A Sparse-Attention Deep Learning Model Integrating Heterogeneous Multimodal Features for Parkinson's Disease Severity Profiling

TL;DR

This work addresses the challenge of characterising Parkinson's disease severity by integrating heterogeneous multimodal data within an interpretable deep learning framework. The authors introduce SAFN, a Class-Weighted Sparse-Attention Fusion Network that uses modality-specific encoders, symmetric cross-attention, and a sparsity-constrained fusion layer to fuse MRI morphometry, clinical assessments, and demographics. To tackle real-world class imbalance, SAFN optimizes with a Class-Balanced Focal Loss, while attention gates provide intrinsic interpretability of modality contributions and feature importance. On 703 participants from PPMI, SAFN achieves state-of-the-art performance (accuracy ≈ 0.98, PR-AUC ≈ 1.00) and identifies clinical measurements as the primary drivers of prediction, aligning with diagnostic principles. The study demonstrates robust, reproducible multimodal profiling with interpretable decision processes, offering a scalable path toward clinical decision support and guiding future work in cross-site validation and longitudinal modelling.

Abstract

Characterising the heterogeneous presentation of Parkinson's disease (PD) requires integrating biological and clinical markers within a unified predictive framework. While multimodal data provide complementary information, many existing computational models struggle with interpretability, class imbalance, or effective fusion of high-dimensional imaging and tabular clinical features. To address these limitations, we propose the Class-Weighted Sparse-Attention Fusion Network (SAFN), an interpretable deep learning framework for robust multimodal profiling. SAFN integrates MRI cortical thickness, MRI volumetric measures, clinical assessments, and demographic variables using modality-specific encoders and a symmetric cross-attention mechanism that captures nonlinear interactions between imaging and clinical representations. A sparsity-constrained attention-gating fusion layer dynamically prioritises informative modalities, while a class-balanced focal loss (beta = 0.999, gamma = 1.5) mitigates 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 plus or minus 0.02 and a PR-AUC of 1.00 plus or minus 0.00, outperforming established machine learning and deep learning baselines. Interpretability analysis shows a clinically coherent decision process, with approximately 60 percent of predictive weight assigned to clinical assessments, consistent with Movement Disorder Society diagnostic principles. SAFN provides a reproducible and transparent multimodal modelling paradigm for computational profiling of neurodegenerative disease.

Figures (8)

• Figure 1: Overall methodological workflow for Parkinson’s disease classification. Multimodal data undergo preprocessing, normalisation, and stratified sampling before model training. Machine-learning, deep-learning, and the proposed Class-Weighted SAFN models are trained under identical experimental settings and evaluated using Accuracy, Balanced Accuracy, ROC--AUC, PR--AUC, Precision, Recall, and F1-score metrics.
• Figure 2: Architecture of the proposed Class-Weighted Sparse-Attention Fusion Network for Parkinson’s disease classification. Four input modalities—MRI Cortical Thickness, Clinical, MRI Volumetric, and Demographic features—are processed through modality-specific tokenizers or MLP encoders. MRI Cortical Thickness and Clinical embeddings interact via symmetric cross-attention, followed by sparse attention–gated multimodal fusion with learnable modality weights ($\alpha_1$–$\alpha_4$). The resulting fused representation $\mathbf{H}$ is passed to a classification head to predict PD probability. Solid arrows indicate data flow, while dashed arrows denote loss-related optimisation pathways.
• Figure 3: Grouped bar chart illustrating the performance of traditional machine-learning models and deep learning models across four key evaluation metrics: Accuracy, Balanced Accuracy, ROC-AUC, and PR-AUC. Error bars represent the standard deviation across five cross-validation folds. The proposed Class-Weighted SAFN exhibits consistently high performance with minimal variance compared to baseline models.
• Figure 4: Radar chart comparing the performance of all baseline models and the proposed Class-Weighted SAFN across seven evaluation metrics (Accuracy, Balanced Accuracy, ROC-AUC, PR-AUC, Precision, Recall, and F1-score). Values correspond to mean scores over five-fold cross-validation. The SAFN model forms the outermost profile, indicating consistently superior performance.
• Figure 5: Performance visualisation of traditional ML classifiers (Logistic Regression, SVM, and Random Forest). For each model, the figure displays the averaged confusion matrix, mean precision–recall curve, and mean ROC curve, computed across 5-fold cross-validation.