Title : Early detection of amyotrophic lateral sclerosis through acoustic speech biomarkers
Abstract:
Amyotrophic Lateral Sclerosis (ALS) is a fatal neurodegenerative disease that has an average diagnostic delay of 10-18 months after symptom onset. Speech impairment affects approximately 80% of patients and may precede clinical recognition. This makes this early acoustic analysis a low cost way to aid in diagnosis. The study applied an interpretable machine learning pipeline to complete the VOC-ALS database. This is a publicly available database of smart phone recorded voice signals from 102 ALS patients and 51 healthy controls. The acoustic features are F0 mean, SD, jitter, shimmer, and harmonic to noise ratio. This has been extracted using Praat in eight vocal tasks. Ten fold stratified cross validation gave an AUC of 0.712 in gradient boost, 0.706 in random forest, 0.692 SVM , and 0.63 Logistic Regression. Jitter, shimmer, HNR, and F0 variability during syllable repetition tasks (/ka/, /ta/, /pa/) emerged as the most diagnostically informative for ALS. These results suggest that interpretable, smartphone acoustic recordings for ALS can achieve clinically meaningful AUC of greater than 0.70. With advances in pause tracking and with better recording quality this can be informative for bulbar dysfunction detection through a clinical setting or via phone applications. This paves the way for at home ALS testing which can greatly benefit those that are taking precautions to make sure that they are safe. If anyone is found to have ALS they can take the medications as early as possible in order to improve their quality of life and slow the progression of ALS.
