Developing deep learning-based solutions for IMU time series with application in Tele-Rehabilitation

Tele-rehabilitation has emerged as a versatile and valuable approach, particularly in the post-Covid-19 era. Among the various techniques employed for data capture in tele-rehabilitation, Inertial Measurement Unit (IMU) sensors have gained prominence due to their ability to capture precise movements. These sensors are favored for their compact size, low power consumption and privacy-preserving features, making them a preferred choice in the field. Despite existing literature explores the segmentation of IMU time series in tele-rehabilitation studies, there is a noticeable gap in offering general solutions to address segmentation across a large variety of rehabilitation devices and motor exercises. This thesis targets to address this gap by proposing a novel solution based on deep learning techniques to effectively segment IMU time series in the context of tele-rehabilitation. The proposed solution involves a series of pre-processing steps, a convolutional neural network (CNN), and post-processing procedures. The proposed model was then tested over a rich dataset obtained from the Henesis srl's ARC intellicare tele-rehabilitation device, incorporating data from 62 participants both healthy and pathological engaged in 41 distinct exercises, each monitored by three IMU sensors. The results indicate that the proposed model accurately identifies movements' initiation and completion with an average true positive rate of 90%; however, there is still room for improvement in mitigating a certain amount of false discovery rate (largely varying depending on the specific exercise). This work contributed towards the improvement of the company's system, while future investigations will involve a comparative analysis between the proposed model and the existing company solution. Moreover, it can also have an impact on other systems using IMU sensors to monitor rehabilitation treatments.