Predictive Models in the Diagnosis of Parkinson’s Disease Through Voice Analysis

Parkinson’s disease (PD) is a chronic and progressive long-term degenerative disorder of the central nervous system. The symptoms usually develop slowly, starting with symptoms in the motor system and potentially evolving into non-motor symptoms as the disease worsens. It is estimated that 90% of patients with Parkinson’s end up having speech disorders. The following work delves into the domain of classification models for Parkinson’s disease through speech recordings using machine learning (ML). Diverse datasets were explored including (1) the Italian Parkinson’s Voice and Speech; (2) Mobile Device Voice Recordings at King’s College London; (3) Synthetic Vowels of Speakers with Parkinson’s Disease and Parkinsonism; and (4) Voice Samples for Patients with Parkinson’s Disease and Healthy Controls. Two types of feature extraction are used, incorporating traditional acoustic features (interpretable) and embeddings extracted using pre-trained Deep Neural Networks (non-interpretable): (1) TRILLsson; (2) Wav2Vec 2.0; (3) HuBERT. For the training algorithms, a myriad of methods were used, from simple logistic regression to XGBoost. The two types of features adopted displayed robustness, as the use of interpretable features reached up to 95% accuracy, while non-interpretable embeddings had a top-performing accuracy of 99%, surpassing the state-of-the-art.