Discriminating Relapsing-Remitting From Progressive Multiple Sclerosis Using Multidimensional Speech Biomarkers: An...

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Differentiating relapsing-remitting from progressive multiple sclerosis (MS) is clinically challenging.
• Current methods often rely on complex or invasive biomarkers.
• Objective phenotyping of MS subtypes is needed.

Purpose of the Study:

• To develop a machine learning model using acoustic speech analysis for classifying MS phenotypes.
• To create a non-invasive, low-cost tool for MS phenotyping.

Main Methods:

• Collected speech recordings from 300 people with MS (PwMS).
• Extracted 44 acoustic features and used elastic net for parameter selection.
• Trained and validated six supervised machine learning classifiers, including Random Forest, on acoustic features and demographic data.
• Tested the best model on an independent clinical validation cohort (n=100).

Main Results:

• Elastic net identified nine acoustic features, age, and sex as relevant parameters.
• Acoustic-only model (excluding age and sex) showed no performance degradation.
• The Random Forest model achieved an AUC of 0.84, NPV of 0.95, and LR- of 0.15 in the validation cohort.
• The model effectively ruled out progressive MS using acoustic features alone.

Conclusions:

• Multidimensional acoustic analysis combined with machine learning provides a viable non-invasive triage tool for MS phenotyping.
• The developed model is accessible, low-cost, and validated in an independent cohort.
• This approach supports objective and efficient MS phenotyping.