Objectives: Lithium and quetiapine are known to be effective treatments for bipolar disorder. However, little information is available to inform prediction of response to these medications. Machine-learning methods can identify predictors of response by examining variables simultaneously. Further evaluation of models on a test sample can estimate how well these models would generalize to other samples.
Methods: Data (N = 482) were drawn from a randomized clinical trial of outpatients with bipolar I or II disorder who received adjunctive personalized treatment plus either lithium or quetiapine. Elastic net regularization (ENR) was used to generate models for lithium and quetiapine; these models were evaluated on a test set.
Results: Predictions from the lithium model explained 17.4% of the variance in actual observed scores of patients who received lithium in the test set, while predictions from the quetiapine model explained 32.1% of the variance of patients that received quetiapine. Of the baseline variables selected, those with the largest parameter estimates were: severity of mania; attention-deficit/hyperactivity disorder (ADHD) comorbidity; nonsuicidal self-injurious behavior; employment; and comorbidity with each of two anxiety disorders (social phobia/society anxiety and agoraphobia). Predictive accuracy of the ENR model outperformed the simple and basic theoretical models.
Conclusion: ENR is an effective approach for building optimal and generalizable models. Variables identified through this methodology can inform future research on predictors of response to lithium and quetiapine, as well as future modeling efforts of treatment choice in bipolar disorder.
Keywords: bipolar disorder; lithium; machine learning; predictive modeling; quetiapine.
© 2019 John Wiley & Sons A/S. Published by John Wiley & Sons Ltd.