Learning Outcomes that Maximally Differentiate Psychiatric Treatments

ObjectiveMatching each patient to the most effective treatment option(s) remains a challenging problem in psychiatry. Clinical rating scales often fail to differentiate between treatments because most treatments improve the scores of all individual items at only slightly varying degrees. As a result, nearly all clinical trials in psychiatry fail to differentiate between active treatments. MethodsWe introduce a new statistical technique called Supervised Varimax (SV) that corrects this problem by accurately detecting large treatment differences directly from original clinical trial data. The algorithm combines the individual items of a clinical rating scale that only slightly differ between treatments into a few scores that greatly differ between treatments. We applied SV to multi-center, double-blind and randomized clinical trials called CATIE and STAR*D which were long thought to identify few to no differential treatment effects. ResultsSV identified optimal outcomes harboring large differential treatment effects in Phase I of CATIE (absolute sum = 1.279, p = 0.002). Post-hoc analyses revealed that olanzapine is more effective than quetiapine and ziprasidone for hostility in chronic schizophrenia (difference = -0.284, pFWER = 0.047; difference = -0.283, pFWER = 0.048), and perphenazine is more effective than ziprasidone for emotional dysregulation (difference = -0.313, pFWER = 0.020). SV also discovered that bupropion augmentation is more effective than buspirone augmentation for treatment-resistant depression with increased appetite from Level 2 of STAR*D (difference = -0.280, pFWER = 0.003). ConclusionsSV represents a powerful methodology that enables precision psychiatry from clinical trials by optimizing the outcome measures to differentiate between treatments.