Abstract
AIM: Most psychiatric and neurodevelopmental disorders are heterogeneous. Neural abnormalities in patients might differ in magnitude and kind, giving rise to distinct subtypes that can be partly overlapping (comorbidity). Identifying disorder-related individual differences is challenging due to the overwhelming presence of disorder-unrelated variation shared with healthy controls. Recently, Contrastive Variational Autoencoders (CVAEs) have been shown to separate disorder-related individual variation from disorder-unrelated variation. However, it is not known if CVAEs can also satisfy the other key desiderata for psychiatric research: capturing disease subtypes and disentangling comorbidity. In this paper, we compare CVAEs to other methods as a function of hyperparameters, such as model size and training data availability. We also introduce a new architecture for modeling comorbid disorders and test a novel training procedure for CVAEs that improves their reproducibility. METHODS: We use synthetic neuroanatomical MRI data with known ground truth for shared and disorder-specific effects and study the performance of the CVAE and non-contrastive baseline models at detecting disorder-subtypes and disentangling comorbidity in brain images varying along shared and disorder-specific dimensions. RESULTS: CVAE models consistently outperformed non-contrastive alternatives as measured by correlation with disorder-specific ground truth effects and accuracy of subtype discovery. The CVAE also successfully disentangled neuroanatomical loci of comorbid disorders, due to its novel architecture. Improved training procedure reduced variability in the results by up to 5.5×. CONCLUSION: The results showcase how the CVAE can be used as an overall framework in precision psychiatry studies, enabling reliable detection of interpretable neuromarkers, discovering disorder subtypes and disentangling comorbidity.