Predicting the Punching Force in Wushu Sanda After Neuromuscular Electrical Stimulation by Employing the KAN Neural Network Combined with Neuromuscular Electricity

Guixian Wang, Haojie Li, Lei HuangSensors 2025 · DOI

TL;DR

This study explores how neuromuscular electrical stimulation (NMES) affects punching force in Wushu Sanda athletes, finding that lower-limb NMES significantly boosts performance. They then introduce a novel application of the Kolmogorov-Arnold Network (KAN) to predict punching force based on surface electromyography (sEMG) signals, demonstrating its potential for real-time monitoring in sports training.

What Problem Does This Paper Address?

Wushu Sanda, a high-intensity combat sport, relies heavily on punching force. While neuromuscular electrical stimulation (NMES) is known to enhance muscle strength and power, its specific effects on complex movements like Sanda punches, and particularly the optimal stimulation sites, are not fully understood. Traditional methods for measuring force (like force plates) are limited in dynamic training environments. Furthermore, predicting punching force from physiological signals like sEMG is challenging due to the complex, non-linear relationship between these signals and the actual force output. Existing linear models struggle to capture these nuances, necessitating more advanced predictive frameworks.

Methods

The researchers employed a randomized repeated-measures design with 30 Wushu Sanda athletes. Participants underwent three conditions: upper-limb NMES (U-NMES), lower-limb NMES (L-NMES), and a sham stimulation control. During these sessions, surface electromyography (sEMG) signals from relevant muscle groups and punching force parameters were collected.

The core innovation lies in the application of a Kolmogorov-Arnold Network (KAN) for predicting punching force. Unlike traditional models that rely on manually extracted features, this KAN architecture, inspired by the TimeKAN model, directly processes raw sEMG time-series data. It uses a Depthwise Convolution layer to capture local temporal structures, followed by Multi-order KANs to model complex, non-linear relationships between sEMG patterns and force outcomes. The network was trained using Mean Squared Error (MSE) as the loss function and the Adam optimizer. For comparison, Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks were also trained and evaluated.

Datasets

The study utilized a dataset collected from 30 Wushu Sanda athletes. This dataset comprised:

  • Surface electromyography (sEMG) signals recorded at 1000 Hz.
  • Punching force parameters measured using the Boxing Power Assessment System (BPAS), which included metrics like relative peak force, impulse, and early explosive force.

This dataset was generated specifically for this study and is not a pre-existing benchmark.

Key Insights and Ideas

  • Lower-limb NMES is a potent enhancer: The study strongly suggests that stimulating the lower limbs (L-NMES) provides a more significant boost to punching force metrics (relative peak force, impulse, early explosive force) than upper-limb stimulation or sham. This highlights the critical role of lower limb power transfer in Sanda.
  • KAN bridges the sEMG-force gap: The KAN model demonstrates a remarkable ability to learn the complex, non-linear relationship between sEMG signals and actual punching force, outperforming traditional RNN and LSTM models. This suggests KAN can decode the underlying neuromechanical determinants of force.
  • sEMG as a diagnostic tool: The research positions sEMG, when analyzed by advanced models like KAN, as a more insightful tool than force sensors alone for understanding neuromuscular readiness and predicting performance, especially in dynamic, real-world training scenarios.
  • NMES primarily impacts peak force, not timing: The findings indicate that NMES effectively increases the magnitude of force output but does not significantly alter the temporal aspects of force production (e.g., time to reach peak force).

Key Results

  • NMES significantly enhances punching force: All NMES conditions (U-NMES and L-NMES) showed significant improvements over sham stimulation for relative peak force, impulse, and early explosive force (p < 0.001).
  • L-NMES superiority: Lower-limb NMES resulted in superior improvements compared to both sham and U-NMES for relative peak force (28.2 ± 3.2 N·kg⁻¹), impulse (16.6 ± 2.3 N·s), and early explosive force (754 ± 94 N).
  • KAN's predictive performance: The KAN model achieved notable predictive accuracy, with R² values reaching up to 0.59 for relative peak force and 0.58 for impulse under L-NMES conditions.
  • KAN outperforms RNN/LSTM: Across all metrics and conditions, KAN consistently showed higher R² values and lower RMSEs compared to RNN and LSTM models. For instance, under L-NMES, KAN achieved an R² of 0.59 for relative peak force, significantly higher than LSTM (0.48) and RNN (0.39).

Limitations and Future Work

  • Moderate predictive accuracy: While promising, the KAN model's predictive accuracy (R² values generally between 0.4 and 0.6) highlights the need for further optimization.
  • Sample size and generalizability: The study was exploratory with 30 participants. Future research with larger sample sizes and diverse athlete populations is needed to enhance generalizability.
  • Optimized model architectures: The authors suggest exploring more advanced KAN architectures and longer-term intervention protocols to improve precision.
  • Ecological validation: Further work is needed to validate the model's performance in more ecologically relevant training environments.
  • Model interpretability: While KAN shows strong predictive power, further research into interpreting its learned relationships could provide deeper biomechanical insights.

Summary updated

Abstract

<b>Objective:</b> To predict the punching force in Wushu Sanda following neuromuscular electrical stimulation (NMES) using the KAN neural network combined with neuromuscular electricity. <b>Methods:</b> Thirty healthy Wushu Sanda athletes underwent a randomized repeated-measures design with three conditions: upper-limb NMES, lower-limb NMES, and Sham stimulation. Surface electromyography (sEMG) signals and punching force parameters were collected. A KAN neural network model was designed to integrate sEMG features and predict punching force. Model performance was evaluated using RMSE, MAE, and R<sup>2</sup> metrics. <b>Results:</b> NMES significantly enhanced punching force metrics (all <i>p</i> < 0.05). lower-limb NMES showed superior improvements in relative peak force (28.2 ± 3.2 N·kg<sup>-1</sup>), impulse (16.6 ± 2.3 N s), and early explosive force (754 ± 94 N) compared to Sham and upper-limb NMES. The KAN model demonstrated robust predictive performance, particularly under lower-limb NMES conditions, with R<sup>2</sup> values of 0.59 for relative peak force and 0.58 for impulse. <b>Conclusions:</b> NMES, especially lower-limb stimulation, effectively boosts punching force. The KAN neural network provides a promising and innovative approach for predicting punching force in Wushu Sanda, providing a foundation for future optimization of real-time monitoring tools.
BibTeX
@inproceedings{wang2025predicting,
  author = {Guixian Wang and Haojie Li and Lei Huang},
  title = {{Predicting the Punching Force in Wushu Sanda After Neuromuscular Electrical Stimulation by Employing the KAN Neural Network Combined with Neuromuscular Electricity}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25195979},
}
Figure