FROM AI-BASED NEUROLOGICAL DIAGNOSIS TO INTELLIGENT NEUROREHABILITATION: ADVANCED DEEP LEARNING ARCHITECTURES LEVERAGING BIOMECHANICAL AND GAIT ANALYSIS DATA FOR SMART DIAGNOSIS, THERAPEUTIC DECISION-MAKING, AND NEUROLOGICAL DISORDER MANAGEMENT

Abstract

Neurological disorders comprise a diverse group of conditions that significantly compromise motor control, cognition, and quality of life, with gait abnormalities and biomechanical dysfunctions being particularly salient clinical indicators. Traditional diagnostic and monitoring techniques such as clinical rating scales and imaging are often subjective, costly, and limited in scalability. In recent years, artificial intelligence (AI), particularly deep learning (DL), has demonstrated transformative potential in extracting nuanced patterns from high-dimensional multimodal data, including wearable sensor signals, three-dimensional gait kinematics, and video-based motion capture. These advances open new avenues for objective, scalable, and automated neurological assessment that extend beyond conventional evaluation. Current research demonstrates that DL models can effectively process sensor-derived gait data to classify pathological versus healthy locomotion, identify disease subtypes, and estimate clinical severity with promising accuracy levels, often outperforming traditional machine learning baselines.   Despite this progress, most existing studies focus predominantly on isolated diagnostic tasks (e.g., Parkinson’s disease identification or freezing-of-gait detection), evaluated on single-center datasets, which limits the scope of clinical deployment. Furthermore, model interpretability and between-center generalization remain key challenges that diminish trust in AI recommendations.     In this work, we propose a comprehensive AI-enabled framework that integrates deep learning architectures with biomechanical and gait analysis data to bridge the continuum from diagnosis to intelligent neurorehabilitation and clinical decision-making. We leverage multimodal data streams wearable inertial measurement units (IMUs), ground reaction forces, kinematic features, and clinical indicators and design hierarchical fusion models that combine convolutional, recurrent, and attention-based modules for robust feature extraction. Our framework supports multi-task learning for simultaneous disorder classification, severity scoring, and therapeutic intervention suggestions. We introduce explainable components that highlight biomechanically meaningful features driving AI decisions, fostering clinical interpretability and acceptance. Evaluations on benchmark gait and neurological datasets demonstrate improved generalization, diagnostic accuracy, and alignment with clinical severity scales compared to modality-specific and traditional methods. By unifying objective AI biomarker extraction with intelligent therapeutic guidance, our approach represents a significant step toward personalized, adaptive neurorehabilitation solutions that can be deployed in real-world clinical environments, ultimately enhancing clinical outcomes and patient quality of life.