Objective: To derive a speech-based score reflecting Huntington’s disease (HD) progression from deep neural representations of raw voice recordings and evaluate its potential as a digital biomarker.

Background: Speech an emerging digital biomarker [1,2] in HD. Most existing approaches infer disease severity from speech using engineered acoustic features to predict predefined clinical scales such as the Unified Huntington’s Disease Rating Scale (UHDRS). We previously developed NDSNet [3], a deep neural network trained to predict several UHDRS scores directly from raw speech waveforms. While this approach models the relationship between speech and clinical status more directly than feature-based methods, it remains constrained by predefined clinical scales. Here, we investigate whether the latent representations learned by NDSNet capture a deeper signature of disease progression that can be used to derive a speech-based biomarker beyond existing clinical measures.

Method: Speech recordings were collected from participants enrolled in the BioHD (NCT01412125) and RepairHD (NCT03119246) studies, including 176 participants and 298 visits spanning the HD spectrum (HC=78, HD-ISS 0–1=78, stage 2=73, stage 3=69). Recordings consisted of scripted and spontaneous speech segments of up to 10 seconds. For all visits included in this study, speech recordings, structural MRI data, and clinical UHDRS assessments were available. Structural MRI provided volumetric estimates of approximately 120 brain regions. A neurodegeneration axis reflecting disease severity was identified using partial least squares regression, relating regional brain volumes to the composite UHDRS (cUHDRS) [4] while adjusting for demographic covariates (age, sex, and education). Speech segments were projected into the latent space of NDSNet, and a regression model was trained to predict this neurodegeneration axis from latent speech representations using cross-validation with patient-level splits.

Results: The resulting speech-derived score correlated with cUHDRS (Spearman ρ≈0.60, p<10⁻¹²) and differentiated HD stages. Regional analyses indicated that the score reflected degeneration patterns involving striatal, motor, and temporal regions, consistent with known neural substrates of speech impairment in HD.

Conclusion: Deep neural representations of raw speech capture a biologically meaningful signature of HD progression.

References: [1] Riad R, Lunven M, Titeux H, et al.
Predicting clinical scores in Huntington’s disease: a lightweight speech test.
Journal of Neurology. 2022;269:5008-5021.

[2] Nunes AS, Pawlik M, Mishra RK, et al.
Digital assessment of speech in Huntington disease.
Frontiers in Neurology. 2024.

[3] Le Moine Veillon C, et al.
Deep neural speech representations capture clinical progression in Huntington’s disease.
Lancet Digital Health. Under review.

[4] Schobel SA et al.
Motor, cognitive, and functional declines contribute to a single progressive factor in early HD.
Neurology 2017; 89: 2495–502.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/a-biomarker-of-huntingtons-disease-progression-derived-from-deep-neural-speech-representations/