A machine learning approach to predict wrist posture in telerehabilitation with haptic devices

Stroke survivors often experience fine motor impairments that prevent them from participating in activities of daily living, adversely impacting their quality of life. Telerehabilitation with haptic devices has the potential to engage survivors in sensorimotor therapy of their hand and wrist, while also collecting pertinent information about their movement towards remote assessment by a medical professional. Nonetheless, it remains challenging to measure patients’ joint angles during interaction with haptic devices, which undermines their prospective use in telerehabilitation. We propose a simple set-up where patients wear a smartphone on their forearm while manipulating the haptic device. In this setting, data from inertial sensors embedded in the smartphone would be integrated with data from the haptic device via a machine learning algorithm to predict the patients’ wrist angle. We demonstrate the feasibility of this approach in experiments with 19 healthy subjects. We measure their wrist angle as they perform a motor task with a Novint Falcon haptic device while wearing sensors on their limb, and train a linear regression model that predicts the wrist angle. The model predicts wrist angles with an accuracy of 88.8%. This effort constitutes a significant step toward automatic assessment of joint movements in fine motor telerehabilitation.