AI-005 - PREDICTING REMISSION IN ACUTE BIPOLAR DEPRESSION: A MACHINE LEARNING STUDY WITH CROSS-TRIAL REPLICATION

J. Marrero-Polanco¹, J. Joyce¹, C. Grant², P. Croarkin¹, A. Athreya¹, W. Bobo¹; ¹Mayo Clinic, ²Mayo Clinic, Rochester, MN, USA.

Background: Variability in depressive symptom response in patients with bipolar I depression (BD-D) makes the predictability of outcomes to acute pharmacotherapy challenging. This study hypothesized that machine learning (ML) models integrating patient clinical features can predict remission after acute BD-D pharmacotherapy.

Methods: A ML workflow utilizing XGBoost or penalized regression algorithms integrated sociodemographic variables as well as baseline and 4-week changes in depressive symptoms (measured using Montgomery Åsberg Depression Rating Scale – MADRS) was developed to predict remission to BD-D pharmacotherapy at 8 weeks. The workflow was trained on data from BD-D patients treated with olanzapine (OLZ, n = 168) and was validated on independent cohorts treated with olanzapine/fluoxetine combination (OFC, n = 131) or lamotrigine (LMG; n = 126). As 53.3% is the average non-remission rate in these drugs (null information rate [NIR]), the prediction performance significance was assessed by comparing the ML derived accuracy to the NIR, which serves as a proxy for chance.

Results: Models trained on only clinical features from the OLZ cohort predicted remission in the holdout test set with a mean [95% CI] AUC of 0.76 [0.66-0.83] (NIR: 0.59; P = 0.15 [0.01, 0.36]). During cross-trial replication, the model predicted remission in OFC and LMG cohorts with AUCs of 0.70 [0.68-0.72] (NIR: 0.52; P < 0.03 [0.001-0.05]) and 0.73 [0.71-0.76] (NIR: 0.52; p < 0.003 [0.0003-0.005]), respectively. Four MADRS symptom week-4 changes were identified as top predictors of remission at 8 weeks, independent of the change in the MADRS total score. These prognostic symptoms were reported sadness, reduced sleep, reduced appetite, and concentration difficulties.

Conclusion: This work demonstrated the ability of the ML workflow to predict remission from depressive symptoms in patients with BD-D with cross-trial and cross-drug replications. The accuracies of these models were significantly better than the NIR, providing a basis for future validations in prospective studies. Future work that incorporates biological features (e.g., -omics, imaging) holds promise for improved predictability of remission in BD-D patients treated with antipsychotics, mood stabilizers, or combinations like antipsychotic with antidepressant.