Parkinson's Disease Detection from Speech Signals: Comparison of Three Different Approaches with Cross-Corpus Evaluation

Abstract

This study presents a cross-corpus evaluation of three different approaches for the detection of Parkinson's disease (PD) from speech signals. Three publicly available datasets (PC-GITA, MDVR-KCL, and mPower), each with different recording environments, languages, and speech tasks, were used, and model generalizability was assessed using a Leave-One-Dataset-Out (LODO) protocol. Three approaches were compared: (i) a Mel-spectrogram-based Hybrid CNN+ViT (EfficientNetB0 + Vision Transformer) deep learning model, (ii) a Wav2Vec 2.0-based transfer learning approach, and (iii) classical ML models (SVM, KNN, MLP, RF) trained on ViT-PCA combined with acoustic features (MFCC, delta, and delta-delta). Results were reported at both segment and subject levels, and the effect of segment duration (10, 15, 20, and 25 seconds) on performance was systematically investigated. In the S1 scenario (Test: mPower), the MLP model achieved the highest accuracy (88.91%) with 20-second segments, whereas the RF model obtained the highest AUC (0.961). The Hybrid CNN+ViT model attained an AUC of 0.935 and 85.81% accuracy with 15-second segments. In the S3 scenario (Test: PC-GITA), the Wav2Vec 2.0-based Deep Neural Network (DNN) model achieved an AUC of 0.726, outperforming all spectrogram-based approaches. The findings suggest that speech analysis may be considered for use in Parkinson’s disease screening applications; however, generalization across different speech tasks remains a significant open challenge.