Multimodal CNN-PD: A Parkinson's Disease Diagnostics Framework Using Multimodal Convolutional Neural Network

As a common neurodegenerative condition, Parkinson's disease markedly impairs motor abilities, cognitive function, and quality of life. Establishing an early and precise diagnosis, even in its prodromal phase, is a key clinical focus to enable timely treatment and enhance patient management. Deep learning (DL) and machine learning (ML) have demonstrated significant potential in boosting diagnostic accuracy for Parkinson's disease (PD). Nevertheless, the scarcity of large-scale, well-annotated datasets remains a major obstacle, underscoring the critical need for multimodal data integration to enhance model robustness and generalizability. This study proposes MultimodalCNN-PD, a novel multimodal convolutional neural network framework, to improve PD classification accuracy. The model integrates Magnetic Resonance Imaging data with clinical metadata, including motor and cognitive assessments, demographics, and genetic biomarkers. The architecture incorporates Convolutional Block Attention Modules within a ResNet-18 backbone to refine spatial and channel-wise features and introduces a Meta Guided Cross Attention (MGCA) mechanism to align imaging and metadata through multi-head attention. An ensemble-based feature selection strategy further extracts the most discriminative clinical features. The model was robustly evaluated on the Parkinson’s Progression Markers Initiative dataset, employing a subject-level five-fold cross-validation scheme and a held-out test set. It achieved a multiclass classification accuracy of 95.68% in distinguishing Normal Control, prodromal PD, and diagnosed PD, outperforming existing state-of-the-art models. Its strong generalizability was confirmed through external validation on the OASIS-3 dataset, where it attained 94.10% accuracy despite differences in demographics and data acquisition. Ablation studies verified the performance contributions of the CBAM, MGCA, and ensemble feature selection modules. By effectively integrating neuroimaging and clinical metadata, this framework sets a new benchmark for multiclass PD diagnosis and demonstrates considerable potential as a robust, clinically applicable AI tool for early detection and personalized management of neurodegenerative diseases.