Comprehensive multimodal approach for Parkinson’s disease classification using artificial intelligence: insights and model explainability

• Published in: Soft computing (Berlin, Germany) Vol. 29; no. 3; pp. 1845 - 1877
• Main Authors: Balaha, Hossam Magdy, Hassan, Asmaa El-Sayed, Ahmed, Rawan Ayman, Balaha, Magdy Hassan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.02.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Classification; Computation; Computational Intelligence; Control; Datasets; Deep learning; Disease; Dopamine; Effectiveness; Engineering; Ensemble learning; Fatigue; Feature selection; Machine learning; Mathematical Logic and Foundations; Mechatronics; Neural Networks; Optimization; Parkinson's disease; Performance evaluation; Performance measurement; Recall; Robotics; Computer-aided diagnosis (CAD); Parkinson’s disease (PD); eXplainable AI; Deep learning (DL); Vision transformers (ViT)
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-025-10463-9

Parkinson’s disease (PD) is a debilitating neurodegenerative disorder affecting millions worldwide. Early detection is vital for effective management, yet remains challenging. In this study, we investigated four distinct datasets for PD detection. Through comprehensive experimentation employing ensemble methods and feature selection, we achieved high classification accuracies across the datasets. For the Oxford Parkinson’s Disease Detection Dataset, an accuracy of 95.67%, precision of 97.59%, recall of 84.5%, specificity of 99.32%, and F1-score of 90.57% were achieved. For the Alzheimer Parkinson Diseases 3 Class Dataset, the “Stacking” approach surpasses individual models, reaching an accuracy of 99.85%, precision of 99.81%, recall of 99.81%, specificity of 99.86%, and F1 of 99.81%. For the NewHandPD dataset, Regarding the Spiral category, The “Base-P32-384” model surpasses others with an accuracy of 97.35%, precision of 96.50%, recall of 98.57%, and F1-score of 97.53%. The collective “Stacking” approach proves highly effective regarding the Circle category, achieving 100% across all performance metrics. Regarding the Meander category, the “Base-P16-224” model achieves an accuracy of 97.35%, precision of 99.26%, recall of 95.71%, specificity of 99.19%, and F1 of 97.45%. The Mobile Device Voice Recordings at King’s College London (MDVR-KCL) dataset contains two datasets. Regarding the “SpontaneousDialogue” dataset, accuracy, BAC, precision, recall, specificity, and F1-score were computed, resulting in values of 94.03%, 92.83%, 90.78%, 100.0%, and 85.67%, respectively. Regarding the “ReadText” dataset, accuracy, BAC, precision, recall, specificity, and F1-score were computed, resulting in values of 91.89%, 90.62%, 87.5%, 100.0%, and 81.25%, respectively. Our findings highlight the efficacy of leveraging diverse data sources and advanced machine learning techniques to enhance PD detection accuracy.