A Machine Learning with Ordinal Ranking Approach for Dysarthria Severity Classification in Parkinson’s Disease

Purpose: This study proposes a simple Support Vector Machine (SVM) classifier with ordinal ranking to improve classification of Parkinson’s disease (PD) dysarthria severity across four levels: normal, mild, moderate, and severe. Methods: Forty Mandarin-speaking participants were recruited from National Cheng Kung University Hospital, Taiwan. A systematic selection of prosodic, glottal, phonetic, and articulatory features was performed to capture severity-related speech patterns. Four approaches were compared, including SVM multiclass, Support Vector Regression (SVR), Deep Neural Networks (DNN), and a Large Language Model (LLM) with LoRA. Results: The proposed SVM with ordinal ranking model achieved 100% accuracy in distinguishing healthy individuals from those with PD dysarthria and 75% accuracy in Leave-One-Subject-Out (LOSO) cross-validation for four severity levels, with a Root Mean Square Error (RMSE) of 0.613. Both SVM and DNN models integrated with ordinal ranking outperformed their counterparts without ordinal ranking, advancing the state-of-the-art in multiclass severity classification. Among all tested methods, SVM with ordinal ranking achieved the highest accuracy and lowest error rate. Conclusion: A systematic selection of speech features was used to capture severity-related patterns, and incorporating ordinal ranking into SVM enables higher accuracy and robust severity classification than conventional approaches. This approach surpasses more complex deep learning models, offering higher accuracy and simple implementation on small datasets. These characteristics make it highly suitable for clinical diagnosis and as a foundation for personalized home-based rehabilitation in PD dysarthria.