Machine learning with multiple modalities of brain magnetic resonance imaging data to identify the presence of bipolar disorder

Abstract

BackgroundBipolar disorder (BD) is a chronic psychiatric mood disorder that is solely diagnosed based on clinical symptoms. These symptoms often overlap with other psychiatric disorders. Efforts to use machine learning (ML) to create predictive models for BD based on data from brain imaging are expanding but have often been limited using only a single modality and the exclusion of the cerebellum, which may be relevant in BD. MethodsIn this study, we sought to improve ML classification of BD by combining information from structural, functional, and diffusion-weighted imaging. Participants (108 BD I, 78 control) with BD type I and matched controls were recruited into an imaging study. This dataset was randomly divided into training and testing sets. For each of the three modalities, a separate ML model was selected, trained, and then used to generate a prediction of the class of each test subject. Majority voting was used to combine results from the three models to make a final prediction of whether a subject had BD. An independent replication sample was used to evaluate the ability of the ML classification to generalize to data collected at other sites. ResultsCombining the three machine learning models through majority voting resulted in an accuracy of 89.5 % for classification of the test subjects as being in the BD or control group. Bootstrapping resulted in a 95 % confidence interval of 78.9 %–97.4 % for test accuracy. Performance was reduced when only using 2 of the 3 modalities. Analysis of feature importance revealed that the cerebellum and nodes of the emotional control network were among the most important regions for classification. The machine learning model performed at chance on the independent replication sample. ConclusionBD I could be identified with high accuracy in our relatively small sample by combining structural, functional, and diffusion-weighted imaging data within a single site but not generalize well to an independent replication sample. Future studies using harmonized imaging protocols may facilitate generalization of ML models.