Abstract
Objectives
Interpatient variability in bipolar I depression (BP-D) symptoms challenges the ability to predict pharmacotherapeutic outcomes. A machine learning workflow was developed to predict remission after 8 weeks of pharmacotherapy (total score of ≤8 on the Montgomery Åsberg Depression Rating Scale [MADRS]).
Methods
Supervised machine learning models were trained on data from BP-D patients treated with olanzapine (N = 168) and were externally validated on patients treated with olanzapine/fluoxetine combination (OFC; N = 131) and lamotrigine (LTG; N = 126). Top predictors were used to develop a prognosis rule informing how many symptoms should change and by how much within 4 weeks to increase the odds of achieving remission.
Results
An AUC of 0.76 (NIR:0.59; p = 0.17) was established to predict remission in olanzapine-treated subjects. These trained models achieved AUCs of 0.70 with OFC (NIR:0.52; p < 0.03) and 0.73 with LTG (NIR:0.52; p < 0.003), demonstrating external replication of prediction performance. Week-4 changes in four MADRS symptoms (reported sadness, reduced sleep, reduced appetite, and concentration difficulties) were top predictors of remission. Across all pharmacotherapies, three or more of these symptoms needed to improve by ≥2 points at Week-4 to have a 65% chance of achieving remission at 8 weeks (OR: 3.74, 95% CI: 2.45–5.76; p < 9.3E-11).
Conclusion
Machine learning strategies achieved cross-trial and cross-drug replication in predicting remission after 8 weeks of pharmacotherapy for BP-D. Interpretable prognoses rules required only a limited number of depressive symptoms, providing a promising foundation for developing simple quantitative decision aids that may, in the future, serve as companions to clinical judgment at the point of care.