上肢
物理医学与康复
冲程(发动机)
医学
特征(语言学)
特征选择
康复
选择(遗传算法)
人工智能
结果(博弈论)
计算机科学
作者
Anastasios Tzepkenlis,Antonio Frisoli,Cristian Camardella
标识
DOI:10.1109/tnsre.2026.3744492
摘要
Accurate and interpretable prediction of rehabilitation outcomes is critical for enabling data-driven clinical decision support in post-stroke robotic therapy. This study systematically investigates feature selection and model explainability to improve the clinical utility of machine learning models that predict post-therapy motor function scores, specifically the Fugl-Meyer Assessment (FMA), Action Research Arm Test (ARAT), and Motricity Index (MI), in stroke patients undergoing upper limb robot-assisted rehabilitation. Building on prior work that identified XGBoost as the top-performing algorithm, we applied Sequential Forward Selection (SFS) with beam search to identify compact, high-performance feature subsets from a 71-feature dataset collected from 40 stroke patients using the MOTORE end-effector robotic platform. Complementary explainability analyses using SHAP values and Permutation Importance (PI) were integrated to identify a consensus feature set of 3–4 variables per outcome. A systematic ablation over the four feature groups shows that clinical variables carry most of the predictive signal, that adding kinematic or time-related variables to the clinical group leaves the error essentially unchanged, and that the benefit of the robotic data emerges only once feature selection is applied. SFS-optimized models reduced prediction error by 26–42% relative to the full 71-feature model and by 20–36% relative to a clinical-only model, with normalized RMSE values of 0.087 (ARAT), 0.046 (FMA), and 0.086 (MI). The compact SHAP+PI feature sets remained usable (nRMSE: 0.130, 0.067, 0.093) while reducing the input space to 3–4 variables. Learning curve analysis shows that error is still decreasing at 40 patients, so the present estimates should be read as an upper bound on the achievable error. Prediction errors were benchmarked against established minimal clinically important difference (MCID) thresholds, demonstrating sufficient precision to distinguish patients likely to achieve meaningful motor recovery. These findings support the development of transparent, clinically deployable decision support systems for robotic rehabilitation.
科研通智能强力驱动
Strongly Powered by AbleSci AI