Parkinson's Disease Detection Based on Running Speech Data From Phone Calls

Area of Science:

• Neurology and Computational Linguistics
• Biomedical Signal Processing
• Machine Learning for Healthcare

Background:

• Parkinson's Disease (PD) diagnosis is often delayed due to subtle early symptoms.
• Objective, longitudinal tracking of PD symptoms is needed for timely intervention.
• Speech impairment is a key indicator of PD, even in early stages.

Purpose of the Study:

• To develop and validate a privacy-aware, technology-based method for early Parkinson's Disease detection.
• To assess the feasibility of using passively captured voice data from routine phone calls for PD screening.
• To evaluate the performance of language-aware machine learning models in classifying PD patients and healthy controls.

Main Methods:

• Voice features were extracted from passively recorded speech during phone calls.
• Language-aware machine learning classifiers were trained using voice features and demographic data.
• A multilingual cohort of 498 subjects (392 healthy controls/106 PD patients) was utilized for training and validation.

Main Results:

• The best models achieved Area Under the ROC Curve (AUC) scores ranging from 0.63 to 0.83 in cross-validation across different language sub-cohorts.
• Out-of-sample testing on 63 subjects demonstrated strong performance, with AUCs of 0.84, 0.93, and 0.83 for English, Greek, and German sub-cohorts, respectively.
• The proposed approach demonstrated superior performance compared to existing language-aware PD detection methods.

Conclusions:

• The developed method offers a high-frequency, privacy-aware, and unobtrusive tool for screening Parkinson's Disease.
• Analysis of voice samples from routine phone calls is a viable method for early PD detection.
• This technology has the potential to significantly transform PD assessment and accelerate diagnosis.