# The role of surface EMG in predicting responsiveness of muscles to FES therapy after cervical SCI

## TL;DR

This study explores using surface EMG signals and machine learning to predict which muscles will respond to FES therapy in people with cervical spinal cord injuries.

## Contribution

A novel approach combining baseline sEMG features and machine learning to predict FES therapy response in cervical SCI patients.

## Key findings

- A Random Forest classifier achieved 76% accuracy in predicting muscle response to FES therapy.
- Key sEMG features included slope sign changes, mean and median frequency, and second-order spectral moment.
- Model performance improved when stratifying patients by motor completeness (AIS A-B vs. C-D).

## Abstract

Cervical spinal cord injury (SCI) can severely impair upper extremity (UE) functions, limiting independence and quality of life. Prior clinical trials showed that functional electrical stimulation (FES) therapy can reduce UE impairment. However, the response to FES therapy is not consistent across all treated myotomes. Our objective was to predict the muscle response to FES therapy using electrophysiological biomarkers from baseline surface electromyography (sEMG) signals, in order to support treatment decisions at the point of care.

We recruited 17 participants with cervical SCI, who were about to undergo FES therapy. Target UE muscles were identified for each participant by treating therapists. Baseline sEMG signals were recorded from the target muscles during resting and maximal voluntary contractions. Time- and frequency-domain features were extracted. The manual muscle testing (MMT) score was tracked through the therapy cycle, and used to categorize each muscle as a responder or non-responder. We explored classifiers including support vector machines, k-nearest neighbors, random forest, and logistic regression, with leave-one-participant-out cross validation. Models were trained on sEMG features alone, on clinical variables alone, and combinations of both.

The final dataset consisted of sEMG recordings of 132 muscles from 17 participants, and 33% of the muscles were considered responders. A Random Forest classifier with a forward-selected feature set yielded the best performance (Matthews correlation coefficient = 0.41, F1 score = 0.68, accuracy = 76%, precision = 0.72, recall = 0.42, and true negative rate = 0.92). With patient stratification based on motor completeness (AIS A-B vs. C-D), the model performance further improved. Included signal features were slope sign changes, mean and median frequency, and second-order spectral moment.

Our results suggest that baseline sEMG signals combined with machine learning models may be used to predict muscle response to FES therapy in individuals with cervical SCI. The models were trained on a small and unbalanced sample and can be optimized with more participants in the future. This work contributes to improving the level of personalization and efficacy of FES therapy, and ultimately improve quality of life after SCI.