Development of a deep learning model for IMU time-series segmentation for tele-rehabilitation

Telerehabilitation is becoming increasingly important, offering patients with severe disabilities or chronic illnesses the opportunity to access rehabilitation services directly from their homes. This approach is particularly significant for individuals facing geographical or mobility limitations, enabling effective care through the use of telecommunication technologies combined with advanced sensor systems, such as Inertial Measurement Units (IMU). IMU sensors play a critical role in capturing precise motion data, which, when processed with cutting-edge Information and Communication Technology (ICT) and Machine Learning techniques, allows a more efficient and personalized rehabilitation experience. This project aims to develop an offline method for segmenting repetitions of telerehabilitation exercises, using motion data acquired from the ARC intellicare system, a medical device designed by Henesis S.r.l. for remote rehabilitation. This system employs multiple wearable IMU sensors to monitor patient movements during therapy sessions, collecting time series data to assess patient progress over time. The thesis begins by proposing a baseline approach based on Support Vector Machines (SVM), in which different configurations of feature extraction and kernel functions are tested to determine the most effective method for segmenting exercise repetitions. Following this, the work explores a more advanced deep learning strategy, adapting the U-Time model, originally proposed by Perslev et al. (2019) for sleep stage segmentation. U-Time, a deep learning model based on the U-net architecture, is particularly well-suited for time series analysis and is modified here to fit the specific requirements of IMU time series in a telerehabilitation scenario. Extensive experiments were conducted on the U-Time model to evaluate different architectural sizes and input feature dimensionalities. These experiments aimed to optimize both the depth and the structure of the model, ensuring the best trade-off between accuracy and computational efficiency. Following this, a fine-tuning process was applied to further refine the model’s performance. The results demonstrate that the U-Time model, once optimized, significantly outperforms the traditional SVM approach in terms of segmentation accuracy, with an average F1-score of 73% across 41 different exercises in the ARC library. This research highlights the potential of deep learning in enhancing the precision of telerehabilitation, paving the way for more scalable and efficient remote rehabilitation services in the future.