Surface EMG-based classification and prediction of Fugl-Meyer upper extremity scores in subacute stroke

 Nothing here is of any use to recovery!  Fugl-Meyer has no objectivity at all, so you can't map recovery protocols to your scores!

'Assessments' like Fugl-Meyer NEVER GET ANYONE RECOVERED! I'd have you all fired for incompetency in not solving stroke!

Surface EMG-based classification and prediction of Fugl-Meyer upper extremity scores in subacute stroke

• Hao-Ping Lin,
• Lina Zhao,
• Daniel Woolley,
• Xue Zhang,
• Weidi Liang,
• Hsiao-ju Cheng,
• Lixin Zhang &
• Nicole Wenderoth 

We’re sharing this article early to provide faster access to peer-reviewed, accepted research. It is citable and carries a permanent DOI. This version is subject to further edits and will be replaced automatically by the final Version of Record. All legal disclaimers apply.

Abstract

Background

Stroke is a leading cause of long-term disability worldwide, and with rising incidence and a global shortage of rehabilitation professionals, there is a growing need for scalable methods to assess motor impairment. Surface electromyography (sEMG) has emerged as a promising modality for capturing motor function. Despite its potential, sEMG remains underused in clinical practice and has been predominantly studied in chronic stroke populations, leaving the subacute phase largely unexplored. Its ability to complement standardized assessments is not yet well established.

Methods

Bilateral sEMG was recorded from 23 subacute stroke participants performing four standardized wrist and hand tasks. Task-level features were extracted to train machine learning models to classify the presence of an impairment (affected or less-affected) and the level of impairment, as measured by the Upper Extremity Fugl-Meyer Assessment (FMA-UE) and its wrist and hand subsections (FMA-WH). Models were evaluated using leave-one-subject-out cross-validation (LOSO-CV). Benchmark regression models used therapist-rated subscores as inputs and paired t-tests compared their absolute errors against those of the sEMG-based models. Model interpretability was examined using SHapley Additive exPlanations values (SHAP) to identify sEMG features contributing most strongly to predicted impairment levels.

Results

For impairment classification, Wrist Extension yielded the highest performance in classifying impairment (Accuracy 0.87 ± 0.16; Area Under the Receiver Operating Characteristic [AUC-ROC] 0.96 ± 0.07). An evaluation of all binary and triple task combinations revealed that Wrist Extension combined with Pincer Grasp achieved the best overall classification (Accuracy 0.92 ± 0.12; AUC-ROC 0.99 ± 0.03). For impairment level estimation, the sEMG-based predictions reached a Root Mean Squared Error (RMSE) of 3.12 for FMA-WH and 6.68 for FMA-UE. SHAP analysis with the sEMG-based model revealed that higher extensor activation strongly drove higher predicted scores.

Conclusions

Using sEMG signals obtained from a consumer-grade armband during hand and wrist tasks enabled estimation of partial and full FMA-UE scores, achieving prediction errors below the minimally clinically important difference (MCID) of the full FMA-UE. Furthermore, a two-task protocol (Wrist Extension + Pincer Grasp) achieved the highest classification accuracy, demonstrating that clinical assessment burden could be reduced without compromising performance. This highlights the potential for scalable and portable assessment solutions, though larger longitudinal validation is required.