Hybrid sEMG decomposition method integrating gradient convolution kernel compensation and correlation-constrained FastICA

Surface electromyography (sEMG) decomposition is a critical technique for decoding spinal motor neuron control strategies. However, traditional decomposition methods, including the gradient convolution kernel compensation (gCKC) and the integrated FastICA-CKC framework, often suffer from degraded performance under moderate-to-high contraction forces or low signal-to-noise ratios (SNRs). This study proposes a hybrid decomposition framework that integrates gCKC and correlation-constrained FastICA to enhance the precision of high-density sEMG decomposition through a two-stage optimization strategy. First, the gCKC algorithm is used to preliminarily estimate the motor unit spike trains (MUSTs). Subsequently, correlation-constrained FastICA is used to refine the initial results by correcting erroneous or missed spikes caused by noise or superimposition of motor unit action potentials (MUAPs). On simulated signals (5 contraction levels × 5 SNRs), the proposed method demonstrated significant improvements at 40% and 50% of maximum voluntary contraction (MVC) when compared to multiple benchmarks, including the standalone gCKC, the FastICA-CKC, and the state-of-the-art 2CFastICA algorithms. For instance, at 50% MVC, the Rate of Agreement (RoA) (96.9% ± 1.8%) showed an average increase of 8.2% and 2.3% compared to using gCKC alone (88.7% ± 2.1%) and the FastICA-CKC method (94.6% ± 3.5%), respectively. Similarly, sensitivity (98.9% ± 0.6%) and precision (97.7% ± 1.3%) were improved by 2.9% and 6.1% relative to standalone gCKC, and by 0.4% and 2.0% relative to FastICA-CKC. The proposed method also achieved a significantly higher decomposition yield (number of identified units) than 2CFastICA, while maintaining comparable sensitivity and precision. In experimental tests, the motor units identified by the proposed method achieved an average pulse-to-noise ratio (PNR) of 23.2 ± 0.8 dB and silhouette distance (SIL) of 96.16% ± 0.43%, exhibiting superior overall performance against all compared algorithms. Statistical tests confirmed that these performance differences were highly significant. The results indicate that the proposed hybrid framework achieves a more favorable balance between “decomposition yield” and “per-unit accuracy.” The integration of convolutional modeling and correlation-constrained FastICA can more effectively mitigate interference from MUST superposition during moderate-to-high force contractions than existing benchmark methods, offering a more robust decomposition scheme for neural-machine interfaces and motor control research.