Parkinson’s disease classification and prediction via adaptive sparse learning from multiple modalities

Abstract

Parkinson’s disease (PD) is a prevalent neurodegenerative disorder where early clinical diagnosis is critical for patients. In this study, we introduce a new embedded adaptive sparse learning method that integrates multimodal data, labeling, and clinical score information for early PD diagnosis. Based on the manifold learning theory, the proposed method utilizes a low-dimensional manifold to represent the structure of high-dimensional data. We carry out similarity learning and feature selection from different modalities while using the $l_{2,p}$ specification for adaptive sparsity control. Also, we dynamically learn the similarity between features to find more effective information in features. an effective optimization iterative algorithm is proposed to solve this problem. To validate the effectiveness of the proposed method, we conducted extensive experiments on the Parkinson’s Progression Markers Initiative (PPMI) database, including PD vs. normal controls (NC), scans without evidence of dopaminergic deficit (SWEDD) vs. NC, PD vs. SWEDD, and PD vs. SWEDD vs. NC. Using baseline data for experiments, our method achieved accuracies of 82.38%, 85.56%, 86.47%, and 65.54%, respectively. When using 12-month data, the accuracies were 77.43%, 95.26%, 95.35%, and 77.98%, respectively. Overall, our method outperformed the other methods. Additionally, the performance of the three deep models using the feature subsets selected by our method surpassed that achieved using the original data, further validating the effectiveness of our method.