Detecting Daily Living Gait Amid Huntington's Disease Chorea using a Foundation Deep Learning Model

TL;DR

This study tackles accurate gait detection in daily living for Huntington's disease with chorea using wrist-worn accelerometers. It introduces J-Net, a segmentation-based architecture that combines a foundation self-supervised model with a specialized segmentation head to achieve sample-level gait classification, outperforming a baseline classifier, especially at higher chorea severities. J-Net achieved a ROC-AUC of $0.97$ in HD in-lab data and produced daily-living walking estimates that correlated with disease severity ($r=-0.52$, $p=0.02$), while avoiding spurious gait detections that plagued prior methods. The approach generalizes to Parkinson's disease after fine-tuning and is complemented by open-source data and code, enabling broader application of daily-living gait assessment in neurodegenerative disorders.

Abstract

Wearable sensors offer a non-invasive way to collect physical activity (PA) data, with walking as a key component. Existing models often struggle to detect gait bouts in individuals with neurodegenerative diseases (NDDs) involving involuntary movements. We developed J-Net, a deep learning model inspired by U-Net, which uses a pre-trained self-supervised foundation model fine-tuned with Huntington`s disease (HD) in-lab data and paired with a segmentation head for gait detection. J-Net processes wrist-worn accelerometer data to detect gait during daily living. We evaluated J-Net on in-lab and daily-living data from HD, Parkinson`s disease (PD), and controls. J-Net achieved a 10-percentage point improvement in ROC-AUC for HD over existing methods, reaching 0.97 for in-lab data. In daily-living environments, J-Net estimates showed no significant differences in median daily walking time between HD and controls (p = 0.23), in contrast to other models, which indicated counterintuitive results (p < 0.005). Walking time measured by J-Net correlated with the UHDRS-TMS clinical severity score (r=-0.52; p=0.02), confirming its clinical relevance. Fine-tuning J-Net on PD data also improved gait detection over current methods. J-Net`s architecture effectively addresses the challenges of gait detection in severe chorea and offers robust performance in daily living. The dataset and J-Net model are publicly available, providing a resource for further research into NDD-related gait impairments.

Figures (9)

• Figure 1: Illustration of the gait detection algorithm for HD patients (top) and the J-Net segmentation model architecture (bottom).
• Figure 2: Examples of 10-second, 3-axis windows of accelerometer signals, with the Y-axis representing the vertical direction when the wearable device is properly worn on a vertically hanging hand. Shown are examples for: (a) Gait of a healthy control (HC), (b) Gait of an individual with Huntington's Disease (HD) without chorea (chorea 0), (c) Gait of an individual with slight chorea (chorea 1), (d) Gait of an individual with moderate chorea (chorea 2), (e) Gait of an individual with severe chorea (chorea 3), misclassified as non-gait by the baseline model but correctly classified as gait by the J-Net segmentation model, and (f) Non-gait (standing) of an individual with severe chorea (chorea 3), misclassified as gait by the baseline model but correctly identified as non-gait by J-Net.
• Figure 3: Diurnal walking patterns (minutes per hour). The top row shows classification model predictions for HC (a) and HD (b), while the bottom row shows segmentation model predictions for HC (c) and HD (d). As expected, Walking decreases at night. Counter-intuitively, the classification model indicates HD patients walked more than HC controls (median daily walking time, $p<0.005$), a finding not supported by the segmentation model ($p=0.234$).
• Figure 4: Daily walking percentage for each subject for each day predicted by the classification model and by the segmentation model. The y-axis denotes the classification predicted walking percentage, while the x-axis represents the segmentation predicted walking percentage. Each point corresponds to an individual on each day included in the analysis, with blue dots representing the HC group and red dots representing the HD group. The difference between the models leads to a clear separation between healthy controls (HC) and individuals with Huntington’s disease (HD). The dashed line represents the line y=x, where the predictions of the two models would be equal.
• Figure S1: Confusion Matrix illustrating performance per chorea level of the J-Net segmentation model.