Clinical drug development in psychiatry is challenging due to heterogeneous patient populations and the uncertainty of measuring neuropsychiatric constructs with symptom rating scales. Here we describe the development and implementation of an enrichment algorithm that identifies canonical versus anomalous symptom presentations, at the individual subject level, based on MADRS ratings obtained at screening and baseline. Data from 5 randomized, placebo-controlled, phase 3 trials in bipolar I disorder was used (N = 2026 subjects and 15,239 MADRS assessments). A variance-covariance difference (VCD) vector was developed to encode individual symptom structure using the 10 items of MADRS from the two sequential assessments. An anomaly score, calculated from each subject's VCD vector was derived by isolation forest to quantify the degree of disparity from the hypothesized canonical item structure. A retrospective application of the algorithm reliably identified a threshold anomaly score above which the psychometric properties of MADRS deteriorate. Consistent with increasing the certainty of MADRS ratings, subjects with a canonical symptom structure at baseline demonstrated greater effect sizes post-baseline in a phase 2 placebo-controlled trial of non-racemic amisulpride (SEP-4199) for bipolar depression, analyzed retrospectively. Our analyses show that the developed algorithm can reduce the symptom structure heterogeneity at baseline and thus improve the measurement certainty of psychiatric symptoms in clinical trials. This novel enrichment method has been prospectively implemented in a Phase 3 clinical study of SEP-4199 and is consistent with regulatory guidelines aimed at increasing the statistical power and lowering patient-burden in clinical trials.Clinical Trials Registry: NCT00868452, NCT00868699, NCT01284517, NCT01986101, NCT03543410, NCT05169710