A Ranking-based Weakly Supervised Learning model for telemonitoring of Parkinson's disease

Telemonitoring is the use of electronic devices to monitor patients remotely. A model is needed to translate the data collected by a patient's mobile device into a predicted score for disease severity assessment. Labeled samples are scarce, which makes it difficult to train a supervised learning model. On the other hand, there is an abundance of samples without precise labels but whose relative rank can be known from domain knowledge. We propose a Ranking-based Weakly Supervised Learning (RWSL) model to integrate both types of data. We apply RWSL to predict Parkinson's disease severity based on mobile-collected tapping activity data of patients. RWSL achieves high predictive accuracy and outperforms competing methods.