Comorbidity-Informed Transfer Learning for Neuro-developmental Disorder Diagnosis

Neurodevelopmental disorders (NDDs) are characterized by impairments in cognition, communication, behavior, and adaptive functioning, placing increasing strain on healthcare resources for neuroimaging-based diagnosis. While deep learning-based computer-aided diagnosis (CAD) systems offer potential solutions, the complex spatiotemporal dynamics inherent in functional magnetic resonance imaging (fMRI) data render feature representations highly susceptible to noise and confounding factors, significantly limiting diagnostic efficacy. Here, we propose the comorbidity-aware transfer learning (CATL) framework for NDD diagnosis using fMRI. CATL introduces an enhanced representation generation network that integrates transfer learning with pseudo-labeling to disentangle task-relevant temporal features from confounding patterns in fMRI data, followed by feature reconstruction via an encoder-decoder architecture. These refined representations are then classified using a lightweight convolutional neural network (CNN) to build the CAD model. Critically, CATL explicitly models shared neurobiological pathways underlying NDD comorbidities through a unified semi-supervised transfer learning paradigm, providing novel mechanistic insights into cross-disorder neural dynamics. On benchmark datasets, CATL achieves mean diagnostic accuracies of 77.34% for autism spectrum disorder (ASD) and 73.28% for attention-deficit/hyperactivity disorder (ADHD), outperforming state-of-the-art transfer learning methods by 7.86% (ASD) and 0.66% (ADHD).