Salvatore Giugliano, Tabea Thies, Elisa Mallick, Johannes Tröger, Myriam Spisto, Raffaele Dubbioso & Giovanna Sannino
SOCO 2026
Abstract:
Amyotrophic Lateral Sclerosis (ALS) is characterized by progressive bulbar dysfunction, necessitating objective, low-burden monitoring tools. This study presents an exploratory analysis investigating whether parsimonious, physiologically interpretable speech measures can serve as candidate digital biomarkers for ALS. Using smartphone recordings from 272 participants (165 ALS, 107 Healthy Controls) in the SAND challenge training cohort, we extracted composite acoustic scores—Articulation, Tempo, and Stability—from oral diadochokinesis (DDK) tasks, alongside vowel space metrics from sustained phonations. We evaluated the in-sample discriminative potential of these markers for disease detection and dysarthria severity classification. Results indicate task-specific clinical utility: /ka/-based metrics demonstrated the highest potential for differentiating ALS from controls (in-sample AUC up to 0.74), while articulatory stability derived from the /ta/ task emerged as the strongest candidate indicator of dysarthria severity within the ALS cohort (AUC = 0.81). To optimize discriminative performance, we systematically evaluated multi-feature combinations, achieving an in-sample AUC of 0.86 for severity classification. Crucially, as this study relies on in-sample threshold optimization without a hold-out test set, these multi-feature findings represent an optimistic upper bound. Nonetheless, by shifting from complex “black-box” models to transparent indices of motor control, this work provides a vital conceptual foundation. We propose a targeted set of interpretable speech features that, pending rigorous cross-validation on independent datasets, hold significant promise for remote assessment in future multi-center ALS trials.