Contrastive Learning Model for Wearable-based Ataxia Assessment

Objective: Frequent and objective assessment of ataxia severity is essential for tracking disease progression and evaluating the effectiveness of potential treatments. Wearable-based assessments have emerged as a promising solution. However, existing methods rely on inertial data features directly correlated with subjective and coarse clinician-evaluated rating scales, which serve as imperfect gold standards. This approach may introduce biases and restrict flexibility in feature design. To address these limitations, this study introduces a novel contrastive learning-based model that leverages motor severity differences in wearable inertial data to learn relevant features. Methods: The model was trained on inertial data collected from 87 individuals with diagnostically heterogeneous ataxias and 44 healthy participants performing the finger-to-nose task. A pairwise contrastive loss function was proposed to learn representations capturing relative differences in ataxia severity, which were evaluated through downstream regression and classification tasks. Results: The learned features demonstrated strong cross-sectional (r = 0.84) and longitudinal (r = 0.68) associations with clinical scores and robust measurement reliability (intraclass correlation coefficient = 0.96). Additionally, the model exhibited strong known-group validity, distinguishing between ataxia and healthy phenotypes with an area under the receiver operating characteristic curve of 0.95. Conclusion: The proposed contrastive model captures robust representations of disease severity with reduced reliance on clinical scales, outperforming state-of-the-art methods that derive features directly from clinical scores. Significance: Combining wearable sensors with contrastive learning enables a more objective, scalable, and frequent method for assessing ataxia severity, with the potential to enhance patient monitoring and improve clinical trial efficiency.