Interpretable Phase-Aware Multi-Scale Conv-Transformer Modeling for Recognition and ADL-Based Upper-Limb Function Assessment in Stroke Patients

Stroke remains a leading cause of long-term disability worldwide, with upper-limb motor impairment significantly limiting patients' ability to perform activities of daily living (ADLs). Inertial measurement units (IMUs) have emerged as promising tools for movement monitoring and rehabilitation. Although IMU-based monitoring methods exist, current approaches suffer from several limitations. First, they fail to model multi-scale temporal dependencies underlying primitive-level motor patterns; second, they rely on oversimplified mappings between ADL data and FMA-UE scores. Third, they lack clinically interpretable representations of stroke-specific movement characteristics. In this study, we propose a novel deep learning framework for fine-grained recognition and explainable evaluation of upper-limb movements in post-stroke individuals. The model integrates InceptionTimebased multi-scale temporal convolution, Efficient Channel Attention (ECA), and a Transformer module with Rotary Position Encoding (RoPE), enabling the capture of both short-term motion primitives and long-range dependencies. To enhance interpretability, we propose the Stage-Aware Channel Perturbation Attribution (SCPA) framework, which quantifies channel contributions across distinct motor stages through structured perturbation analysis. Extensive experiments on a public stroke movement dataset demonstrate the model’s state-of-the-art performance, achieving an accuracy of 66.7% in recognizing functional motor primitives and uncovering clinically meaningful patterns, outperforming previous approaches by 4%. Through comprehensive interpretability analysis, we identify robust digital biomarkers ADL-MotorPrints that effectively characterize upper-limb motor states. Among them, the inter-task attribution correlation mean exhibits a strong negative correlation with FMA-UE scores (ρ = –0.88), highlighting its potential as a quantifiable marker of functional impairment. These findings lay the groundwork for next-generation stroke rehabilitation systems, enabling objective, interpretable, and clinically aligned assessment of upper-limb function in ecologically valid ADL scenarios.