DualMyo: Transformer-based sEMG-to-Digit Classification for Handwriting Decoding

Overview and motivation for sEMG handwriting decoding

Surface electromyography (sEMG) measures myoelectric activity from muscles noninvasively and is widely used for human–machine interaction. Existing sEMG applications commonly target coarse motor actions such as grasping or wrist rotation. However, decoding fine motor behaviors, notably handwriting and digit generation, remains challenging due to complex spatio-temporal patterns in the signals and variability across recording sessions.

The work summarized here frames handwriting decoding from sEMG as an opportunity to improve prosthetic control and communications interfaces. Instead of relying on classical signal-processing pipelines, the authors explore a deep-learning alternative that treats multi-channel sEMG as a structured time series and seeks to learn richer representations suitable for digit classification.

DualMyo model architecture and key components

The proposed model, named DualMyo, is a multi-channel, dual-stream architecture based on the Transformer paradigm. Key architectural elements reported in the source include:

• Patch Embeddings: the model divides multi-channel sEMG time series into local patches and converts these patches into learned embeddings to provide compact representations of local spatio-temporal patterns.
• Rotary Positional Embeddings (RoPE): RoPE are incorporated to encode relative positional information within the time series, helping the model capture temporal dependencies across patches.
• Dual-stream design: the architecture processes multi-channel inputs in a way intended to capture both intra-channel and cross-channel dynamics, although detailed layer-by-layer design specifics were not reproduced here beyond the dual-stream and Transformer characterization.

The authors present DualMyo as a specialized Transformer variant tuned for the peculiarities of myoelectric signals, aiming to model complex dynamics across channels and time windows relevant to fine-motor decoding.

Training protocol and experimental setup

The source reports experiments that evaluated DualMyo on an EMG-to-digit classification task corresponding to handwriting decoding. Details provided include intra-session and cross-session evaluations and a fine-tuning protocol to adapt to session variability. Exact dataset sizes, number of subjects, preprocessing steps, hyperparameters beyond the described fine-tuning regimen, and other experimental specifics were not fully enumerated in the source summary and therefore are not restated here.

Intra-session performance results

According to the authors, DualMyo demonstrated strong intra-session performance when trained and tested within the same recording session. This suggests that the Transformer-based approach with Patch Embeddings and RoPE can effectively learn discriminative representations for digit classification from multi-channel sEMG when session conditions remain stable.

Cross-session challenges: drift and sensor displacement

The authors emphasize the practical obstacles that arise when applying sEMG decoders across sessions. Two major sources of performance degradation identified are:

• Signal drift: baseline shifts and gradual changes in recorded myoelectric signals over time.
• Sensor displacement: shifts in electrode placement between sessions that alter channel-specific signal characteristics.

These factors reduce cross-session generalization and are central problems for real-world deployment of sEMG-based handwriting decoders.

Lightweight fine-tuning strategy and cross-session adaptation

To mitigate cross-session variability, the authors tested a lightweight fine-tuning approach. Reported findings include:

• Fine-tuning DualMyo for 10 epochs using a small calibration set (two examples per digit) enables the model to adapt to session-specific changes.
• With this minimal per-session calibration, the model reportedly achieved approximately 91% accuracy, indicating substantial recovery of performance across sessions with limited new data and brief adaptation.

The source frames this approach as a practical trade-off: modest additional calibration can compensate for drift and sensor displacement without retraining from scratch.

Limitations, implications, and next steps

The authors note limitations and necessary future work explicitly. Key points reported in the source:

• The study is presented as a preprint and has not undergone peer review.
• While intra-session performance is strong and lightweight fine-tuning restores cross-session accuracy in the reported experiments, further validation is required for real-time operation, multi-subject generalization, and closed-loop neuromuscular control.
• The authors position DualMyo as a promising step toward adaptive sEMG-based handwriting interfaces but underscore that scaling to live, multi-user prosthetic control or communication systems will require additional study.

Data and code availability

The authors provided a link to a public repository for data and code in the source record (GitHub reference). The preprint also includes a DOI and indicates the work is shared under a CC-BY 4.0 license. Specifics on dataset composition, subject numbers, and full training hyperparameters were not exhaustively listed in the source summary and would need to be obtained from the full manuscript or the linked repository.