Enhancing Model Generalizability In Parkinson's Disease Automatic Assessment: A Semi-Supervised Approach Across Independent Experiments

Machine learning in Parkinson's disease assessment uses data from clinically-coded movements, such as finger tapping, to objectively measure motor impairment. Video-based models showed promise in several experiments, but the lack of a unified test benchmark hinders proving generalizability. Additionally, new telemedicine systems may easily collect large amounts of unsupervised data, while obtaining ground truth labels for supervised learning remains time-consuming and requires specialized clinicians. This study explores semi-supervised learning to enhance the generalizability of a Light Gradient Boosting model for video-based finger tapping staging, while reducing its need for supervised data labelling. Specifically, this work employs the Self-training schema in two trials using openly-available finger tapping datasets from three independent experiments. This method significantly improves model performance across various metrics, achieving notable accuracy gains (e.g., from 87.62% to 92.05%) when tested on unseen data from a different experiment. Semi-supervision proves valuable when limited labelled data (less than 10%) from the test distribution are available during training.