Artificial Intelligence in Psychiatric Clinical Trials: Opportunities and Limitations

Artificial intelligence tools often used in drug development today were not optimal for stratifying psychiatric trial populations, shared Joseph Geraci, PhD. A recent study evaluated how artificial intelligence could characterize patient subpopulations from small clinical trial and trial-like data sets, with 2 goals: to reveal new information about major depression and schizophrenia, and to show that stratifying a patient population and identifying the factors driving each subpopulation was highly nontrivial.¹ A cancer data set was also analyzed, and specialized tools provided a measurable lift to standard models.

Standard machine learning algorithms classified and modeled well when objects were relatively simple, such as sports teams or images of cats and dogs, but patient populations presented greater heterogeneity and noise. Labels such as responder, nonresponder, and placebo responder were also less clean than image labels. “The complexity is that the labels themselves are not completely reliable because there’s so much phenomenon loaded behind it,” Geraci said. His mathematical methods were designed so a model could remain honest about which patients in a small data set actually adhered to their labels, allowing researchers to learn from them.

The key element was that the model could also abstain when it lacked adequate information about a patient. “Some patients are going to look like other patients and because of the heterogeneity, you’re just not going to have the sample size to see every type of patient in a psychiatric data set,” Geraci said. Abstention was particularly important because clinical trial data sets were small.

Data availability also differed by field where the research was done. Cancer trials could draw on genetic, methylation, and other biological data from tumor sampling, whereas psychiatric trials relied on clinical scales that, although reliable, varied with the clinicians who asked the questions and rated patients. Depression and schizophrenia also each represented multiple diseases grouped under a single umbrella, and the analysis included the Clinical Antipsychotic Trials of Intervention Effectiveness schizophrenia data set.² The model identified consistency within groups of patients sharing certain variables, isolating the portion of the data with reproducible, replicable evidence.

Dr Geraci is chief technology officer, chief scientific officer, and founder of NetraMark.

References

1. Geraci J, Qorri B, Cumbaa C, et al. Interpretable subgroup discovery with abstention in small, heterogeneous clinical trials: a retrospective multi-dataset study. AI. 2026;7(9):349.

2. Lieberman JA, Stroup TS, McEvoy JP, et al. Effectiveness of antipsychotic drugs in patients with chronic schizophrenia. N Engl J Med. 2005;353(12):1209-1223.