Optimizing the Prediction of Depression Remission: A Longitudinal Machine Learning Approach.

Abstract

Decisions about when to change antidepressant treatment are complex and benefit from accurate prediction of treatment outcome. Prognostic accuracy can be enhanced by incorporating repeated assessments of symptom severity collected during treatment. Participants (n = 714) from the Genome-Based Therapeutic Drugs for Depression study received escitalopram or nortriptyline over 12 weeks. Remission was defined as scoring ≤ 7 on the Hamilton Rating Scale. Predictors included demographic, clinical, and genetic variables (at 0 weeks) and measures of symptom severity (at 0, 2, 4, and 6 weeks). Longitudinal descriptors extracted with growth curves and topological data analysis were used to inform prediction of remission. Repeated assessments produced gradual and drug-specific improvements in predictive performance. By Week 4, models' discrimination in all samples reached levels that might usefully inform treatment decisions (area under the receiver operating curve (AUC) = 0.777 for nortriptyline; AUC = 0.807 for escitalopram; AUC = 0.794 for combined sample). Decisions around switching or modifying treatments for depression can be informed by repeated symptom assessments collected during treatment, but not until 4 weeks after the start of treatment.