Modeling of user behavior using artificial intelligence techniques for the design and adaptation of rehabilitation therapies assisted by robotic devices

Abstract

Recent reports indicate that the prevalence of stroke has increased by 50% over the last 17 years, and at present, 1 in 4 people are at risk of getting a stroke in their lifetime. Hemiparesis is a common after-effect of stroke, leading to a decrease in limb function and significantly affecting quality of life. However, personalized therapy adapted to the needs of each patient can lead to the recovery of mobility of the affected limbs. This doctoral thesis focuses on the study and modeling of user behavior to design and adapt robotic rehabilitation systems to the user's needs. First, an underwater sensor system has been developed to measure the kinematics of the lower limbs. In addition, a forces model acting during gait in aquatic environments has been developed with the aim of designing and subsequently controlling a lower limb exoskeleton immersed in water. Moreover, a Machine Learning model has been proposed to overcome the difficulties encountered in measuring the vertical ground reaction force in an aquatic environment. The model takes as inputs the kinematics of the lower limbs in the sagittal plane and the vertical acceleration of the user. Furthermore, a Deep Learning model has been developed to recognize a set of gait activities to generalize the previous model and to adapt the assistance of a lower limb exoskeleton to different activities of daily living. Also, in order to adapt and personalize robot-assisted upper limb therapies, two approaches are presented. First, a method has been proposed to dynamically adapt the difficulty of upper limb rehabilitation therapies. In addition, a method to provide assistance according to the needs of each patient based on demonstrations by an experienced therapist has been proposed. Part of the main results of this thesis have been published in two journals included in the Journal Citation Reports (JCR). The publication "Estimating vertical ground reaction forces during gait from lower limb kinematics and vertical acceleration using wearable inertial sensors" focuses on the development of a Machine Learning model to estimate the vertical reaction force during gait using a motion capture system based on inertial sensors. The publication "Gait Activity Classification with Convolutional Neural Network using Lower Limb Angle Measurements from Inertial Sensors" focuses on the development of a classifier based on a Deep Learning model to recognize activities of daily living during gait, employing techniques to understand the decision-making process and analyzing the feasibility of adapting the model to multiple situations.