LLM-guided multimodal attention network for robust multiclass Parkinson’s disease diagnosis

Abstract

Objectives: Accurate identification of Parkinson’s disease (PD), particularly during its prodromal stage, remains a major clinical challenge due to heterogeneous symptom presentation and overlapping neurological patterns. This study proposes an LLM-Guided Multimodal Attention Network (LLM-MAN) to improve PD staging by jointly modeling structural MRI and clinical/cognitive metadata. Methods: We develop a unified multimodal framework that encodes structural MRI using a ResNet-18 backbone enhanced with Convolutional Block Attention Modules (CBAM) for discriminative neuroimaging feature extraction, and represents clinical/cognitive metadata using an LLM-based text encoder (pre-trained BERT) for contextualized semantic modeling. A Meta-Guided Cross-Attention (MGCA) module is introduced to align clinical semantic knowledge with imaging features, enabling robust cross-modal fusion for multiclass classification (Normal Control, prodromal PD, and diagnosed PD). The model is evaluated on the Parkinson’s Progression Markers Initiative (PPMI) dataset and further validated on an independent external cohort. Results: On the PPMI dataset, LLM-MAN achieved an accuracy of 95.68 % for distinguishing Normal Control, prodromal PD, and diagnosed PD. External validation on an independent cohort yielded 94.10 % accuracy, indicating strong generalization performance across datasets. Conclusions: LLM-guided multimodal fusion via MGCA provides reliable and interpretable approach for PD staging, substantially improving prodromal PD identification by integrating semantic clinical knowledge with neuroimaging representations.