Artificial Intelligence Applications to Lower Limb Rehabilitation Robotics

Artificial Intelligence (AI) is increasingly transforming healthcare by enabling data-driven decision support, predictive modeling, personalized treatment planning, and continuous patient monitoring. From medical imaging and clinical risk stratification to wearable sensing and digital therapeutics, AI systems are reshaping how complex physiological signals are interpreted and translated into actionable knowledge. However, despite remarkable progress, several challenges limit the deployment of AI in real world clinical environments. Medical data is often heterogeneous, noisy, sparsely labeled, and characterized by high inter patient variability. Moreover, healthcare applications demand robustness, interpretability, safety, and generalization under distribution shifts requirements that are significantly more stringent than those in many conventional machine learning benchmarks. These challenges become particularly evident in the domain of lower limb rehabilitation robotics. Neurological and musculoskeletal disorders such as stroke, Parkinson’s disease, spinal cord injury, and age related degeneration frequently impair gait and mobility, severely affecting patients’ independence and quality of life. Robotic exoskeletons and assistive devices have emerged as promising tools to support motor recovery through repetitive, task oriented, and adaptive training. By integrating biomechanical sensing, actuation, and control, these systems can deliver intensive and personalized therapy. Nevertheless, the effectiveness of robotic rehabilitation critically depends on accurate modeling and interpretation of human locomotion dynamics. Human gait is a complex, nonlinear, and multi scale process involving coordinated interactions between joints, muscles, sensory feedback, and environmental context. Lower limb robotic systems must detect gait phases, recognize locomotion modes (e.g., level walking, stair ascent/descent, ramp negotiation), estimate kinematics, and adapt assistance in real time. Additionally, subtle deviations in gait patterns may serve as early indicators of pathological conditions, stress states, or neurodegenerative diseases such as Parkinson’s disease. Designing AI models that can robustly capture these temporal dependencies from wearable sensors such as inertial measurement units (IMUs), electromyography (EMG), and pressure sensors remains an open research problem. 3Traditional machine learning approaches often struggle to generalize across subjects and conditions due to limited labeled datasets and the intrinsic variability of clinical populations. Furthermore, many existing architectures either fail to capture long range temporal dependencies or are computationally inefficient for deployment on embedded robotic hardware. These limitations motivate the exploration of advanced neural network architectures capable of modeling structured temporal dynamics while maintaining efficiency and robustness. This thesis aims to advance the field of AI-driven lower limb rehabilitation robotics by addressing two central challenges. The first challenge concerns the investigation and comparative analysis of modern neural network architectures for human locomotion recognition. Specifically, this work systematically studies recurrent neural networks, convolutional temporal models, attention-based Transformers, state space models, and hybrid architectures for gait classification, terrain recognition, and regression tasks such as slope and stair height estimation. By evaluating these models on multimodal locomotion datasets, the thesis analyzes their ability to capture short and long range dependencies, handle sensor noise, and generalize across different walking conditions. Particular attention is given to computational complexity, and suitability for real time robotic control. The results contribute to a clearer understanding of which modeling paradigms are most appropriate for continuous locomotion analysis and exoskeleton assisted rehabilitation. The second challenge addresses the role of Self-PreTraining (SPT) in medical time series modeling. In many healthcare scenarios, labeled data is scarce and expensive to obtain, while large amounts of unlabeled physiological signals are available. Self-pretraining strategies offer a promising solution by enabling models to learn meaningful representations from unlabeled data before fine tuning on downstream supervised tasks. However, the mechanisms through which self-pretraining improves performance—particularly in structured sequence models remain insufficiently understood, especially in medical contexts. This thesis investigates how self-pretraining affects sequence classification architectures, with a focus on Transformer-based models and their internal attention dynamics. Through controlled experiments and ablation studies, the work examines whether performance gains stem from better initialization of specific parameters, improved representation learning, or enhanced optimization landscapes. By analyzing weight norms, parameter displacement, freezing strategies, and positional encoding interactions, the thesis provides empirical insights into how SPT shapes learned representations and facilitates downstream adaptation. Importantly, the benefits of self-pretraining are evaluated not only in locomotion recognition tasks but also in broader medical time series applications, including stress detection and Parkinson’s disease detection. Stress detection from physiological signals requires the 4identification of subtle temporal patterns in heart rate variability, motion, and other biosignals. Parkinson’s detection, particularly through gait and movement analysis, demands high sensitivity to micro variations in motor dynamics. In both cases, labeled datasets are limited and heterogeneous, making them ideal candidates for representation learning through self-pretraining. Experimental results demonstrate that SPT can improve classification accuracy across multiple medical time series datasets, particularly in low data regimes. The findings suggest that self-pretraining enhances the model’s ability to encode structured temporal information and accelerates convergence during fine tuning. Furthermore, different masking strategies and architectural configurations are analyzed to determine how pretraining design choices influence downstream performance. The thesis highlights that the impact of SPT varies with model depth, dataset characteristics, and signal structure, providing practical guidance for its application in healthcare oriented sequence modeling. This work contributes to the development of robust, scalable, and data efficient AI methodologies for lower limb rehabilitation robotics and related medical time series tasks. By integrating architectural advances in long sequence modeling with principled analysis of self-pretraining mechanisms, the thesis bridges theoretical insights and applied clinical challenges. The proposed approaches support more accurate human locomotion recognition, improved stress and Parkinson’s detection, and enhanced adaptability of robotic rehabilitation systems. In doing so, this research advances the vision of intelligent, human centered rehabilitation technologies that leverage AI not merely as a predictive tool, but as an enabling framework for personalized, adaptive, and clinically meaningful motor recovery.