hal-05719778 A deep learning model for speech-based prediction of clinical scores in people (…)

Summary Background Sensitive monitoring tools are needed to track progression in neurodegenerative diseases and assess interventions before overt brain damage occurs. We propose speech as a non-invasive, easily collected biomarker to capture disease-related variation over time. We developed and validated Neurodegenerative Disease Speech Network (NDSNet), an automated deep learning model that generates individual speech-derived estimates of contemporaneous clinical scores at each visit in people with Huntington’s disease, from presymptomatic stages (Huntington’s Disease Integrated Staging System [HD-ISS] stages 0–1) to symptomatic stages (HD-ISS stages 2–3). Methods We included data from people with Huntington’s disease and healthy controls from three prospective longitudinal studies (Bio-HD, REPAIR-HD, and MIG-HD) with speech recordings and Unified Huntington’s Disease Rating Scale (UHDRS) scores. NDSNet combines a pre-trained wav2vec 2.0 model and a recurrent attentive network in a contrastive learning framework for processing audio waveforms of speech. We trained and cross-validated (10-fold) NDSNet to predict the observed UHDRS scores in people with Huntington’s disease across visits, with external validation in a replication cohort (the TPMH study). We then compared NDSNet predictions with striatal atrophy on MRI, the best-established marker of Huntington’s disease progression. Findings Our developmental cohort included 191 people with Huntington’s disease and 58 healthy controls, with speech data collected between 2001 and 2025 (MIG-HD data collected in 2001–13 and REPAIR-HD and Bio-HD data collected in 2018–25). The replication cohort included 110 people with Huntington’s disease, of whom 78 were not included in the developmental cohort, with speech data collected between Oct 10, 2022, and Feb 5, 2024. In the developmental cohort, we analysed speech recordings obtained from 146 people with Huntington’s disease (62 with brain MRI). Relative error between NDSNet predictions and observed clinical scores was 11⋅4% (95% CI 9⋅7–12⋅5) overall. The intraclass correlation coefficient (ICC) between NDSNet-predicted and observed motor scores (ICC 0⋅87 [95% CI 0⋅83-0⋅91]) was similar to that for clinician ratings (ICC 0⋅847). Predictions showed strong temporal association and responsiveness to observed clinical change at the individual level. Performance remained consistent in 67 people with Huntington’s disease in the replication cohort (after excluding 32 participants already included in the developmental cohort; relative error 15⋅0% [13⋅0-19⋅0]; ICC 0⋅67 [0⋅38-0⋅71]). Predicted scores showed MRI associations similar to those of observed clinical scores across symptomatic stages, and stronger associations with striatal atrophy at HD-ISS stages 0–1. Interpretation NDSNet predictions captured clinically relevant progression from speech in people with Huntington’s disease, showed neuroanatomical grounding across disease symptomatic stages, and estimated the subclinical state at HD-ISS stages 0–1. These findings support the use of NDSNet as a scalable tool to complement standard assessments for longitudinal monitoring in Huntington’s disease.