Source-linked AI summary
EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
Yuwei Jia, Cheng Zhong, Jinyang Yu, Zhe Cui
TL;DR
Public EMG datasets differ in sensor layouts, frequency support, and size, complicating their joint use for pretraining. EMGBlend couples geometry-aware channel processing, recording-supported spectral targets, and balanced source exposure; at fixed capacity and window budget, eleven-source training improves all DB7 settings, while cross-person force decoding remains difficult.
Problem
Public EMG datasets differ in arrays, sampling rates, passbands, and annotations, while naive pooling can conflate sensor identities, assign unobservable targets, and overexpose some sources.
Source: §1 · p. 1Method
EMGBlend combines shared channel patches and geometry-aware attention with supported-band spectral targets and source-exposure correction.
Source: §1 · p. 1Results
At fixed capacity and window budget, eleven-source pretraining improves all DB7 settings; Full improves PiMForce, while cross-person force decoding remains an open challenge.
Source: §4 · p. 4Takeaways & Limitations
Matched controls support each mechanism, and fixed-budget source controls show multi-source pretraining improves gesture recognition while remaining competitive for force decoding.
Source: §4 · p. 4, §3.4 · p. 4The external system comparison does not isolate SSL design because model capacity, pretraining data, and training budget are unmatched.
Source: §3.4 · p. 4
Abstract
from arXiv · showhide
Public surface electromyography (EMG) datasets vary widely in electrode layout, channel count, frequency support, and size. Simply mixing them for pretraining can misalign channel semantics, introduce spectral targets that some devices cannot observe, and let large or high-channel-count datasets dominate learning. We introduce EMGBlend, a self-supervised framework designed around these differences. It combines shared channel patches with geometry-aware attention, restricts spectral targets to each recording's supported frequency band, and balances exposure across data sources. We pretrain a 109M-parameter model on 11 public EMG sources and evaluate it on gesture recognition, continuous-force regression, and contact classification. EMGBlend consistently outperforms matched random initialization and waveform reconstruction controls. Fixed-budget source controls show that multi-source pretraining improves gesture recognition and remains competitive for force decoding. Ablations confirm that geometry, band-aware targets, and source balancing each contribute to transfer, although cross-person NinaPro force estimation remains difficult. Overall, EMGBlend shows how heterogeneous EMG datasets can be combined through explicit mechanism design rather than simple concatenation. Code is available at https://github.com/tamanano/EMGBlend
1. INTRODUCTION
Public EMG datasets differ in sensor layouts, frequency support, and scale, so pooling them can misalign channels, create unobservable spectral targets, and overexpose some sources. EMGBlend couples interface, target, and exposure mechanisms and evaluates the resulting initialization across gesture and force tasks.
1. INTRODUCTION
- Naive pooling can conflate sensor identities, assign targets outside devices’ observable bands, and overexpose sources with more windows or channels.Existing approaches address relevant interface, target, or transfer ingredients, but the paper says they do not jointly address montage, frequency-support, and source-exposure mismatch.Source: §1 · p. 1
- EMGBlend combines shared channel patches with geometry-aware attention, supported-band spectral codes, and source-exposure correction for heterogeneous EMG pooling.The mechanisms require neither shared task labels nor paired pose.Source: §1 · p. 1
- With the same capacity, updates, and number of windows, eleven-source pretraining improves all four DB7 settings over a single source and remains comparable on PiMForce.Evaluations include gesture transfer, frozen gesture recognition, continuous-force and contact decoding, waveform-objective controls, mechanism removals, fixed-budget source controls, and a capacity sweep.Source: §1 · p. 1
2. HETEROGENEITY-AWARE EMG PRETRAINING
EMGBlend pools heterogeneous EMG through a variable-montage interface, recording-valid spectral targets, and source-exposure correction. Its pretraining predicts discrete codes for masked channel patches, while downstream transfer uses gesture, force, and contact-decoding setups.
2. HETEROGENEITY-AWARE EMG PRETRAINING
- EMGBlend couples a montage-flexible interface, recording-valid targets, and source-exposure correction, with task labels used only downstream.Each source may differ in channel count, sampling rate, frequency support, and sensor geometry.Source: §2 · p. 1
2.1. Source-aware pooling across acquisition settings
- The loader tempers channel-patch exposure imbalance across sources, using α = 1/2 in the standard recipe.Channel-count bucketing limits padding, and at most 64 channels are sampled per window; available-pool size and realized training exposure are distinct.Source: §2.1 · pp. 1–2
2.2. Geometry-aware variable-montage encoder
- A shared Conv1D patch encoder and alternating temporal and spatial attention use sensor geometry to support variable montages without requiring shared device identity or anatomical registration.Unknown geometry uses learned embeddings; pairwise-distance attention bias encodes sensor arrangement.Source: §2.2 · p. 2, Figure 1 · p. 2
2.3. Supported-band spectral-code learning
- A vector-quantized tokenizer supplies discrete codes, while spectral reconstruction is restricted to frequencies supported by each recording.Averaging over valid bins avoids downweighting narrower passbands; whole-channel and temporal masking hide embeddings, and the encoder predicts codes at hidden valid positions.Source: §2.3 · p. 2, Figure 2 · p. 2
2.4. Transfer to gestures and force
- Frozen gesture features use regularized linear classifiers, while force decoding maps frozen features to finger targets with ridge regression.Pose-guided gesture adaptation retains query alignment and pose reconstruction; PiMForce contact decisions threshold predictions using validation users.Source: §2.4 · p. 2
3. EXPERIMENTS
The experiments evaluate EMGBlend across gesture recognition, force decoding, and unified controls. Results include gains over matched random initialization and controls, alongside a difficult cross-person force endpoint.
3.1. Training and evaluation settings
- The 462,980-window pretraining pool spans 11 sources, with 2–448 released channels, sampling rates of approximately 157–2,048 Hz, and 752–171,828 windows per source.Channels are capped at 64 when sampled; UCI’s nominal timestamp grid is not used to infer physical bandwidth, and UCI is omitted from transfer comparisons because its downstream subjects appeared in unlabeled pretraining.Source: §3.1 · p. 3
3.2. Gesture recognition and transfer
- 81.76 ± 0.37% seen LP and 61.68 ± 0.30% unseen LP are reported for EMGBlend after pose-guided adaptation.The local EMBridge reproduction reports 76.62% seen LP and 49.37% unseen LP; matched random initialization reports 68.96% and 48.48%, respectively.Source: Table 1 · p. 3, §3.2 · p. 3
3.3. Continuous force decoding
- Within-person DB2/DB3 favor SSL over both controls, while cross-person DB2 remains difficult: SSL R2 is −0.2926 versus −0.2218 for random and −0.0363 for handcrafted features.On PiMForce, SSL reduces cross-user MAE from random’s 2.3802 to 2.1204 (10.92%), versus HC’s 2.6151; MAE includes contact and non-contact windows, while BA/F1 assess thresholded contact state.Source: §3.3 · p. 4
3.4. External baselines and unified controls.
- The system-level comparison uses common downstream splits and metrics, but model capacity, pretraining data, and training budget are not matched, so it does not isolate SSL design.For DB5, the study uses a shared downstream protocol rather than PhysioWave’s original split.Source: §3.4 · p. 4, Table 3 · p. 4
- With the 109M backbone, eleven-source pool, and 60k updates held fixed, Full exceeds each one-mechanism removal on all 11 endpoints.Source-pool controls find that matched-volume multi-source pretraining improves gesture recognition and remains competitive for force decoding; scaling is non-monotonic.Source: §3.4 · p. 4
4. CONCLUSION
EMGBlend combines montage-flexible inputs, supported-band targets, and source-balanced sampling for heterogeneous EMG pretraining. Matched controls support its mechanisms, and eleven-source training improves all DB7 settings at fixed capacity and window budget, though cross-person force decoding remains challenging.
4. CONCLUSION
- At fixed capacity and window budget, eleven-source pretraining improves all DB7 settings.Matched controls support each mechanism.Source: §4 · p. 4
- EMGBlend couples montage-flexible inputs, supported-band targets, and source-balanced sampling for heterogeneous EMG pretraining.Source: §4 · p. 4
- Full improves PiMForce, while cross-person force decoding remains an open challenge.Source: §4 · p. 4