Empirical Analysis of Acoustic Parameters of Parkinson’s Subjects and Predicting Health Status with Machine Learning Approach
Objective:
It aims to demonstrate the salient changes in vocal features in PD patients by applying specific data visualization techniques in the data set of 195 distinct voice recordings and further propose a relatively more suitable ML method for accurately diagnosing PD.
Background:
Parkinson’s Disease (PD), a neurodegenerative disease, generally affects motor skills and cognition abilities. Diagnosing early-onset PD is difficult due to the lack of definitive laboratory tests. Recent studies have shown that voice parameters can be crucial biomarkers for PD detection at early stages. Implementing Machine Learning (ML) algorithms to voice recordings of a group of individuals can help isolate the PD-affected from the healthy controls, obviating the necessity of repeated clinic visits for observing motor symptoms.
Design/Methods:
A data set (collected from UCI ML repository) of values of voice parameters from 195 voice recordings have been used here. Necessary data preparation steps have been performed using Python and relevant libraries such as Pandas and Numpy. Specific Python libraries, such as Matplotlib and Seaborn have been used to create visualizations including heatmaps, scatterplots, etc. KNN, SVC, Logistic Regression, etc. ML techniques have been used for diagnostic predictions.
Results:
The results strongly exhibit that certain acoustic parameters, such as NHR, jitter and shimmer are higher in PD patients. PD patients also show decreased fundamental frequency and HNR range. RPDE, D2, PPE, Spread1 and Spread2 measurements vary more with relatively higher values in PD patients. Random Forest Classifier proves to be the most feasible ML approach for classification with an overall accuracy score of 97%.
Conclusions:
In the paper, the limitations of ML-based PD detection have also been mentioned. Thus, it contributes to the ongoing research on assessing the potential of a Machine Learning-based screening test for diagnosing Parkinson’s Disease before monitoring symptoms in clinical trials.