Overall approach

The main analysis process had six steps in this study. First, subject-level morphological (gray matter volume, GMV) differences were derived using the normative model. Second, individualized GMV differences were parsed into latent disease factors using non-negative matrix factorization (NMF). As NMF needed a predefined number of latent factors, we also proposed a strategy to automatically determine the optimal number (K) of latent disease factors. Third, the association between the identified disease factors with clinical characteristics was examined. Forth, robust analysis. Fifth, the association between subject-level differences and group-level differences obtained using the traditional case–control approach investigated. Finally, performing clustering analysis to identify OCD subtypes. More details were described in the corresponding steps below (Fig. 1).

A) Illusion of the dimensional approach where a patient is expressed as a combination of potential disease factors. B) The flow chart of our framework. The individualized GMV differences are derived from the normative model and then parsed into latent disease factors with the optimal number of latent disease factors (K). Then, we associate factor compositions (weights) with clinical characteristics. Robust analysis is done to examine reproducibility of our results. We investigate the association between individualized differences and group-level differences using a regression model. Finally, clustering analysis is performed to reveal OCD subtypes.