MultimodalCNN-PD: a Parkinson's disease diagnostics framework using multimodal convolutional neural network.

Background: Parkinson's disease (PD) is a prevalent neurodegenerative disorder that severely affects motor and cognitive functions. Early diagnosis, particularly during the prodromal phase, is critical for effective intervention. Methods: This study presents MultimodalCNN-PD++, a deep learning model that integrates Magnetic Resonance Imaging (MRI) with clinical metadata (including motor/cognitive assessments, demographic data, and genetic biomarkers) to enhance PD classification. The model employs a lightweight EfficientNetB0 backbone, Mobile Convolutional Block Attention Modules (Mobile CBAM), and an enhanced Meta-Guided Cross-Attention (MGCA++) mechanism. A three-stage hierarchical feature selection method identifies the most discriminative clinical features, while metadata is processed with BioClinicalBERT using Low-Rank Adaptation (LoRA). Results: Validated on the Parkinson's Progression Markers Initiative (PPMI) dataset, the model achieved 97.5% accuracy in distinguishing Normal Control, prodromal PD, and diagnosed PD cases, with reduced parameters and computational costs. External validation on the OASIS-3 dataset confirmed robust generalizability (96.2% accuracy) despite demographic and acquisition protocol variations. Ablation studies highlighted the contributions of Mobile CBAM, MGCA++, hierarchical feature selection, and BioClinicalBERT-LoRA. Discussion: This framework sets a new benchmark for multiclass PD diagnosis, demonstrating strong potential as a clinically deployable AI tool for early detection and personalized management of neurodegenerative diseases.