Task-aligned outcome learning in psychiatry: reducing endpoint dilution.

Psychiatric research relies on well-defined outcomes for standardization, comparability, and replication, yet investigators often fix broad endpoints before knowing which symptom domains carry task-relevant signal. Even when psychometrically sound and clinically useful, composite measures can dilute predictive information and attenuate treatment effects when predictability or responsiveness concentrates in only a subset of symptoms-thus making studies appear negative despite meaningful change. This Perspective proposes a task-aligned, two-stage machine-learning framework for learning the appropriate outcome. In the first stage, constrained discovery derives a clinically interpretable outcome from a prespecified item pool. In the second, confirmatory evaluation tests the prespecified hypothesis either on a fixed learned outcome, when the aim is to assess a previously derived endpoint in a closely matched study, or on a relearned outcome generated by the same prespecified procedure, when the aim is to test whether that procedure can recover a task-aligned endpoint across different studies. The framework complements psychometrics and open-science practices, shifting focus from broad unsupervised composites to empirically supported targets, with safeguards to keep results interpretable and rigorous.