Diffusion modeling for characterizing neuropathology in psychotic disorders: a Machine Learning study

Psychotic spectrum disorders exhibit overlapping symptoms and heterogeneous neuropathology, complicating subtype differentiation. Conventional diffusion MRI models may conflate concurrent microstructural alterations. This study investigated whether integrating complementary diffusion compartments improves discrimination within the psychosis spectrum. A framework combining Diffusion Tensor Imaging (DTI), Diffusion Kurtosis Imaging (DKI), Free Water Imaging (FWI), and Intravoxel Incoherent Motion (IVIM) was evaluated using multi-shell diffusion MRI data from 1,267 Human Connectome Project participants (1,094 healthy controls; 173 patients). Parameter recovery was assessed through simulations, while regional diffusion features were analyzed using neural networks with nested, site-stratified five-fold cross-validation, ablation, scanner harmonization, and confound controls. The integrated framework distinguished patients from controls with a pooled out-of-fold AUC-ROC of 0.968 (95% CI: 0.952–0.981), compared with 0.896–0.929 for reduced models. Conventional DTI and DKI models showed near-chance discrimination between schizophrenia-spectrum and non-schizophrenia psychosis (AUC-ROC = 0.543–0.564), whereas the integrated framework achieved 0.716 (95% CI: 0.635–0.792), significantly outperforming DTI and DTI+DKI (paired bootstrap p = 0.011 and 0.007, respectively). Ablation demonstrated complementary contributions from extracellular free-water and perfusion metrics beyond conventional diffusion measurements. Same-cohort matched analysis retained strong patient-control discrimination (AUC-ROC = 0.958), although acquisition confounding remains a limitation. These findings establish the incremental predictive value of multi-compartment diffusion integration for distinguishing psychotic subtypes, providing a foundation for microstructural profiling and patient stratification, pending independent validation.