Assessing the Generalizability of Foundation Models for the Recognition of Motor Examinations in Parkinson's Disease

Current machine learning approaches focusing on motor symptoms in Parkinson's disease are commonly trained on small datasets and often lack generalizability from developmental setups to clinical applications. Foundation models using large, unlabeled datasets of healthy participants through self-supervised learning appear attractive for such setups with limited samples, despite the potential impact of motoric symptoms. Acting as an exemplar, this study aims to evaluate the robustness of fine-tuned models in recognizing movements related to motor examinations across datasets and recording setups. Accelerometer data of 51 participants with Parkinson's disease in three different training and fine-tuning setups were used to tailor the general model to the disease. Training the model on pre-trained weights, both partially (F₁ = 0.70) and fully (F₁ = 0.69), statistically significantly outperformed training the model from scratch (F₁ = 0.55) in a nested cross-validation. For evaluation, the model's ability to process data recorded from 24 patients in clinic was tested. The models achieved lower mean F₁ scores of 0.33 (train from scratch), 0.43 for full, and 0.48 for partial fine-tuning, but demonstrated improved generalizability and robustness regarding the orientation of sensors compared to training from scratch. Utilizing foundation models for accelerometer data trained on healthy participants and fine-tuned for clinical applications in movement disorders appears as an effective strategy for optimized generalizability with small datasets.