Objective: To investigate the diagnostic utility of machine learning-driven frequency-domain analysis of gait acceleration for the objective identification of Huntington’s Disease (HD).

Background: Gait disturbances frequently manifest as early clinical indicators of HD. However, standard clinical evaluations rely heavily on subjective rating scales that may miss subclinical kinematic alterations. Frequency-domain analysis, a robust technique in complex biomedical signal processing, excels at characterizing rhythmic and continuous cyclic patterns.

Method: This cross-sectional study evaluated 148 participants (78 HD patients and 70 healthy controls) performing a standardized overground walking protocol. Participants wore inertial measurement units (IMUs) positioned across key anatomical sites: bilateral feet, shanks, thighs, lumbar spine, chest, and wrists. Raw acceleration signals from each sensor were transformed into the frequency domain to systematically extract distinct spectral features. A Universal Background Model-Gaussian Mixture Model (UBM-GMM) classifier was trained to distinguish HD patients from controls. Classification efficacy was analyzed across isolated individual sensors and mathematically optimized, multi-sensor spatial configurations.

Results: Low-frequency spectral features emerged as the most highly discriminative kinematic biomarkers for HD detection, effectively capturing the characteristic arrhythmicity of choreiform gait. When evaluating sensors individually, the right thigh IMU yielded the highest standalone diagnostic accuracy (81%, Area Under the Curve [AUC] = 0.85). Integrating data across an optimized, multi-sensor array—comprising bilateral thighs, the lumbar spine, the left foot, and the right arm—substantially elevated the overall classification accuracy to 93% (AUC = 0.95), highlighting the immense diagnostic value of multi-nodal spatial integration.

Conclusion: Frequency-domain analysis of whole-body gait acceleration provides a highly accurate, non-invasive, and objective machine learning framework for early HD detection. Proximal lower limb and axial sensors supply the most informative kinematic data. This scalable methodology generates robust digital biomarkers to augment early clinical assessment and track HD progression objectively.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/machine-learning-based-frequency-domain-analysis-of-gait-acceleration-for-huntingtons-disease-detection/