Exploiting smartphone voice recording as a digital biomarker for Parkinson’s disease diagnosis

With the progresses of smart medicine, more and more easily-measured signals are becoming digital biomarkers for healthcare monitoring. This paper employs smartphone voice recordings for Parkinson’s disease (PD) diagnosis and concludes that voice recording energized by machine learning serves as a reliable digital biomarker for accurately identifying individuals with PD. Specifically, smartphone voice recording is first preprocessed, including denoising and segmenting, then features are extracted from the processed voice segments, and different feature selection methods are employed to screen the optimal feature subsets. Finally, machine learning models utilizing tailored optimal feature subsets are able to yield diagnostic results. On two different databases, the achieved average diagnosis accuracy is greater than 90% which demonstrates that smartphone voice recording is valid to be a digital biomarker for screening PD patients. This paper further analyzes the optimal feature subset for promoting PD recognition performance, and finds the alignment between acoustic analysis and data science in PD diagnosis.