SPECTRE: Spectral Pre-training Embeddings with Cylindrical Temporal Rotary Position Encoding for Fine-Grained sEMG-Based Movement Decoding

TL;DR

SPECTRE addresses the challenge of fine-grained sEMG-based movement decoding by introducing a domain-specific self-supervised learning framework that emphasizes physiologically meaningful spectral features and explicitly models forearm sensor topology. It combines a channel-independent CNN embedder, Cylindrical Rotary Position Embedding (CyRoPE), and a spectral pre-training objective based on STFT-derived pseudo-labels (via K-means clustering) to learn robust representations. Ablation studies show that both spectral pre-training and CyRoPE are critical to performance, with SPECTRE achieving state-of-the-art results on multiple datasets, including amputation scenarios, and demonstrating strong transfer with limited labeled data. The work underscores the importance of incorporating domain knowledge in SSL for biosignals and provides a practical foundation for advanced, robust myoelectric interfaces.

Abstract

Decoding fine-grained movement from non-invasive surface Electromyography (sEMG) is a challenge for prosthetic control due to signal non-stationarity and low signal-to-noise ratios. Generic self-supervised learning (SSL) frameworks often yield suboptimal results on sEMG as they attempt to reconstruct noisy raw signals and lack the inductive bias to model the cylindrical topology of electrode arrays. To overcome these limitations, we introduce SPECTRE, a domain-specific SSL framework. SPECTRE features two primary contributions: a physiologically-grounded pre-training task and a novel positional encoding. The pre-training involves masked prediction of discrete pseudo-labels from clustered Short-Time Fourier Transform (STFT) representations, compelling the model to learn robust, physiologically relevant frequency patterns. Additionally, our Cylindrical Rotary Position Embedding (CyRoPE) factorizes embeddings along linear temporal and annular spatial dimensions, explicitly modeling the forearm sensor topology to capture muscle synergies. Evaluations on multiple datasets, including challenging data from individuals with amputation, demonstrate that SPECTRE establishes a new state-of-the-art for movement decoding, significantly outperforming both supervised baselines and generic SSL approaches. Ablation studies validate the critical roles of both spectral pre-training and CyRoPE. SPECTRE provides a robust foundation for practical myoelectric interfaces capable of handling real-world sEMG complexities.

Figures (4)

• Figure 1: Schematic overview of the SPECTRE framework. Raw multi-channel sEMG is first processed by a channel-independent CNN embedding layer to produce spatio-temporal patches . These patches are fed into a Transformer Encoder incorporating CyRoPE. During pre-training (Top Right), a random subset of patches is masked. The encoder processes unmasked patches. The pre-training objective is to predict clustered STFT pseudo-labels corresponding to the masked patches using a prediction head. During fine-tuning, a [kinematics] or kinematic token is often used, and its output representation from the Transformer is passed to an MLP head to predict the downstream task target, e.g., finger joint angles.
• Figure 2: Schematic diagram of CyRoPE. In the two left figures, different colors represent different indices in the time and spatial dimensions. The time dimension and spatial dimension (EMG channel index) are first embedded separately using rotary position embedding, then merged into a single CyRoPE that simultaneously contains both temporal and spatial position information.
• Figure 3: Decoding results for a representative non-amputated subject. Top to bottom: Ground truth finger kinematics; decoded kinematics using SPECTRE model without self-supervised pre-training; decoded kinematics using SPECTRE with self-supervised pre-training.
• Figure 4: Decoding results for a representative individual with amputation (subject 4). Top to bottom: prompted/target kinematics (expected hand movements shown to the subject), against which the decoded movements are compared; decoded kinematics using SPECTRE model with finetuning on labeled dataset; decoded kinematics using SPECTRE model with self-supervised pre-training; decoded kinematics using SPECTRE model with self-supervised pre-training and fine-tuning.