ReactEMG Stroke: Healthy-to-Stroke Few-shot Adaptation for sEMG-Based Intent Detection

TL;DR

This work tackles the calibration bottleneck in sEMG-based stroke intent detection for assistive hand devices by proposing healthy-to-stroke few-shot adaptation. It leverages a healthy-domain foundation model, ReactEMG, and evaluates three fine-tuning strategies (Head-only, LoRA, Full) against zero-shot and stroke-only baselines on a new 3-participant stroke dataset with distribution shifts. Results show that healthy-pretrained adaptation improves both transition and raw accuracy under realistic shifts, with average gains of up to 0.19 in transition accuracy and 0.09 in raw accuracy, and data-efficiency benefits vary by participant. The findings support transferring reusable healthy EMG representations to stroke users to reduce calibration burden and enhance robustness for real-time intent detection in neurorehabilitation robotics.

Abstract

Surface electromyography (sEMG) is a promising control signal for assist-as-needed hand rehabilitation after stroke, but detecting intent from paretic muscles often requires lengthy, subject-specific calibration and remains brittle to variability. We propose a healthy-to-stroke adaptation pipeline that initializes an intent detector from a model pretrained on large-scale able-bodied sEMG, then fine-tunes it for each stroke participant using only a small amount of subject-specific data. Using a newly collected dataset from three individuals with chronic stroke, we compare adaptation strategies (head-only tuning, parameter-efficient LoRA adapters, and full end-to-end fine-tuning) and evaluate on held-out test sets that include realistic distribution shifts such as within-session drift, posture changes, and armband repositioning. Across conditions, healthy-pretrained adaptation consistently improves stroke intent detection relative to both zero-shot transfer and stroke-only training under the same data budget; the best adaptation methods improve average transition accuracy from 0.42 to 0.61 and raw accuracy from 0.69 to 0.78. These results suggest that transferring a reusable healthy-domain EMG representation can reduce calibration burden while improving robustness for real-time post-stroke intent detection.

Figures (2)

• Figure 1: Stroke data collection setup. a participant wears a Myo armband on the paretic forearm and the MyHand orthosis while a graphical user interface presents timed movement cues and records synchronized labeled EMG.
• Figure 2: Convergence comparison for participant S2. Average transition accuracy on the five held-out stroke test sets versus training epoch for stroke-only training and three healthy-initialized adaptation strategies. Dotted lines denote zero-shot performance of the frozen healthy-pretrained model on stroke (brown) and on the held-out able-bodied evaluation set (purple). “Healthy retention” reports performance of the adapted checkpoint on the able-bodied evaluation set.