Machine learning for the diagnosis of Parkinson’s disease using speech analysis: a systematic review

As voice technology continues to advance, clinical analysis of acoustic measures in speech patterns has the potential for early detection of Parkinson’s disease (PD), the second most prevalent neurodegenerative disorder for humans globally. Often returning inconsistent findings when examined by a specialist, voice has been demonstrated to be a vital biomarker in diagnosis. Many models of early onset PD in humans have been developed based on speech patterns. We conducted a search on three databases and included studies that evaluated the performance of machine learning algorithms in predicting the Parkinson’s disease through speech analysis via area under the receiver operating characteristic curve (ROC AUC) values. Studies were manually assessed for risks of selection bias, performance bias, and of the eight studies selected, the model with the highest number of samples and best ROC AUC of 92.4% was Gradient Boosted. Whereas the model with the lowest number of samples had the lowest ROC AUC of 50% for all models in the study, including Bayes Net, Zero R, Naïve Bayes, and Random Forest. Gradient Boosted models seemed to have higher accuracy of predictions in most studies. We determined that some machine learning algorithms can accurately predict the Parkinson’s disease using voice data on specific datasets, showing potential for use in clinical settings. This allows for earlier detection of Parkinson’s disease and improvement of clinical decisions.