Based on an interview with Dr. Tristan Glatard
Understanding mental health can mean looking far beyond symptoms or diagnoses. The brain does not experience mental health in isolation from the rest of the body. Stress can affect both mental and cardiovascular health, signals pass between the brain and the heart, and communication between the brain and gut works in both directions (Contrada, 2025; Ohara & Hsiao, 2025; Valenza et al., 2025). For Dr. Tristan Glatard, Scientific Director of the Krembil Centre for Neuroinformatics at the Centre for Addiction and Mental Health (CAMH) and inaugural BMO Chair in Artificial Intelligence and Mental Health, this is one reason to think “a little bit beyond” traditional labels and scientific silos. Mental health remains an important focus, but studying it may require looking across boundaries that research and healthcare have historically treated as separate.
Artificial intelligence (AI) offers powerful tools for finding patterns across increasingly complex health data. Researchers can use it to analyze clinical records, brain images, and other biological and social information, develop models that predict clinical outcomes, and create tools that support clinicians or interact with clients. But throughout Dr. Glatard’s conversation with Elemental Magazine, an important question kept returning: not simply what AI can do, but whether its results are reliable enough to trust in healthcare.
What Could AI Change in Mental Health Care?
For someone seeking mental health care, the promise of AI is not necessarily a futuristic robot therapist. More often, it may work in less visible ways: helping clinicians anticipate what support a client might need, making administrative tasks more efficient, or expanding access to certain forms of care.
One familiar example is conversational AI. Researchers are developing purpose-built systems based on approaches such as cognitive behavioural therapy and motivational interviewing, which could help expand access to evidence-based support. Dr. Glatard emphasized that this is different from simply asking a general-purpose chatbot such as ChatGPT to act as a therapist. Early research is promising, but evidence for effectiveness and approaches to evaluation remain inconsistent (Hua et al., 2025). As he put it, the goal is to “apply AI to problems that we have in mental health care.” Importantly, not all AI applications interact directly with clients. “Ambient scribes,” for example, can capture clinician-client conversations and use AI to assist with clinical documentation, potentially making note-taking and related administrative tasks more efficient (Bracken et al., 2025).
Another major area, and one that is central to Dr. Glatard’s research, is prediction. AI models can combine different kinds of information to estimate outcomes such as someone’s risk of returning to hospital or their likelihood of responding to a particular treatment. Depending on the research question, these data might include clinical and social information, brain imaging, biological measures, and data from wearable devices. At CAMH, researchers are studying AI approaches to outcomes such as hospital readmission and treatment response. Recent work has also explored whether information from cognitive testing and brain imaging can help predict treatment resistance in late-life depression.
These predictive applications lead to a less visible but crucial part of AI research. Before a model can meaningfully inform care, researchers need to know whether it is actually learning what they think it is learning, whether its performance will hold up with new data, and whether the outcome it predicts is useful for clinical decision-making.
How Can a Good Prediction Go Wrong?
Imagine an AI model that correctly predicts an outcome 80% of the time. At first glance, 80% sounds impressive. But the number alone leaves important questions unanswered. What did the model actually learn? What was it compared against? Would it perform just as well with a new group of people?
One challenge is that a model can make the right prediction for the wrong reason. Dr. Glatard illustrated this with Parkinson’s disease. People with Parkinson’s tend, on average, to be older than healthy comparison groups. If researchers build a model using groups that differ substantially in age, the model might learn age-related differences rather than features of Parkinson’s disease itself. Its predictions could look highly accurate while being driven partly by the wrong information. Researchers call this kind of hidden influence a confound.
Even testing a model can be surprisingly tricky. One way to think about AI development is like studying for an exam. The training data are the practice questions, complete with answers, that allow the model to learn. The test data are supposed to be the final exam, showing whether the model can succeed on information it has not already learned from. But imagine checking the final exam after every study session. You look at your score, change how you study, take the same exam again, and repeat. Eventually, your performance on that particular exam may become excellent, but that does not necessarily mean you have mastered the subject.
Something similar can happen in AI research. If researchers repeatedly test a model, adjust it based on the results, and test it again on the same dataset, information from the supposedly unseen test data gradually influences how the model is built. Its reported performance may then look stronger than what researchers would see with truly new data.
Researchers also need to ask what “good” performance actually means. Imagine that 70% of people in a dataset are not readmitted to hospital. A model could achieve 70% accuracy simply by predicting “no readmission” for everyone. It would sound reasonably accurate while providing little useful information.
An appropriate benchmark gives that percentage some context. Researchers might compare a model against simple strategies such as always predicting the most common outcome, predictions made by chance, or an existing clinical approach. A new model becomes more informative when it can improve on what simpler or more conventional approaches can already tell us.
The metric used to judge a model matters too. Suppose a hospital develops an AI model to identify clients at high risk of readmission. A model optimized only for overall accuracy might look successful because it correctly predicts the many clients who will not return, while failing to identify many of the clients who will actually be readmitted. If the clinical goal is to identify people who may benefit from additional support before discharge, researchers may be particularly interested in how well the model identifies clients who are later readmitted, while also considering the consequences of incorrectly classifying someone as high risk.
In other words, a model can be mathematically successful according to one measure while being less useful for the clinical decision it was meant to support. Deciding which outcome matters, whether the model is actually capturing that outcome, and how its performance should be judged is therefore part of the scientific question itself.
Taken together, these issues form what might be called the evaluation problem: how can researchers tell whether strong-looking model performance actually reflects a clinically meaningful signal, rather than confounding factors, misleading performance metrics, or quirks of the particular data used to develop and test the model? Rigorous evaluation is therefore not simply a final check after an AI model has been built. It is part of determining whether the model’s results are trustworthy and clinically relevant.
How Do We Know an AI Model Really Works?
This evaluation problem sits at the heart of the Brain Health Data Challenge Platform (https://www.bhdc-pdsc.ca/), a Brain Canada-funded initiative led by Dr. Glatard. The platform supports AI prediction competitions in mental and brain health while developing infrastructure and practices intended to make those competitions secure, reproducible, and equitable.
Data challenges separate, at least partly, the people building models from the process that ultimately judges them. Participants can receive training data with known outcomes and a public test dataset where the outcomes remain hidden, but a leaderboard lets them see how well their models are performing. After the competition closes, final models can be evaluated on a separate private test dataset that participants have not repeatedly checked during development. In other words, the final exam remains a final exam. As Dr. Glatard put it, those responsible for the evaluation should have “no skin in the game” when it comes to which modelling strategy succeeds. The aim is not to demonstrate that one researcher’s preferred algorithm works. It is to find which approach best addresses the problem.
Data challenges can also bring many approaches to the same question. Instead of one research group trying a handful of models, researchers and data scientists from different backgrounds can tackle the problem independently. Meanwhile, clinicians, researchers focused on equity in AI, and people with lived experience can help shape the questions and evaluation criteria. This helps ensure that a challenge is not only technically interesting, but also clinically relevant and attentive to the communities its results may affect.
Privacy adds another layer. Health datasets can contain deeply sensitive information, making open sharing difficult. One strategy being explored by the platform is synthetic data, artificially generated datasets that preserve useful statistical patterns without corresponding directly to an actual person’s health record. Researchers can develop models using these simulated data, while promising approaches can later be evaluated privately using real clinical data that never have to be publicly released.
What Comes Next for AI in Mental Health?
Dr. Glatard sees some of AI’s greatest potential in more personalized care. Better prediction could eventually help clinicians decide which treatment may be most helpful for a particular client, identify who may need additional support, or better understand how conditions develop and change over time.
At the same time, better prediction creates responsibilities. Health data are deeply personal, and breaches of privacy may carry especially serious consequences for people in vulnerable or precarious circumstances. AI systems can also inherit and amplify biases already present in the datasets used to train them, potentially worsening existing inequities in mental healthcare (Timmons et al., 2025). A model that performs well overall may still perform poorly for particular communities.
For students and early-career researchers, Dr. Glatard’s advice reflects the same careful approach: understand the fundamentals, including how methodological choices affect results, but also “learn by doing.” Working with real datasets exposes researchers to ambiguities and unexpected challenges that simplified examples often cannot capture.
Perhaps that combination captures what responsible AI research requires more broadly: enthusiasm for what new technology makes possible, paired with a willingness to question what appears to work. What is the model really learning? Is it answering a clinically meaningful question? Has its performance been evaluated rigorously? And do we understand its clinical usefulness and its limitations?
Good predictions can still go wrong. Asking these questions helps distinguish impressive model performance from evidence that can responsibly inform care. And as AI helps researchers look across the traditional boundaries of mental, brain, and physical health, knowing how to evaluate what it finds may be just as important as building the models themselves.

References
Bracken, A., Babu, A. R., Whelehan, S., Merghani, K., Sheehan, E., & Feeley, I. (2025). Ambient AI reduces documentation time and enhances quality in a simulated inpatient setting. The Surgeon. https://doi.org/10.1016/j.surge.2025.10.008
Contrada, R. J. (2025). Stress and cardiovascular disease: The role of affective traits and mental disorders. Annual Review of Clinical Psychology, 21, 139–168. https://doi.org/10.1146/annurev-clinpsy-081423-023833
Hua, Y., Na, H., Li, Z., Liu, F., Fang, X., Clifton, D., & Torous, J. (2025). A scoping review of large language models for generative tasks in mental health care. npj Digital Medicine, 8, 230. https://doi.org/10.1038/s41746-025-01611-4
Ohara, T. E., & Hsiao, E. Y. (2025). Microbiota-neuroepithelial signalling across the gut-brain axis. Nature Reviews Microbiology, 23, 371–384. https://doi.org/10.1038/s41579-024-01136-9
Timmons, A. C., Duong, J. B., Walters, S. N., Carta, K. E., Jumonville, G. A., Carrasco, A. S., Romero, D. N., Ahle, M. W., Comer, J. S., Khurd, I. P., & Chaspari, T. (2025). Bridging fair-aware artificial intelligence and co-creation for equitable mental healthcare. Nature Reviews Psychology, 4, 793–807. https://doi.org/10.1038/s44159-025-00491-5
Valenza, G., Matić, Z., & Catrambone, V. (2025). The brain-heart axis: Integrative cooperation of neural, mechanical and biochemical pathways. Nature Reviews Cardiology, 22, 537–550. https://doi.org/10.1038/s41569-025-01140-3
