Objective: This study analyzes the speech and facial  characteristics of Huntington’s disease (HD) and explores their relationship with clinical manifestations to develop machine learning models for disease severity assessment.

Background: Previous studies have shown that voice alterations and facial dyskinesia appear early in HD and correlate with disease progression. Using machine learning algorithms to analyze voice and facial features offers a more objective and sensitive approach.

Method: From June 31, 2022, to December 31, 2024, patients diagnosed with HD, mutant HTT gene carriers, and healthy controls were recruited from the neurodegenerative clinic at Beijing Tiantan Hospital. Demographic, clinical data, and motor assessment using the Unified Huntington’s Disease Rating Scale (UHDRS) motor component, were collected. A standardized protocol for speech and facial expression was created. 42 speech features, such as SD Delta 2 (sound quality), Loud stdNorm (voice loudness), and VoicedSegsPerSec (paragraph features), and 17 facial expression features, including eyebrow variability, blink frequency, and dyskinesia were extracted using Python for analysis. The diagnostic value of speech and facial expression features for HD was evaluated with receiver operating characteristic (ROC) curves. Lasso regression identified factors linked to UHDRS-motor scores, and Random Forest models were used to distinguish HD from controls.

Results: 63 participants (30 HD patients, 33 controls) were included. HD patients showed significant differences in pitch variation and loudness compared to controls, especially during long text readings. The area under the curve (AUC) for several features was greater than 0.8, with SD Delta2 being the best feature (AUC = 0.916). Additionally, HD patients exhibited reduced blink frequency, longer blink intervals, and more facial dyskinesia, with AUCs ranging from 0.730 to 0.916. Lasso regression identified key features (e.g., right eyebrow variability y, Loud stdNorm, and VoicedSegsPerSec) associated with UHDRS-motor scores (R² = 0.7795). The Random Forest model successfully distinguished HD from controls with 100% accuracy.

Conclusion: This study demonstrates that HD patients have unique voice and facial expression features, and the machine learning models derived from these features as a tool for the accurate diagnosis and the progression monitoring of HD.

References: 1.Riad R, Lunven M, Titeux H, et al. Predicting clinical scores in Huntington’s disease: a lightweight speech test. J Neurol. 2022;269(9):5008-5021. doi:10.1007/s00415-022-11148-1
2.Kouba T, Frank W, Tykalova T, et al. Speech biomarkers in Huntington’s disease: A cross-sectional study in pre-symptomatic, prodromal and early manifest stages. Eur J Neurol. 2023;30(5):1262-1271. doi:10.1111/ene.15726
3.Diehl SK, Mefferd AS, Lin YC, et al. Motor speech patterns in Huntington disease. Neurology. 2019;93(22):e2042-e2052. doi:10.1212/WNL.0000000000008541

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/exploring-the-characteristic-features-of-speech-and-facial-expressions-in-patients-with-huntingtons-disease-based-on-machine-learning/