Digital speech assessments and machine learning for differentiation of neurodegenerative diseases

Abstract

Introduction

Speech impairment is a prevalent symptom of neurological disorders, including Parkinson’s disease (PD), Progressive Supranuclear Palsy (PSP), Huntington’s disease (HD), and Amyotrophic Lateral Sclerosis (ALS), with mechanisms and severity varying across and within conditions. Scalable digital health tools and machine learning (ML) are essential for diagnosing and tracking neurodegenerative disease.

Methods

A total of 92 individuals were included in this study (21 PSP, 21 PD, 18 HD, 15 ALS, and 16 healthy elderly controls (CTR)). The Rainbow Passage was collected on a digital device and analyzed to extract 12 speech features representing speech production. A set of Elastic Net ML models was trained on these speech features to differentiate between diagnostic classes. A specialized Support Vector Machine ML model was then developed to differentiate PSP from PD.

Results

Elastic Net models achieved a balanced accuracy of 77% over 5 diagnostic classes (group-specific sensitivities of 76% for PSP, 67% for PD, 83% for HD, 73% for ALS, and 88% for CTR) and 83% over 4 diagnostic classes (group-specific sensitivities of 83% for PSP-PD, 83% for HD, 73% for ALS, and 94% for CTR). The PSP vs. PD classification model demonstrated a balanced accuracy of 85%, with sensitivity of 88% for PSP and 82% for PD. Key speech features differentiated clinical conditions, with Total Voiced Time being the strongest positive feature for combined PSP-PD. In HD, ALS, and CTR, Ratio Extra Words, Pauses per Second, and Intelligibility were the most strongly differentiating features, respectively. Articulatory Rate emerged as the most distinguishing feature between PD and PSP.

Conclusion

Our findings highlight the potential of digital health technology and ML in identifying and monitoring speech features in neurodegenerative diseases.