A Sparse-Attention Fusion Network integrating heterogeneous multimodal features for Parkinson’s disease classification

Accurately characterising the heterogeneous presentation of Parkinson’s disease (PD) requires effective integration of neuroimaging and clinical information, yet many multimodal models suffer from limited interpretability, class imbalance, or suboptimal fusion of imaging and tabular data. We propose the Class-Weighted Sparse-Attention Fusion Network (SAFN), an interpretable deep learning framework designed for robust multimodal Parkinson’s disease classification. SAFN integrates MRI-derived quantitative neuroimaging biomarkers (cortical thickness and volumetry), clinical assessments, and demographic variables using modality-specific encoders and a symmetric cross-attention mechanism to model non-linear interactions between imaging-derived and clinical signal representations. A sparsity-constrained attention-gating fusion layer adaptively prioritises informative modalities, while a Class-Balanced Focal Loss ( β = 0.999 , γ = 1.5 ) addresses dataset imbalance without synthetic oversampling. Evaluated on 703 participants (570 PD, 133 healthy controls) from the Parkinson’s Progression Markers Initiative using subject-wise five-fold cross-validation, SAFN achieves an accuracy of 0.98 ± 0.02 and a PR–AUC of 1.00 ± 0.00, achieving competitive performance relative to established machine-learning and deep-learning baselines. Interpretability analysis reveals clinically coherent attribution patterns, with modality gating assigning approximately 60% of the fused representation to clinical assessments, consistent with Movement Disorder Society diagnostic principles, while MRI-derived biomarkers provide complementary structural information within the multimodal fusion framework. Attribution estimates were validated against SHAP and permutation-based analysis, which independently identified clinical assessments as the dominant modality, and cross-fold ranking stability was quantified explicitly. These results demonstrate SAFN as an interpretable and robust multimodal fusion framework for Parkinson’s disease classification within a harmonised research cohort setting and indicate that further external, prospective, and usability-focused validation is required before considering clinical decision-support use.