Research on Passive Assessment of Parkinson’s Disease Utilising Speech Biomarkers

Abstract

Speech disorders, collectively referred to as hypokinetic dysarthria (HD), are early biomarkers of Parkinson’s disease (PD). To assess all dimensions of HD, patients could perform several speech tasks using a smartphone outside a clinic. This paper aims to adapt the parametrization process to running speech so that a patient is not required to interact actively with the device, and features can be extracted directly from phone calls. The method utilizes a voice activity detector followed by a voicing detection. The algorithm was tested on a database of 126 recordings (86 patients with PD and 40 healthy controls) of monologue mixed with noise with different signal-to-noise ratios (SNR) to simulate the real environment conditions. Pearson correlation coefficients show a strong linear relationship between speech features and patients’ scores assessing HD and other motor/non-motor symptoms – p-value < 0.01 for the normalized amplitude quotient (NAQ) with Test 3F Dysarthric Profile (DX index) and Unified Parkinson’s Disease Rating Scale (part III) in 20 dB SNR conditions, p-value < 0.01 for the jitter and shimmer with the Mini Mental State Exam (10 dB SNR). A model based on the Extreme Gradient Boosting algorithm predicts the DX index with a 10.83% estimated error rate (EER) and the Addenbrooke’s Cognitive Examination-Revise (ACE-R) score with 13.38% EER. The introduced algorithm can potentially be used in mHealth applications for passive monitoring and assessment of PD patients.

Speech disorders, collectively referred to as hypokinetic dysarthria (HD), are early biomarkers of Parkinson’s disease (PD). To assess all dimensions of HD, patients could perform several speech tasks using a smartphone outside a clinic. This paper aims to adapt the parametrization process to running speech so that a patient is not required to interact actively with the device, and features can be extracted directly from phone calls. The method utilizes a voice activity detector followed by a voicing detection. The algorithm was tested on a database of 126 recordings (86 patients with PD and 40 healthy controls) of monologue mixed with noise with different signal-to-noise ratios (SNR) to simulate the real environment conditions. Pearson correlation coefficients show a strong linear relationship between speech features and patients’ scores assessing HD and other motor/non-motor symptoms – p-value < 0.01 for the normalized amplitude quotient (NAQ) with Test 3F Dysarthric Profile (DX index) and Unified Parkinson’s Disease Rating Scale (part III) in 20 dB SNR conditions, p-value < 0.01 for the jitter and shimmer with the Mini Mental State Exam (10 dB SNR). A model based on the Extreme Gradient Boosting algorithm predicts the DX index with a 10.83% estimated error rate (EER) and the Addenbrooke’s Cognitive Examination-Revise (ACE-R) score with 13.38% EER. The introduced algorithm can potentially be used in mHealth applications for passive monitoring and assessment of PD patients.