Deep Neural Networks for Learning Intent from sEMG Signals to Support Hardware Devices for Post-Stroke Neurorehabilitation

Zakariyya Brewster; Divy Wadhwani; Emily Yan; Aidan Wang; Karma Namgyal; Shuting Xie; Markiyan Konyk; Tala Abdelmaguid

Source-linked AI summary

Deep Neural Networks for Learning Intent from sEMG Signals to Support Hardware Devices for Post-Stroke Neurorehabilitation

Zakariyya Brewster, Divy Wadhwani, Emily Yan, Aidan Wang, Karma Namgyal, Shuting Xie, Markiyan Konyk, Tala Abdelmaguid

arXiv:2609.09971v1Sep 9, 2026cs.LG

TL;DR

The paper addresses five-finger intent decoding from impaired-arm sEMG for post-stroke rehabilitation hardware. It evaluates shared preprocessing and multiple neural architectures, then develops a compact four-channel CNN with cross-channel distillation. The distilled model reaches 0.5219 ± 0.0114 subset accuracy and 0.6095 ± 0.0058 macro F1 across five seeds, providing a software foundation for later hardware-in-the-loop evaluation.

Problem
Post-stroke rehabilitation needs reliable five-finger intent estimates from residual sEMG despite impairment-related signal variability and hardware constraints.
Method
The study evaluates LSTM, CNN, and GNN decoders under shared preprocessing, searches compact CNN architectures, and retrains CNN-Micro for four active channels with cross-channel distillation.
Results
0.5219 ± 0.0114 subset accuracy and 0.6095 ± 0.0058 macro F1 were achieved by the distilled four-channel student across five seeds.
Takeaways & Limitations
The selected 123K-parameter model provides a compact four-sensor software path for subsequent Raspberry Pi 5 timing and hardware-in-the-loop rehabilitation studies.
Takeaways & Limitations
The four-channel comparison retains one patient split, uses electrode maps inferred from intended placement, and evaluates decoding rather than therapeutic benefit.

Abstract

from arXiv · show

Finger-specific motor intent is a clinically meaningful control signal for post-stroke neurorehabilitation, where residual muscle activity may remain measurable despite weak or incomplete movement. We study five-finger multilabel intent decoding from impaired-arm high-density surface electromyography (sEMG) in PhysioMio, a bilateral longitudinal dataset collected from stroke patients. A common processing protocol aligns movement labels, applies 20--450 Hz Butterworth filtering and Symlet-4 wavelet denoising, segments overlapping 200 ms windows, and extracts twelve time- and frequency-domain descriptors per channel. Direct LSTM, CNN, and GNN baselines reveal complementary behavior: the LSTM attains the highest subset accuracy (0.545), whereas the GNN attains the highest macro F1 (0.706) and macro AUPRC (0.776). Architecture search then identifies CNN-Large as the strongest single-split CNN, with 0.593 subset accuracy and 0.714 macro F1, while CNN-Micro provides a compact architecture for embedded inference. To match a four-sensor hardware design, we retrain CNN-Micro using channels associated with ECRB, ECRL, FDS, and FDP and exclude the ground electrode from model input. Across five seeds, cross-channel knowledge distillation improves the four-channel student over direct training, reaching $0.5219 \pm 0.0114$ subset accuracy, $0.7612 \pm 0.0038$ finger accuracy, and $0.6095 \pm 0.0058$ macro F1. The selected 123K-parameter model accepts nine windows of 48 features and has been exported to ONNX. These results establish a reproducible software path from post-stroke sEMG to compact five-finger intent prediction for subsequent hardware-in-the-loop evaluation.

1 Introduction

The paper frames five-finger intent decoding from impaired-arm sEMG as a hardware-aware problem for post-stroke rehabilitation. It evaluates model families, transfer strategies, and compact four-channel designs while balancing prediction quality with deployability.

Reliable finger-intent decoding is needed to estimate intended movement from residual sEMG when visible post-stroke movement is weak or incomplete.
Post-stroke multilabel decoding must address abnormal co-contraction, altered recruitment, fatigue, sensor placement, impairment severity, and multiple simultaneous finger activations.
The study evaluates a hardware-aware pipeline spanning label alignment, signal processing, model-family comparison, transfer learning, and compact CNN selection for Raspberry Pi 5 rehabilitation hardware.
The experiments compare LSTM, CNN, and GNN inductive biases, assess architecture search and transfer learning, and test whether four active channels retain useful performance through cross-channel distillation.

2 Related Work

Related work shows that stroke-specific sEMG decoding remains heterogeneous across populations, sensors, labels, and protocols. Transfer learning and deployment-oriented optimization motivate evaluating accuracy together with robustness, model size, latency, and hardware constraints.

Ninapro is a dominant healthy-sEMG benchmark family, whereas PhysioMio provides stroke-specific bilateral and longitudinal high-density recordings.
Stroke-specific decoding studies use heterogeneous tasks and settings, with reported performance varying across impairment levels, gesture sets, sensors, and classifiers.
Recurrent, convolutional, and graph models provide complementary temporal, local-motif, and relational representations for sEMG decoding.
Healthy-to-stroke transfer has produced large gains in some prior stroke gesture-classification settings, motivating architecture- and task-specific transfer evaluation here.
Knowledge distillation and hyperparameter search connect offline decoding performance to compact models, latency, and hardware-constrained inference.

3 Method

The method converts PhysioMio recordings into patient-split multilabel examples through harmonization, filtering, denoising, overlapping windowing, and shared tensor preparation. The same preprocessing configuration supports comparison across model families.

The pipeline aligns raw sEMG with gesture annotations, segments recordings, extracts features, maps tensors, evaluates model families, and supports deployment analysis.
Each sample is represented by channel-by-window feature tensors and a five-bit target for thumb, index, middle, ring, and little finger activation.
The loader groups contiguous movement labels, removes segments shorter than 200 raw samples, and requires complete 200 ms windows without zero-padding.
Patients are assigned deterministic 70/10/20 train, validation, and test partitions, preventing samples from one patient from crossing partitions.
Signals receive fourth-order 20–450 Hz Butterworth filtering, Symlet-4 wavelet soft-threshold denoising, and 200 ms windows with 50% overlap.

3.4 Window-level feature representation

The method summarizes each channel-window pair with twelve handcrafted time- and frequency-domain descriptors, then formats shared sequences for recurrent, convolutional, and graph predictors. Reduced-channel CNNs operate on 48 features per window and exclude the reference electrode.

Each channel-window pair produces twelve handcrafted descriptors spanning time and frequency domains, yielding tensors shaped (C, W, F) with F = 12.
The representation reduces input dimensionality and exposes standard sEMG descriptors to compact students, while raw end-to-end learning remains an untested alternative.
A shared adapter concatenates channel features per window into sequences for common recurrent, convolutional, and graph-model inputs.
The high-density CNN treats 64 × 12 features as 768 temporal-convolution input channels, while reduced inputs use 4 × 12 = 48 features per window.
The hardware-targeted model uses ECRB, ECRL, FDS, and FDP signals, excludes the ground electrode, and receives nine consecutive windows with 48 features each.

3.7 Predictor families

The predictor families impose different structural assumptions on the same feature representation: recurrence accumulates temporal evidence, graphs relate windows, and CNNs model temporal patterns compactly.

LSTM: The LSTM uses two stacked recurrent layers and classifies from the final hidden state of the second layer.This design emphasizes temporal accumulation across windows.
GNN: The GNN represents windows as graph nodes and applies neighborhood aggregation, graph convolution, and global mean pooling before multilabel readout.Its complete undirected graph connects all windows within each sample, injecting a structural prior beyond a flat sequence.
CNN: The CNN students use a 1 × 1 projection, temporal convolution blocks, adaptive pooling, and a compact fully connected head.The family spans approximately 54K to 1.62M parameters, with INT8 footprints from 53 KB to 1.58 MB.
CNN: The CNN family ranges from Nano to XLarge, while the teacher family uses 1D ResNet variants with bottleneck residual blocks.This establishes compact students alongside substantially deeper teacher references.

3.8 Training, tuning, and evaluation protocol

Training uses a shared multilabel objective and reporting framework, while tuning balances validation macro F1 against latency and supports teacher-guided distillation for compact students.

Training: Binary cross-entropy with logits trains the five-label predictors alongside Adam optimization, learning-rate reduction, early stopping, and checkpointing.The objective is defined for logits z ∈ R5.
Evaluation: The protocol reports exact-match subset accuracy, finger-level accuracy, macro classification metrics, macro AUROC, and macro AUPRC.Per-finger scores and task-level curve data support aggregate and finger-wise analysis.
Tuning: Optuna architecture search evaluates CNN students using validation macro F1 and latency relative to a size-specific baseline.Trials at or below baseline latency are unpenalized, whereas slower trials receive increasing penalties and pruning.
Distillation: Distillation blends hard labels with temperature-scaled teacher targets, using paired full-representation teacher inputs and reduced student inputs.The four-channel experiment uses T = 2 and α = 0.5, with thresholds tuned on validation probabilities.

3.9 Transfer-learning paths

The study evaluates transfer across subjects’ arm condition and across input density, using pretrained models or sliced projection weights to initialize impaired-arm and four-channel models.

Healthy-to-impaired transfer: Healthy-to-impaired transfer pretrains CNN and LSTM models on healthy-arm recordings before finetuning them on impaired-arm recordings.The same five-finger multilabel objective is preserved, and the CNN experiment uses the tuned CNN-Base configuration.
Input-density transfer: Input-density transfer slices the first projection-layer weights of a compatible 64-channel CNN-Micro model to initialize the four-channel student.Later shape-compatible layers are copied directly, and this path is evaluated separately from cross-channel distillation.

4 Experiments

Experiments progress from standardized PhysioMio benchmarks to transfer and hardware-constrained four-channel evaluation, with patient-level splits and seed-based final comparisons.

Data and splits: The processing path converts contiguous PhysioMio movement segments into five-bit targets and PyTorch tensors after signal processing and feature extraction.Dataset construction begins from raw parquet recordings grouped by patient.
Data and splits: Patients are split into 70% training, 10% validation, and 20% test partitions, with the main benchmarks using impaired-arm recordings.The four-channel study fixes patient assignments across training modes and includes 34 training, five validation, and nine test patients.
Experiment groups: The experiments compare direct model families, optimized CNN students and ResNet teachers, healthy-to-impaired transfer, and four-channel hardware-constrained training.The four groups cover baseline decoding, model-footprint optimization, transfer mechanisms, and sensor reduction.
Four-channel evaluation: The selected four-channel configuration uses the right channel map and nine windows, then compares direct supervision, sliced transfer, and cross-channel distillation across seeds 0–4.Development runs also compare left, right, and dual channel-map views and context lengths of one, four, and nine windows.
Evaluation protocol: Single-split baseline, CNN-sweep, ResNet, and healthy-transfer tables provide deterministic estimates, whereas final four-channel results are summarized across five seeds.Per-finger thresholds are tuned on validation probabilities, and checkpoint selection uses validation macro F1 rather than test performance.

5 Results

Results show complementary strengths across direct model families, optimized CNNs, transfer learning, and four-channel retraining. CNN-Micro distillation provides the strongest reduced-input performance while preserving a compact hardware-oriented design.

5.2 Optimized CNN students and ResNet teachers: 0.593 subset accuracy and 0.714 macro F1 made CNN-Large the strongest single-split CNN, while CNN-Micro retained 0.578 subset accuracy and 0.697 macro F1 with an estimated 154 KB INT8 footprint.CNN-Micro was selected for hardware-targeted study because it approached CNN-Large using approximately one fifth of its estimated INT8 footprint.
5.3 Healthy-to-impaired transfer: CNN-Base finetuning improved subset accuracy from 0.465 to 0.585, finger accuracy from 0.760 to 0.800, macro AUROC from 0.833 to 0.863, and macro AUPRC from 0.699 to 0.767.The LSTM showed mixed transfer effects, with accuracy increases but decreases in macro F1, AUROC, and AUPRC.
5.4 Four-channel hardware-targeted retraining: 0.5219 ± 0.0114 subset accuracy, 0.7612 ± 0.0038 finger accuracy, and 0.6095 ± 0.0058 macro F1 were achieved by cross-channel distillation in the four-channel study.Distillation improved direct four-channel training by 0.0418 subset accuracy, 0.0287 finger accuracy, and 0.0272 macro F1.
5.4 Four-channel hardware-targeted retraining: The selected four-channel network has 123,317 parameters, accepts a fixed (N, 9, 48) tensor, produces five logits, and was exported to ONNX.The four-channel model trades some offline discrimination for an implementable sensor contract; no on-device timing result is reported.
5.5 Finger-level behavior and external positioning: Per-finger rankings varied: the GNN led thumb, CNN-Large led index and little finger, and the LSTM slightly led middle and ring finger.The passage identifies these biomechanical interpretations as hypotheses requiring targeted validation rather than anatomical evidence.

6 Discussion

The discussion links complementary model behavior, compact CNN design, and cross-channel distillation to a hardware-aware four-sensor decoder, while emphasizing unresolved generalization, timing, and clinical-validation limits.

Model-family comparison: The LSTM leads subset accuracy whereas the GNN leads macro F1 and AUPRC, indicating complementary strengths for complete-output consistency and class-balanced discrimination.These results are consistent with both ordered temporal dynamics and relationships among non-adjacent windows in paretic sEMG.
Architecture search: CNN-Large provides the strongest offline CNN result, while CNN-Micro retains most performance at approximately one fifth of the estimated INT8 footprint.The CNN sweep is non-monotonic: CNN-XLarge performs below CNN-Large despite having more parameters.
Transfer learning: Healthy-to-impaired transfer helps CNN-Base but produces mixed LSTM results, while sliced full-grid CNN-Micro initialization does not outperform training from scratch.The paper attributes the reduced-channel transfer outcome to incomplete preservation of the full model’s spatial basis.
Reduced-channel transfer: Cross-channel distillation improves every reported reduced-input metric over direct training, while the 64-channel source remains stronger by 0.0715 macro F1.The four-channel trade-off exchanges some offline discrimination for an implementable sensor contract.
Deployment boundary: The selected ONNX model establishes interface compatibility rather than real-time hardware performance, because Raspberry Pi latency has not been measured.Current preprocessing reruns filtering and wavelet denoising over rolling history.
Limitations: Generalization remains uncertain because most experiments are single-run estimates, the five-seed four-channel comparison uses one patient split, and therapeutic benefit was not measured.Further work requires patient-level cross-validation, hardware-in-the-loop timing, robustness studies, and evaluation with the target rehabilitation population.

7 Conclusion

The study develops a post-stroke five-finger decoder from impaired-arm sEMG and narrows it to a compact four-channel CNN-Micro design. Cross-channel distillation yields the strongest reduced-input model while preserving a practical ONNX interface.

Scope: The study evaluates progression from model-family comparison through hardware-targeted channel reduction for post-stroke five-finger intent decoding.The conclusion notes complementary temporal behavior across LSTM and GNN baselines and a compact convolutional path from CNN search.
Final result: 0.6095±0.0058 macro F1 and 0.5219±0.0114 subset accuracy are achieved by the distilled four-channel CNN-Micro across five seeds.The model uses ECRB, ECRL, FDS, and FDP features and emits five finger-intent logits.