Technical Summary: ReactEMG Stroke

Problem Statement

Surface electromyography (sEMG) offers a promising non-invasive signal for controlling assistive hand orthoses in post-stroke rehabilitation, enabling "assist-as-needed" systems that respond to a patient's voluntary intent. However, deploying sEMG-based intent detection for stroke survivors faces a critical bottleneck: subject-specific calibration is lengthy, data-hungry, and brittle.

Signals from paretic (affected) muscles differ fundamentally from those of able-bodied users due to weakness, reduced voluntary drive, spasticity, and abnormal synergies (e.g., co-contraction). Consequently, models trained on stroke data often require large amounts of labeled data to be usable, yet they remain fragile to real-world variability such as posture changes, sensor repositioning, and signal drift within a session. Existing few-shot or adaptive strategies typically initialize models from scratch using only stroke data or use healthy data merely to tune configurations, failing to leverage the rich, reusable representations found in large-scale able-bodied sEMG datasets.

Methodology

The authors propose ReactEMG Stroke, a framework for healthy-to-stroke few-shot adaptation. The core premise is to initialize an intent detector using a model pretrained on large-scale able-bodied sEMG data and then fine-tune it for individual stroke participants using a minimal amount of subject-specific labeled data.

1. Base Model and Data

• Base Model: The authors utilize ReactEMG, an encoder-only transformer pretrained on five public datasets comprising over 650 able-bodied participants. ReactEMG uses a multimodal masked modeling objective, learning to reconstruct both EMG signals and aligned intent labels, thereby capturing robust EMG representations.
• Stroke Dataset: A new dataset was collected from three individuals with chronic stroke. Participants wore a Myo armband on the paretic forearm and a MyHand orthosis. The protocol involved timed sequences of attempted hand opening, closing, and relaxation.
• Evaluation Protocol: To test robustness, models were evaluated on held-out test sets featuring realistic distribution shifts:
  • Within-session drift: Signals collected mid-session and end-of-session.
  • Unseen posture: Arm lifted off the table (hovering posture).
  • Sensor placement: Armband rotated ~15°.
  • Device-driven motion: Orthosis actively opening the hand while the patient attempted to close it.

2. Adaptation Strategies

The study compares three fine-tuning strategies against two baselines under identical data budgets:

• Baselines:
  • Healthy Zero-shot: The frozen pretrained model applied directly to stroke data.
  • Stroke-only Training: A model trained from scratch (random initialization) using only the small stroke dataset.
• Adaptation Strategies:
  • Head-only Tuning: Freezing the pretrained backbone and updating only the classification head.
  • LoRA (Low-Rank Adaptation): Freezing pretrained weights and training low-rank update matrices for linear layers, allowing internal feature adjustment with few parameters.
  • Full Fine-tuning: Updating all model parameters.

3. Metrics

Performance was evaluated using:

• Raw Accuracy: Fraction of timesteps with correct predicted labels.
• Transition Accuracy: A stricter metric reflecting real-time control stability. It measures the correctness of intent transitions (requiring a correct prediction within a reaction buffer) and the maintenance of the new class without "flicker" until the next transition.

Key Contributions

1. First Healthy-to-Stroke Transfer for Intent Detection: The authors present the first framework to adapt a large-scale, healthy-domain EMG foundation model specifically for stroke intent detection.
2. Comprehensive Strategy Comparison: Under controlled data budgets, the paper systematically compares head-only tuning, parameter-efficient LoRA, and full fine-tuning against stroke-only training and zero-shot transfer.
3. Robustness Evaluation: The evaluation includes diverse distribution shifts (posture, sensor placement, drift) that mimic real-world clinical deployment challenges, rather than just static in-distribution testing.

Results

• Performance Gains: Healthy-pretrained adaptation consistently outperformed both zero-shot transfer and stroke-only training. The best adaptation methods (LoRA and Full Fine-tuning) improved average transition accuracy from 0.42 to 0.61 and raw accuracy from 0.69 to 0.78 compared to the stroke-only baseline.
• Strategy Dependence: No single strategy dominated all participants.
  • LoRA and Full Fine-tuning generally yielded the highest performance, particularly for participants with more severe impairment (S1, S2).
  • Head-only tuning performed surprisingly well for the less impaired participant (S3), suggesting that less severe impairment may require less aggressive representation changes.
• Data Efficiency: Significant improvements were observed even with extremely low data budgets (e.g., N=1 or N=4 pairs). For less impaired participants, performance with minimal data approached or surpassed that of full-budget training. However, more impaired participants showed a continued benefit from larger data budgets.
• Convergence: Healthy-initialized models achieved meaningful gains within the first few epochs of fine-tuning, whereas stroke-only models trained from scratch improved more gradually.

Significance and Claims

The paper claims that transferring a reusable healthy-domain EMG representation can substantially reduce the calibration burden for stroke rehabilitation while improving robustness to distribution shifts.

The authors emphasize that while scaling healthy foundation models is feasible due to the abundance of safe, cheap, and large-scale able-bodied data, stroke EMG data remains inherently scarce. Therefore, healthy-to-stroke fine-tuning offers a scalable pathway to translate large model capacity into better intent detection for limited stroke datasets. The results suggest that this approach is a viable strategy to overcome the data bottleneck currently hindering the practical deployment of sEMG-based assistive devices.