Diagnosis of Parkinson's Disease Based on Voice and Speech Using Machine Learning and Deep Learning: A Systematic...

Early diagnosis of Parkinson's disease (PD) is complicated. Speech impairment, as an early symptom of PD, offers a noninvasive, scalable biomarker for remote assessment. Speech-based machine learning has shown promise, but methodological quality of existing evidence remains unclear. This review examines the previous reviews on the performance of machine learning models, their strengths and weaknesses, and future research opportunities in diagnosing PD using speech.

We conducted this umbrella review in accordance with PRISMA guidelines, systematically searching PubMed, Scopus, Web of Science, IEEE, and Cochrane Library up to April 2025, using keywords related to PD, machine learning, deep learning, and voice and speech data. Two reviewers independently performed screening and data extraction. Methodological quality and risk of bias were assessed using AMSTAR-2 and ROBIS. Of 233 records screened, 22 reviews were ultimately included.

The most common algorithms evaluated were Support Vector Machine, Artificial Neural Network, K-Nearest Neighbors, Convolutional Neural Network, and Deep Neural Network. Despite reported performance ranges from modest to excellent (AUC 0.5-1.0; F1 score 0.690-0.997), the methodological quality of the reviews was generally low. No meta-analyses were conducted, and most studies insufficiently reported critical clinical factors such as patient medication state, language, disease severity, or recording conditions. Few reviews addressed external validation, calibration, model fairness, or clinical interpretability, elements essential for translation into clinical practice.

Speech-based machine learning holds strong potential as a low-cost, noninvasive, and clinically scalable tool for early PD detection. However, the current evidence is fragmented and not yet robust enough to support clinical adoption.