Artificial Intelligence in Psychopharmacology: Promise and Pitfalls, With Roy Perlis, MD
How AI can aid everyday psychiatry—TRD prediction, suicide risk, and smartphone phenotyping—plus why data quality and workflow still block adoption.
BRAIN TRUST: CONVERSATIONS IN PSYCHOPHARMACOLOGY
Series Editor Joseph F. Goldberg, MD
Joseph F. Goldberg, MD, in this installment of "Brain Trust: Conversations in Psychopharmacology," sits down with Roy Perlis, MD, to discuss the role of artificial intelligence (AI) in clinical psychopharmacology, including its applications in treatment-resistant depression (TRD), suicide risk stratification, digital phenotyping, and the challenges of integrating AI tools into everyday psychiatric practice.
Perlis reflected on 25 years of experience in this domain, noting that early neural network models used for psychiatry, like those applied to fluoxetine trial data, yielded little predictive signal.1 He observed that despite dramatic growth in dataset size—from hundreds of patients in early studies to hundreds of thousands in contemporary electronic health record (EHR) analyses—AI models have not consistently outperformed the well-trained clinician. He argued that one of AI's most practical near-term contributions is not superior prediction but rather democratization: giving the average practitioner access to the same structured clinical reasoning that expert clinicians apply intuitively.
Perlis raised important concerns about the quality of the data from which AI models learn. He noted that EHR notes serve multiple purposes—clinical communication, billing, and medical-legal documentation—and that this dilutes the clinical signal in modeling cases. He warned that models can only find what clinicians document, and that key predictors of outcomes such as TRD may simply not be captured in structured or narrative data. Perlis cited his collaborative work with EHR data across institutions, noting that TRD prediction models trained at one site often fail to generalize to another, suggesting the limiting factor is data quality rather than sample size.2
On digital phenotyping, Perlis identified some potential, emphasizing that passively collected smartphone data (like changes in usage patterns, timing, and movement) are most informative when interpreted as deviations from an individual's own baseline rather than as population-level classifiers. He identified integration into clinical workflows as a primary barrier, not data availability or analytic methods.
Perlis urged caution regarding premature deployment of AI tools, particularly suicide risk models, noting that current models still yield incorrect answers approximately 10% of the time. He also addressed large language model confabulation, clarifying that these systems are trained to complete sentences, not to acknowledge uncertainty, and that grounding models in constrained, citable data sources remains an active area of development.
Dr Goldberg is a clinical professor of psychiatry at The Icahn School of Medicine at Mount Sinai in New York, NY and the immediate-past president of the American Society of Clinical Psychopharmacology.
Dr Perlis is chair and professor of psychiatry at Harvard Medical School and Massachusetts General Hospital, director of the Center for Quantitative Health at Massachusetts General Hospital, and editor in chief for artificial intelligence at JAMA Network Open.
References
1. Winterer G, Ziller M, Linden M.
2. McCoy TH, Castro VM, Perlis RH.
Related Content
