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

Differentiating relapsing-remitting from progressive multiple sclerosis (MS) remains a clinical challenge, often requiring complex or invasive biomarkers. This study aimed to develop a machine learning model using acoustic speech analysis to classify MS phenotypes.

A cohort of 300 consecutive people with MS (PwMS) provided speech recordings. Forty-four acoustic features were extracted. An elastic net model first identified the most relevant parameters from the development cohort (n = 200), which were then used to train six supervised ML classifiers. During the variable selection and ML model training phase, the development sample was divided into a 75/25 split, employing cross-validation to minimize overfitting. The best-performing model was subsequently tested in an independent, temporally unseen clinical validation cohort (n = 100).

Elastic net retained 11 parameters: nine acoustic features together with age and sex. Ablation analysis confirmed that removing demographic variables did not impair performance, yielding a strictly acoustic, bias‑free model. While logistic regression achieved the highest metrics during development, the Random Forest model demonstrated superior generalization in the independent, unseen validation cohort (n = 100). Relying solely on the nine acoustic features, the Random Forest achieved a robust area under the curve (AUC) of 0.84, an exceptional negative predictive value (NPV) of 0.95, and an excellent negative likelihood ratio (LR- = 0.15), effectively ruling out progressive MS.

Multidimensional acoustic analysis combined with machine learning offers a highly accessible, low-cost, and non-invasive triage tool for MS phenotyping, with evidence of independent validation.