Personalized symptom clusters that predict depression treatment outcomes: A replication of machine learning methods

Abstract

Objectives: The purpose of this study is to use independent datasets to externally validate the three symptom clusters of unipolar depression identified by Chekroud, to evaluate personalized treatment trajectories and outcomes based on these symptom clusters, and to verify predictors. Methods: The Quick Inventory of Depressive Symptomatology-Self Report (QIDS-SR16)11 QIDS-SR16: Quick Inventory of Depressive Symptomatology-Self Report. and Hamilton Rating Scale for Depression (Ham-D)22 Ham-D: Hamilton Rating Scale for Depression. data from two placebo controlled, double-blind clinical trials (Dual Therapy and Duloxetine) were used for external validation. Machine learning methods were applied to replicate the three symptom clusters and to produce treatment trajectories. Penalized logistic regressions were conducted to identify top baseline variables that best predicted treatment outcomes. Results: The variables Chekroud identified as comprising sleep, atypical and core emotional clusters are replicated. Treatment trajectories demonstrate that dual treatment (escitalopram and bupropion) performed best across all symptom clusters but did not outperform escitalopram monotherapy over time. For each symptom cluster, there were differences in treatment efficacy among antidepressants. Conclusion: By using different treatment trajectories based on a patient's symptom cluster profile, clinicians could potentially select best fit antidepressants to achieve the biggest benefit. Our results showed that total baseline QIDS, Ham-D score, anxiety disorder diagnosis and course of depressive illness were the best baseline predictors. Results could enhance personalized depression treatment plans and help to improve outcomes. Clinical trials registration: NCT00519428, NCT00360724.

Keywords

• Depression
• Symptom clusters
• Predictive analyses
• Machine learning
• Personalized treatment
• Chekroud