Source-linked AI summary
Comparing Temporal Deep Learning Architectures for Discrete Cross-User Myoelectric Control
Ethan Eddy, Evan Campbell, Scott Bateman, Erik Scheme
TL;DR
Prior work showed that discrete cross-user myoelectric control without recalibration is feasible, but optimization and generalization to randomly prompted continuous data remained underexplored. The paper benchmarks temporal architectures and feature extraction approaches, then evaluates them on a continuous gesture collection. GRU models performed best overall, with a two-layer CNN-GRU configuration reaching 96.1% on MyoDisCo.
Problem
Although prior work demonstrated discrete cross-user control without recalibration, optimization approaches and generalization to randomly prompted data remained little evaluated.
Source: §I · p. 1Method
The study benchmarks five temporal deep-learning architectures and handcrafted versus CNN-based feature extraction on EMG-EPN, then evaluates models on continuous MyoDisCo data.
Source: §I · p. 1Results
GRU performed best across the evaluated architectures, and its two-layer CNN configuration reached 96.1% on MyoDisCo, with no significant differences among the models’ best configurations.
Source: §IV · p. 3, B. Continuous Stream Evaluation · p. 3Takeaways & Limitations
For embedded control systems, model size matters because larger architectures used substantially more parameters without necessarily improving generalizability to MyoDisCo.
Source: §IV · p. 4The authors note that EMG-EPN gesture segmentation is unknown, and minor differences in input-template framing may affect model robustness.
Source: §IV · p. 4
Abstract
from arXiv · showhide
The recognition of dynamic gestures via discrete myoelectric control is becoming an attractive input modality for event-based human-machine interaction. While recent work suggests that discrete cross-user models are possible (i.e., no calibration data is required from the end user), few have evaluated techniques to optimize these systems and their generalizability to new randomly prompted data simulating real-world discrete control is relatively unknown. Using the 612-user EMG-EPN dataset, this work benchmarks five temporal deep-learning approaches, evaluates the tradeoff between performance and model complexity, and compares hand-crafted and machine-learned feature extraction techniques. The CNN-GRU achieved the best accuracy of 93% when tested on 184 previously unseen users. These models were then evaluated on a different, more realistic, unsegmented dataset, wherein users were randomly prompted to elicit discrete gestures during a continuous data collection. This dataset was recorded five years apart and in a separate country from the data used to train the original model. Results show that the cross-user results scale to this new dataset (accuracies of 96%) even though no feedback was given to users during the elicitation of gestures. These findings highlight the exciting potential of discrete zero-shot cross-user myoelectric control and offer valuable insights for the continued advancement of this promising technology.
I. INTRODUCTION
Discrete myoelectric control maps dynamic gestures to event-based commands, a potential fit for human-machine interaction. This work benchmarks temporal models and tests their generalization on randomly prompted continuous data.
I. INTRODUCTION
- Personalized end-user calibration has hindered myoelectric control’s wider use for human-machine interaction despite its history of success with powered prostheses.The EMG signal consists of electrical potentials generated by contracting muscles.Source: §I · p. 1
- Discrete control maps dynamic gestures, such as a double tap or finger snap, to event-based commands rather than making decisions on small data windows.The introduction notes that recent work suggests gesture temporal structure may help address the challenge of cross-user models without end-user training data.Source: §I · p. 1
- The study benchmarks five temporal deep-learning architectures and compares handcrafted with CNN-based feature extraction, then evaluates performance on randomly prompted continuous data.The MyoDisCo evaluation requires gesture detection and segmentation from the continuous stream.Source: §I · p. 1
- Although prior work demonstrated cross-user discrete control without recalibration, optimization techniques and generalization to randomly prompted data remained little evaluated.The paper reports accuracies up to 96% on unseen users in data collected years later, in a different country, with randomly prompted gestures.Source: §I · p. 1
II. METHODS
The study uses two datasets recorded nearly five years apart with the same eight-channel Myo Armband setup. EMG-EPN-612 supplies prompted, labeled gesture repetitions from 612 users.
A. Datasets
- Both datasets were recorded with an eight-channel Myo Armband at 200 Hz, placed approximately alike, and share six gestures including rest.The common gestures are hand open, hand close, wrist flexion, wrist extension, double tap, and rest.Source: A. Datasets · p. 1
1) EMG-EPN-612:
- EMG-EPN-612 contains recordings from 612 users, with gestures prompted within five-second windows and start/end indices included in metadata.Labeled data include 50 repetitions per gesture from 306 users and 25 repetitions from the other 306.Source: 1) EMG-EPN-612 · p. 1
2) MyoDisCo:
MyoDisCo’s continuous gesture collection provides a more realistic test of discrete models, since gestures must be found within ongoing EMG data.
2) MyoDisCo:
- Participants were randomly prompted every 3–5 seconds to perform discrete gestures during a continuous collection, with no feedback.The collection was intended to benchmark models when active gestures must be segmented from the continuous EMG stream.Source: 2) MyoDisCo · p. 2
- The collection was a previously unanalyzed and unreleased part of MyoDisCo, initially recorded to study confounding factors in discrete myoelectric control.The passage identifies this continuous gesture collection as a new part of the dataset.Source: 2) MyoDisCo · p. 2
- Figure 1 shows a 45-second example of raw EMG recorded from one subject during the continuous collection.The figure illustrates the type of ongoing signal from which prompted gestures must be identified.Source: Figure 1 · p. 2, 2) MyoDisCo · p. 2
B. Classification
The classification pipeline compares handcrafted wavelet features with CNN-learned features and feeds gesture templates into temporal architectures. Training uses EMG-EPN data, with no additional tuning or training for MyoDisCo evaluation.
B. Classification
- Gesture data were windowed at 50 ms with 25 ms increments, then represented by wavelet energy or three 1D convolutional layers applied to raw windows.Templates were padded to constant size before their final-window encodings were sent to the classifier.Source: B. Classification · p. 2
- The classification network used three linear layers before a six-neuron output layer, with Adam, cross-entropy loss, gradient clipping, and 50 training epochs.The model used an initial learning rate of 0.001 and a batch size of 1000.Source: B. Classification · p. 2
- Training used 80% of labeled EMG-EPN data from 428 users and validation used the remaining 20% from 128 users; MyoDisCo received no additional tuning or training.The passage reports the training and validation split by users.Source: B. Classification · p. 2
1) Algorithms:
- The study evaluated LSTM, bidirectional LSTM, GRU, and RNN architectures alongside a Transformer, with the RNN using ReLU between layers.The Transformer used a 128-dimensional model, 512-dimensional feedforward layer, eight attention heads, and learned positional embeddings.Source: 1) Algorithms · p. 2
C. Gesture Segmentation
Gesture segmentation extracts candidate gestures from continuous EMG using energy thresholds, while statistical comparisons use nonparametric tests.
C. Gesture Segmentation
- Teager-Kaiser energy is averaged across EMG channels and smoothed before thresholding identifies gesture-ready periods, onsets, and gesture ends.The process waits for energy below tready, detects onset above tonset, then waits for a peak and local minimum before classification.Source: C. Gesture Segmentation · p. 2
D. Statistical Analysis
- Because the data were non-normal, factor differences were tested with Friedman tests and posthoc paired Nemenyi tests.Wilcoxon tests were used when comparisons involved fewer than three groups.Source: D. Statistical Analysis · p. 2
III. RESULTS
The results examine classification accuracy across model, feature, and layer choices. Figure 2 summarizes offline EMG-EPN results and parameter counts.
III. RESULTS
- Classification accuracy is evaluated against MODEL, FEATURE, and LAYERS as independent variables.Figure 2 plots parameter-count distributions across model configurations and, for other measures, per-subject distributions averaged across configurations; stars mark best-configuration average performance by condition.Source: A. Offline Results · p. 2, Figure 2 · p. 3
1) Model:
Model choice affected accuracy, with GRU achieving the highest reported best-configuration accuracy. The architectures also differed substantially in parameter counts.
1) Model:
- Across layers and features, GRU averaged 92.9%, versus 89.8% for RNN and 92.1% for BILSTM.MODEL had a significant effect (χ2(4) = 248, p=2.2 × 10−16); pairwise differences were significant except where the passage specifies Transformer versus BILSTM at p=0.023.Source: 1) Model · p. 3
- 93.5% was the GRU best-configuration accuracy (7 layers, CNN features), significantly above RNN at 91.9%.No significant differences were found for BILSTM or Transformer versus other models in the best-configuration comparison.Source: 1) Model · p. 3
2) Parameters:
- Average parameter counts ranged from 140916 for RNN to 1265780 for BILSTM, with GRU, LSTM, and Transformer between them.Ascending order was RNN (140916), GRU (360257), LSTM (469927), Transformer (804033), and BILSTM (1265780).Source: 2) Parameters · p. 3
3) Layers:
Layer count and feature extraction were evaluated for their effects on accuracy; increasing layers mattered in averaged comparisons, while feature methods were not significantly different.
3) Layers:
- 92.1% accuracy at 5 layers exceeded 90.6% at 1 layer in averaged comparisons, but best configurations showed no significant layer-count effect.The averaged layer effect was significant (χ2(4) = 215, p=2.2 × 10−16); the best-configuration comparison was not (χ2(4)=4, p=0.34).Source: 3) Layers · p. 3
4) Feature Extraction:
- 91.9% for CNN features and 91.4% for handcrafted WENG features were not significantly different (p=0.052).Best configurations with handcrafted and machine-learned features also did not differ significantly (p=0.29).Source: 4) Feature Extraction · p. 3
B. Continuous Stream Evaluation
On MyoDisCo, optimized models achieved similar detection performance, with the GRU’s best configuration reaching 96.1%. Gesture detection had low false-positive and false-negative rates, with a 700 ms average prompt-to-prediction delay.
B. Continuous Stream Evaluation
- Averaged across layers and features, model effects were significant, with pairwise differences involving GRU, BILSTM, LSTM, RNN, and Transformer.Reported pairwise differences included BILSTM (92.7%) versus GRU (94.4%), and GRU versus RNN (90.4%) and Transformer (88.5%).Source: B. Continuous Stream Evaluation · p. 3
- Gesture detection had a 1.7% false-positive rate, a 0.3% false-negative rate, and a 700 ms average prompt-to-prediction delay.The delay includes gesture elicitation.Source: B. Continuous Stream Evaluation · p. 3
- 96.1% was the GRU’s best-configuration accuracy, and best configurations did not differ significantly; Figure 3 marks each model’s average across subjects for its best configuration.The GRU configuration used two layers and CNN features.Source: Figure 3 · p. 3, B. Continuous Stream Evaluation · p. 3
IV. DISCUSSION
The discussion reports that GRU performed best overall, while differences between optimized model configurations were less pronounced. It also highlights segmentation uncertainty and model complexity as considerations for generalization and deployment.
IV. DISCUSSION
- GRU performed best overall, followed by BILSTM, LSTM, Transformer, and RNN, but differences were less pronounced for optimized configurations.The discussion suggests all proposed architectures are viable when configured correctly.Source: §IV · pp. 3–4
- Unknown EPN gesture segmentation may help explain cross-dataset differences, and minor changes in input-template framing may affect robustness.Source: §IV · p. 4
- BILSTM and Transformer require substantially more parameters; CNN feature extraction did not significantly improve performance, although the highest-performing MyoDisCo models used CNNs.The discussion notes that embedded deployment makes model size a design consideration.Source: §IV · p. 4