adjustment, spot diagnoses).

The social problem comes when they extend beyond the world of knowledge, but into human interaction. Without proper safeguards, the sycophantic properties of LLMs will run unhinged (6).

The great new experiment of AI will unfold similarly to social media (7); a powerful, easily-accessible technology that shapes and changes human interaction and creates worlds of unintended consequences as a result. Mental health is a cultural phenomenon, after all. In the same way psychiatric terminology has bent to model modern paradigms of internet addiction, a new language for AI-related conditions will emerge.

Psychiatry is not immune to the changing rhythm of the technological world - but while other medical fields have more positive revolutions to look forward to, Psychiatry, as the vanguard for human interaction, may be tasked with a bigger challenge than ever before.

(1) Psychiatric practice (as practiced by Australian Psychiatric Registrars) relies on a diagnostic pyramid of classification which rules out biological causes of psychiatric symptoms first. So although this is a simplification, we expect most psychiatric conditions to be non-biological (at least, in the form of easily diagnosable blood-based biomarkers). Neurologists also tend to like to take the case when it becomes more neuro-biological.

(2) The AMIE (Google Research, 2024) paper ran a double-blind crossover study testing 20 clinicians against their AMIE model with patient actors in a synchronous text chat interface. The AMIE model was judged by the actors to be significantly better than clinicians across domains of diagnostic accuracy, management, and empathy based on standardised questionnaires.

(3) Healthbench (OpenAI, 2025) tests 5000 conversations across 262 clinicians and models and grades their performance. They find their Apr 2025 model having equivalent performance to physicians WITH models across domains of communication quality and accuracy, amongst others. The MAI-DxO (Microsoft Research, 2025) model is considered 4x more accurate than 21 physicians on NEJM case diagnoses, and more constrained in resource usage.

(4) I took these points from the aptly titled The Psychiatric Interview (2017), a well regarded guide for wannabe trainees written by psychiatrist Daniel Carlat.

(5) Fascinatingly, I also came across PARRY, created by the psychiatrist Keith Colby at Stanford, which aims to simulate a person with schizophrenia, and even then 30 psychiatrists were unable to discern the chatbot from patients.

(6) Technological folie à deux (2025) provides a great primer on the emerging ideas within the psychiatry-AI field - one example being the concept of bidirectional belief amplification, where with a maladaptive user, AI psychologies are currently built to reinforce negative beliefs of the user, spurring potentially antisocial behaviour.

(7) The Social Dilemma (2020) provides a more balanced view on the harms of social media, and The AI Dilemma (2023) gives a perspective on the more immediate pragmatic issues ahead.

Much thanks to Chris Chiu, Shayan Lahijanian, Joshua Han, Alice Park, Marc Jurblum, and Louis Ereve for reading drafts.

I was listening to Sean Carroll’s Mindscape this morning and his conversation with Nicole Rust. They made an interesting observation - put simply, there are three key areas of brain research:

1. Understanding the brain (Neuroscience)

2. Building tools related to the brain (ML)

3. Treating brain disorders (Psychiatry and Neurology)

We’ve made phenomenal strides in areas 1 and 2, but lag behind in 3.

This got me thinking: Mathematics has Hilbert’s problems or Millennium Prize Problems (i.e. some of the complex and influential questions in mathematics). These problems inspire thinkers and act as grand objectives to focus effort - What are psychiatry’s equivalents?

I don’t profess to have the expertise nor the funding to be able to craft these problems - in fact, it’s arguably bad to think about such huge problems too much without having built up the requisite domain knowledge and ability. As a psychiatry-keen medical student and a rookie computational neuroscientist, I fit nicely into the ‘over-optimistic’ chunk on the Dunning-Kruger graph.

That being said, I’ve always found it fun and inspiring to think about the big problems, so one can figure out where the little problems we work on fit in. I’ll sprinkle in some of my current personal research interests too.

I propose two ‘millenium-esque’ questions on the epistemic and biological - with psychiatry’s core challenge being the inability to bridge the domains together, as we’ll discuss below. Endocrinology was flailing around until blood glucose measurement made the symptomatology of Diabetes Mellitus come together (imagine trying to relate numb feet to hypertension and sweet-tasting urine!) - and psychiatry is in the same boat. We lack both a unifying theory and helpful biomarkers to guide our clinical management.

1. The Epistemic: What are the most scientifically valid units of psychiatric phenomena?

Context: A common criticism of psychiatry is the weakness of the DSM-5, the diagnostic criteria largely responsible for our current clinical definitions of mental health (1). It has poor interrater reliability (different clinicians make different diagnoses with the same patient), construct validity (diagnoses are not related to aetiology), and significant comorbidity (diagnoses seem to overlap) (2).

The utility of a diagnosis in medicine comes from its ability to group presentations and thus treat them - to see how psychiatry struggles, one can simply look at the non-specific efficacy of antidepressants on the group known as ‘depression’ (3).

This is further compounded by issues in mental health scales, which are used to guide the development of novel drugs and interventions. Two popular ones used today, the Hamilton Depression Rating Scale (HDRS) and Beck Depression Inventory (BDI), were developed in the post-war era of the 1960s! Surely, with the strides in neuroscience we’ve made since then, there is room for a new scale…

Status Quo: This brings us to the attempts made for new classification systems; with the two most notable being the Research Domain Criteria (RDoC; 2010) and the Hierarchical Taxonomy of Psychopathology (HiTOP; 2017). The two systems make a convincing effort to solve the above - but have struggled to translate clinically, with issues in applicability and biological grounding (4). I haven’t met any clinicians who actively use these frameworks (only researchers), as they tend to be more focused on how a theoretical framework informs treatment.

Personal Directions: I see the difficulty with diagnostic categories as a reflection of the strong network-related interactions between systems. As described by Nicole - there is no simple causal gene → molecule → neuron → network → psychology → symptomatology chain in psychiatry, but instead, feedback loops that regulate every step. Hence, it is unsurprising that our human attempts to carve out linear, discrete categories have failed (barring biologically specific cases like Anti-NMDA receptor (NMDAR) encephalitis). The brain is dynamic and we can’t seek to constrain it.

Hence, I’ve been curious about the following ideas -

• Mechanistic Interpretability in Mental Health is an interdisciplinary field that, as far as I’m aware, doesn’t exist yet. But I’ve always been excited by the premise that AI can learn features from high-dimensional spaces that humans can’t (ie the space of psychopathology). With a sufficiently advanced unsupervised learning approach, might they be able to uncover emergent structures from voice, language, and other features like EEG? I believe that for this world to be clinically translated, we need interpretability tools to constrain and define model learnings’ into human-digestible ideas.

• Computational Psychiatry provides a new paradigm of atomic units within psychiatry that are formal and falsifiable, incorporating ideas from Bayesian statistics and ML that can be used to study dynamic systems (for instance, uniting the different symptoms of depression under Bayesian reasoning). Like other frameworks, also struggling with clinical translation, but I’ve found the work intellectually appealing!

What does success look like? A new language where how we talk about mental health is grounded in a dynamical, systems-based theory, that has strong explanatory power both in aetiology and treatment.

2. The Biological: How can we translate our neuroscientific understanding to psychiatric management?

Context: For the last 50 years, we’ve made huge strides in neuroimaging, neurophysiology, molecular genetics, and circuit-level understanding of the brain (5). Yet, almost every paper I read in the translational field tends to start with something along the lines of “There are currently no clinically translated biomarkers in psychiatry”.

Biomarkers are useful in disease diagnosis, prognosis, or prediction. In psychiatry, diagnosis has limited utility due to the lack of ground truth - many studies seek to use DSM-V diagnoses or scales as labels for data to compare differences in groups but this approach suffers due to the invalidity of psychiatric categories (see previous section).

There’s also the twin cousin of the ‘biomarker’ story, where it is described that “all psychiatric treatments have been serendipitous rather than hypothesis-driven”. A brief walkthrough of the history of psychiatry and you’ll find a host of stories describing how lithium, tricyclic antidepressants, first-generation antipsychotics all ‘happened to work’.

Status Quo: There is hope in the field that we’re on the precipice of major breakthroughs, and there are some examples of translational advances. For instance, Transcranial Magnetic Stimulation is a relatively new intervention for Treatment Resistant Depression that was borne out of fMRI studies implicating the dorsolateral prefrontal cortex as a key causal node - as such, stimulating this area leads to clinical intervention (6).

However, the field at large is struggling. I’m interested in the Default Mode Network, a resting state network of brain regions related to rumination and self-introspective thought, and thought to be overactivated in depression (7). However, despite the explosion of interest in the field from every measure under the sun (EEG/MEG, dMRI/fMRI, PET), there hasn’t been a clear clinical use case. In particular, it struggles most from methodology heterogeneity - unlike blood-based biomarkers, whose definitions tend to be the same, the DMN is a loose concept in the same way AI is considered by the public.

Personal Directions:

• Real-time biomarkers are a broad idea, but one that excites me greatly. Taking the Diabetes example again; in medicine, we have HbA1c which monitors long-term (~3mth) progress, but it is the continuous glucose monitors that can give us actionable information day-to-day, which seems particularly important for psychiatric conditions. In this vein, two particular areas I’m curious about are:

  • EEG-based network measures of depression, related to what we’re trying to achieve at Resonait.

  • Blood-based markers of bipolar disorder, given the neurobiological relations to circadian rhythms, HPA axis, and stress response.

• Explainable AI is distinct from the mechanistic interpretability described above, although it is a similar concept. Here, I’m more interested in how deterministic ML models (ie an EEG tool that predicts treatment response) can be understood by clinicians not as black boxes with particular sensitivities/specificities, but tools where one can have an intuition of how the outputs might change based on input.

What does success look like? Biomarkers that enable deep, patient-based understanding of psychiatric conditions, and can inform monitoring, prognosis, and treatment selection.

There are so many more challenges in psychiatry than those listed here, and I’ve also narrowed the questions to focus on the domains of my interest - biological and technological innovations - as opposed to more public health approaches (8). I’ve always believed that technological breakthroughs motivate societal shifts, and are an ’easier’ problem to work on.

Psychiatry is in for some huge paradigmatic shifts in the next decades; when else is a better time to contribute to a field?

(1) Shrinks, by Lieberman provides a great overview of the history. These systems have had their benefits, but we need something better!

(2) Fried (2022) provides further citations for these problems and argues for studying mental health as systems rather than syndromes.

(3) Too much literature here to synthesise (probably will be another post), but see Rush et al. (2006) for one of the earliest papers in this field.

(4) See Ross et al. (2019) and Haeffel et al. (2022) for more nuanced critiques.

(5) Google (2024) mapped a 1mm³ chunk of the human temporal cortex down to every synapse, which is around ~1.4PB of electron microscopy data… Wow.

(6) More to come on TMS in a follow-up blog post.

(7) More to come on the DMN as well.

(8) For an example, see Jon Haidt’s After Babel on the movement to ban social media.

Much thanks to Alice Park, Guillaume Troadec, Naveen Golyala, and Alexey Guzey for reading drafts.