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

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.

Method

We start with ReactEMG, an encoder-only transformer that jointly embeds 8-channel sEMG and intent labels with a masked-modeling objective, producing per-timestep gesture predictions without per-user calibration. Pretrained on more than 650 able-bodied participants, it serves as our EMG foundation model, which we adapt to each stroke participant from a small dataset of cued open/close attempts collected while wearing the MyHand orthosis. We compare three fine-tuning strategies — head-only, LoRA, and full fine-tuning — against two baselines: a frozen healthy model and stroke-only training from scratch, all under the same data budget.

ReactEMG Stroke pipeline: healthy pretraining followed by per-subject fine-tuning with head-only, LoRA, or full strategies

Intent Detection Performance

Across three chronic-stroke participants and five held-out test sets, the strongest fine-tuning strategy lift average raw accuracy from 0.69 to 0.78 and transition accuracy from 0.42 to 0.61, with consistent gains across all three fine-tuning strategies.

Data Efficiency

Most of the benefit with fine-tuning is unlocked at very low data budgets. Just a handful of labeled open/close attempts per subject is enough to deliver meaningful per-subject improvement, making the approach practical to deploy in a clinical session.