xHD-Vox, an Automated Speech Model for Estimating Motor and Cognitive Scores in Huntington’s Disease: Development and Longitudinal Validation (Preprint)

BACKGROUND Huntington’s disease (HD) is a rare genetic neurodegenerative disease that causes progressive motor, cognitive, and psychiatric symptoms over decades after onset. Care is typically provided in specialized centers with only annual clinical assessments, highlighting the need for more frequent and cost-effective monitoring. OBJECTIVE The aim of this study was to develop and validate xHD-Vox, a fully automated, interpretable, speech-based model for longitudinal monitoring of Huntington’s disease. METHODS We included HD gene carriers from three French prospective cohorts: BIO-HD (NCT01412125), REPAIR-HD (NCT03119246), and MIG-HD (NCT00190450). Participants had ≥40 CAG repeats, available composite Unified Huntington’s Disease Rating Scale (cUHDRS) scores, and audio recordings of forward and backward counting (1-20). Using speech pathologists’ annotations, we extracted 60 validated features and identified a minimal subset to predict cUHDRS and its cognitive, motor, and functional components. Performance was assessed using mean absolute error (MAE) across 50 cross-validation folds (20% participants in validation folds). We then automated feature extraction using Whisper, an open-source speech recognition tool. Final model performance was evaluated on an independent longitudinal test set, using MAE, explained variance (R²), and intraclass correlation coefficient (ICC), and comparison of predicted vs. clinician-assessed one-year score changes at HD-ISS Stages 2-3. RESULTS Our cohort included 181 HD gene carriers totalizing 341 visits with an average of 2 ± 1 annual visits (29% in HD-ISS Stages 0-1, 22% in Stage 2, and 49% in Stage 3). A total of 145 recordings from 90 participants were annotated by speech pathologists. Feature selection identified four key predictors: standardized CAG-age-product (CAP) score, CAG repeat length, rate of number pronounced per second, and the standard deviation of that rate. On the independent test set (24 participants, 3 annual visits), xHD-Vox achieved a MAE of 2.1 for cUHDRS and explained 57% of its variance - compared to 43% when using CAP score alone. Model performance remained stable across three consecutive annual visits. Moreover, xHD-Vox detected a decline consistent with clinicians’ assessment: one-year and two-year mean predicted change fell within 95% CI interval of clinician’s assessed change. CONCLUSIONS We developed xHD-Vox, an interpretable and automated model that predicts clinical progression in HD using a short speech task. Performance was consistent over time and supports its use in mobile applications for remote monitoring. This approach could facilitate scalable, real-time tracking of disease progression, especially in underserved regions, and enable personalized and responsive clinical care.