Automated Prediction of Item-Level Arat Scores From Wearable Sensors

The demand for neurorehabilitation is projected to increase drastically over the next decades, creating a need to automate and enhance the efficiency of clinical processes. Clinical assessments are important to track treatment effectiveness and to tailor therapy approaches. We here develop a scoring prediction system for the Action Research Arm Test (ARAT), a common clinical assessment for upper limb function in stroke. To predict each of the 19 movement items individually, we use a raw time series of 5 wearable movement sensors. A general classification model was trained on 100 ARAT tests from various neurological disorders and achieved a mean balanced accuracy across all closely related movement items of 80 % (range: 59%-91%). Training disorder-specific models for stroke and Parkinson's disease did not improve accuracy but yielded specific feature maps that can inform future research. This study demonstrates the feasibility of predicting ARAT scores at the item level, preparing the ground for clinical decision support systems or even automated ARAT scoring.