Using machine learning to diagnose Parkinson's disease from voice recordings

摘要

Parkinson's Disease (PD) is a debilitating neurodegenerative disease which cannot be diagnosed through standardized blood tests, so a faster, cheaper diagnostic tool is essential. Using machine learning algorithms to analyze the variations in voice patterns is a novel method of predicting the existence of PD in patients. This paper proposes a predictive model that effectively diagnoses PD with maximum accuracy using a dataset that consists of extrapolated data from voice recordings of Parkinson's patients and unaffected subjects. The results of experimental testing showed that a Boosted Decision Tree, which is an ensemble model made from gradient boosted regression trees, was the best model to use on the data, with an accuracy score of 91-95%. It was also discovered through filter-based feature detection that the strongest weighted features were spreadl, spread2, and PPE, all three nonlinear measures of fundamental frequency variation in the voice recordings. These findings can be applied to PD, other motor disorders, or even vocal biometrics.