TL;DR
Foundation models for surface electromyography (sEMG) promise to revolutionize how we decode muscle activity, but do they actually solve the field's biggest hurdle? This systematic evaluation of three leading models reveals a sobering reality: while pretraining provides a significant boost for the same user, the benefits largely vanish when generalizing to new subjects, where classical machine learning baselines often remain competitive.
What Problem Does This Paper Address?
Surface electromyography (sEMG) is the "wild west" of biosignals. Because signals are influenced by everything from a person's unique anatomy to the exact millimeter an electrode is placed, a model trained on Subject A rarely works for Subject B. This is known as the cross-subject generalization problem.
Recently, "Foundation Models"—large transformers pretrained on massive amounts of unlabeled data—have emerged as a potential solution. The hope is that by seeing millions of muscle signal patterns, these models will learn a universal "language" of movement. However, the authors argue that current evaluations are inconsistent. Many papers report "intra-session" results (testing on the same person they trained on), which doesn't reflect real-world clinical or prosthetic needs. This paper sets out to provide a rigorous "reality check" on whether pretraining actually helps models generalize to unseen humans.
Methods
The researchers conducted a systematic head-to-head comparison of three sEMG foundation models: TinyMyo, PhysioWave, and NeuroRVQ.
To isolate whether the performance comes from the pretraining or just the model size, they employed three distinct fine-tuning strategies:
- Full Fine-tuning: Starting with pretrained weights and updating the whole model.
- From-Scratch: Using the same architecture but starting with random weights.
- Linear Probing: Freezing the pretrained model and only training a simple classification layer on top.
Crucially, they also standardized the preprocessing pipelines. They discovered that matching the input frequency to the model's original pretraining frequency (resampling) was vital for performance, as shown in Figure 2.
Datasets
The models were evaluated across three diverse gesture classification datasets:
- NinaPro DB5: A complex dataset with 10 subjects and 52 different gestures.
- EPN612: A large-scale dataset featuring 612 subjects but only 6 gestures.
- UCI EMG Gestures: A mid-sized dataset with 36 subjects and 6 gestures.
The authors tested these under two split strategies: Intra-Session (training and testing on the same person) and Inter-Subject (training on one group and testing on entirely new people).
Key Insights and Ideas
- The Generalization Gap: Pretraining is a "fair-weather friend." It consistently beats training from scratch when you stay within the same recording session. However, in the much harder inter-subject task, that advantage nearly disappears.
- Classical ML is Still King in Low-Data Regimes: On smaller datasets like NinaPro DB5, simple algorithms like Linear Discriminant Analysis (LDA) and Support Vector Machines (SVM) matched or even outperformed the massive foundation models in cross-subject tests.
- Reconstruction vs. Invariance: Most current sEMG models use "masked reconstruction" (trying to fill in missing parts of a signal) for pretraining. The authors suggest this might be the wrong goal: it teaches the model to represent the signal perfectly, including all the subject-specific noise and anatomical quirks that we actually want to ignore.
- Preprocessing is a Silent Killer: Inconsistencies in how data is normalized and resampled can lead to misleading results. The authors found that global z-score or min-max normalization often outperformed the per-channel normalization used in some original papers.
Key Results
The quantitative results highlight a stark contrast between settings:
- Intra-Session: On the large EPN-612 dataset, NeuroRVQ and PhysioWave achieved impressive accuracies of ~97.5%, significantly outperforming the "from-scratch" versions (~93-97%).
- Inter-Subject: When moving to new subjects on the same dataset, the accuracy for NeuroRVQ dropped to 93.26%. While pretraining still helped here, the gain over training from scratch was only about 2.3%.
- The NinaPro "Floor": On the complex NinaPro DB5 dataset (inter-subject), the foundation models struggled immensely, with NeuroRVQ hitting only 23.58% accuracy—essentially the same as a simple LDA baseline (23.63%).
Limitations and Future Work
The primary limitation identified is that current pretraining objectives do not incentivize subject-invariant representations. The models are too good at learning the "fingerprint" of a specific subject's muscles, which leads to overfitting.
The authors suggest three paths forward:
- New Pretraining Objectives: Moving beyond reconstruction toward objectives that explicitly force the model to ignore subject-specific features.
- Better Fine-tuning: Developing strategies that can leverage pretrained knowledge without overfitting to the small number of subjects available in downstream tasks.
- Rigorous Benchmarking: Future sEMG research must prioritize inter-subject evaluation and always include "from-scratch" baselines to prove that pretraining is actually adding value.
Summary updated