AI and Resolving the Heterogeneity of Psychiatric Disease Taxonomy

Artificial intelligence (AI) may play a strong role in drug discovery through biochemistry, said Joseph Geraci, PhD, MS; though AI will not solve all of biology because the combinatorics are complicated, he added. The main obstacle was often not the molecules but the taxonomy of the disease. Heterogeneity meant that a single drug often had to address several distinct diseases at once. Geraci’s work focused on mathematical methods that augmented machine intelligence to resolve patient populations, which allowed pharmaceutical partners to carry findings from earlier trials into later ones and reduce failure.

Classifying complex systems was possible with methods that abstained when they lacked understanding, but this came at a cost. Clinical trial datasets, often fewer than 500 patients, were small and expensive, and models applied to them needed to abstain when they did not understand.¹

Rating scales used by psychiatrists were powerful and useful. Including more of them in a trial improved understanding of the patient population and gave models more features to detect, although findings had to be replicated. Academic psychiatrists who added more scales to their studies achieved better outcomes. Clinical practice was also expected to change as real-world data, wearable data, speech latency, and changes in typing patterns on apps entered care. Such changes could signal onset of depression.

There could be a future where “there’s going to be a proxy to these big genetic and MRI type of tools that come from the scales and other measures you can do locally,” Geraci said. A model that determined whether to prescribe escitalopram, for example, was not yet available, but anhedonia rating scale items, such as enjoyment of a favorite meal, could be incorporated into practice.² Collecting such questions during patient interviews held clear value, according to Geraci: “if they start to collect these questions as they interview their patients, this becomes very valuable data.”

Models would emerge from specialized techniques, and clinicians could contribute through openness about the data they collected and their attitude toward the field. Geraci considered the common understanding of AI misleading and called for techniques specialized for patient populations. His approach represented population factoring rather than conventional stratification because it considered heterogeneity up front and treated the taxonomy of the disease as unresolved.

Dr Geraci is the chief technology officer, chief science 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. Sajjadian M, Uher R, Ho K, et al. Prediction of depression treatment outcome from multimodal data: a CAN-BIND-1 report. Psychol Med. 2023;53(12):5374-5384.