Real-time continuous assessment of fatigue from surface electromyography with deep learning for training load regulation in elite cyclists

Abstract

Background

Accurate assessment of exercise-induced muscle fatigue is essential for optimizing training loads and preventing overtraining in elite athletes. This study presents a deep learning framework for continuous fatigue estimation from raw surface electromyography (sEMG) signals during high-intensity cycling.

Methods

Fourteen elite track cyclists performed a 30-second all-out sprint on a cycle ergometer. Power output was recorded at 1 Hz to derive a continuous fatigue index (percentage decline from peak power). Simultaneously, sEMG signals were recorded at 1000 Hz from four lower-limb muscles. A sliding window approach (3-second windows, 0.25-second stride) was used to construct input samples. A deep learning model integrating convolutional neural networks (CNN), bidirectional long short-term memory (Bi-LSTM), and channel attention mechanisms was developed to predict the fatigue index directly from raw sEMG. Model performance was evaluated using leave-one-subject-out cross-validation with subject-specific fine-tuning, and compared against eight baseline models. Ablation studies were conducted to quantify each component’s contribution.

Results

The proposed CNN–Bi-LSTM–attention model achieved a mean absolute error of 0.048 ± 0.019 and a Pearson correlation coefficient of 0.822 ± 0.123, and was the only model yielding a positive coefficient of determination (R² = 0.493 ± 0.249).

Compared to baseline machine learning models, the proposed model achieved superior prediction accuracy. Furthermore, it reduced prediction errors by 35.1% relative to time-only regression. Ablation studies revealed that removing the CNN or Bi-LSTM modules caused marked performance degradation, whereas removing the attention mechanism minimally affected prediction accuracy but enhanced model interpretability.

Conclusions

The proposed CNN-Bi-LSTM-attention framework enables accurate, continuous fatigue estimation from raw sEMG during high-intensity cycling, offering a promising tool for real-time fatigue monitoring and training load regulation in all-out cycling sprint.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13102-026-02080-2.

Introduction

The 30-second all-out cycling sprint is a commonly used training and testing protocol in competitive track cycling. This task demands maximal power output and places a high metabolic and neuromuscular load on the lower limbs, making it highly representative of race-specific efforts [1]. The high-intensity, short-duration exercise can rapidly induce muscular fatigue [2, 3], which can impair force production capacity, motor control, and overall performance [4–6]. In the context of elite training, fatigue not only compromises acute output but also accumulates over time, increasing the risk of overtraining, injury, and maladaptive responses [7]. Therefore, accurate fatigue assessment is essential for optimizing training loads, enhancing recovery strategies, and supporting long-term performance development in athletes.

Surface electromyography (sEMG) has emerged as a valuable tool for fatigue assessment in both research and applied settings [8, 9]. Commonly used sEMG features for fatigue assessment can be categorized into time-domain features, such as root mean square amplitude (RMS); frequency-domain features, such as median frequency (MF), mean power frequency (MPF), and nonlinear features, such as Lempel–Ziv complexity (LZC), fractal dimension (FD), and sample entropy (SE) [10–13]. These features have shown relatively good performance in static or isometric contraction tasks [14]. Previous studies often employed linear regression to illustrate the relationship between sEMG-derived indicators and muscle fatigue [13]. However, the key issue is that these studies rarely examined the validity of such changes. Moreover, sEMG signals exhibit complex, nonlinear, and even chaotic patterns, reflecting the inherent nonlinearity of neuromuscular electrical activity [15]. Therefore, the effectiveness and robustness of using linear regression models to assess muscle fatigue based on sEMG remain questionable.

With the advancement of artificial intelligence, deep learning models have gained attention for their ability to model complex temporal and spatial dependencies. Specifically, convolutional neural networks (CNNs) have demonstrated strong capability in automatically extracting informative patterns from sEMG signals [16]. CNNs are particularly effective at capturing local temporal and amplitude characteristics of sEMG signals, and have been successfully applied to fatigue analysis. For example, Mu et al. [17] achieved approximately 90% accuracy in a three-class fatigue classification task during static biceps brachii contractions, while Moniri et al. [18] showed reliable prediction of trunk muscle fatigue features using windows up to 25 s. To better capture the sequential dynamics of fatigue, CNNs can be combined with long short-term memory (LSTM) layers, which model long-range temporal dependencies in both forward and backward directions [19]. Despite these advances, continuous estimation of a task-level fatigue index from raw sEMG remains limited, particularly during short-duration maximal-effort tasks such as 30-second all-out cycling sprints. Moreover, deep models are often criticized as “black boxes,” limiting insight into the relative contribution of different muscles. To address this, a channel attention mechanism was introduced to adaptively reweight the contributions of each sEMG channel [20]. This allows the network to emphasize the most informative muscles while suppressing less relevant channels.

This study presents a novel deep learning approach for assessing a continuous fatigue index based on the decline rate of maximal cycling power, using raw multi-channel sEMG signals recorded during a 30-second all-out sprint performed by 14 elite track cyclists. The proposed model integrates CNN, Bi-LSTM, and attention mechanisms to fully exploit spatial-temporal features embedded in the signal. This work aims to provide practical insights for real-time fatigue monitoring, training load optimization, and individualized athletic conditioning.

Methods

Study overview

This study comprised a single experimental session in which each participant performed a 30-second all-out cycling sprint on a stationary ergometer. During the sprint, power output was continuously recorded to derive a time-resolved fatigue index, while multi-channel sEMG signals were simultaneously acquired from four lower-limb muscles. Offline data processing involved sliding-window segmentation and label assignment, followed by two parallel modeling paths: (i) a deep learning approach using raw sEMG signals with a CNN–Bi-LSTM–attention architecture, and (ii) traditional machine learning baselines using handcrafted sEMG features. Model performance was evaluated using leave-one-subject-out cross-validation with subject-specific calibration. An overview of the entire pipeline is illustrated in Fig. 1.

Fig. 1.

Overview of the experimental and modeling pipeline

Participants and data acquisition

Fourteen elite track cyclists (10 males, Age: 21.50 ± 4.67 years; Height: 175.00 ± 8.25 cm; Weight: 75.40 ± 10.91 kg) were recruited for this study. All participants had more than five years of professional training experience and were actively competing at an elite level. Prior to participation, written informed consent was obtained from all subjects. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Tongji University (Approval No. 2020tjdx006). All participants were screened to ensure they were free from musculoskeletal injury, acute fatigue, sleep deprivation, or recent strenuous exercise within 24 h prior to testing.

All experiments were conducted in a controlled laboratory environment (24 °C). Each participant performed a standardized warm-up followed by a single 30-second all-out cycling sprint on a Wattbike Pro cycle ergometer. During the sprint, participants were instructed to pedal as fast and forcefully as possible throughout the entire duration. The ergometer’s electromagnetic resistance was fixed at level 1, while air resistance was individually adjusted to allow maximal power output.

Power output and cadence were recorded at 1 Hz using the Wattbike Performance Computer. Simultaneously, surface electromyography (sEMG) signals were recorded at 1000 Hz using a ME6000 P8 system (Mega Electronics Ltd., Finland) from four right lower-limb muscles: rectus femoris (RF), biceps femoris (BF), tibialis anterior (TA), and lateral gastrocnemius (GL). Electrode placement followed SENIAM recommendations. After shaving, abrading, and cleaning the skin with alcohol, Ag/AgCl electrodes were placed along the direction of muscle fibers with an interelectrode distance of 2 cm. Synchronization of power output and sEMG signals was ensured using a TJ-Motion 2.0 sync trigger.

Data processing and feature extraction

Linear interpolation was applied to resample the power signal to 1000 Hz, matching the sampling frequency of the sEMG signals. Data analysis was conducted from the third-second during which maximal power output was reached for each participant. Task-level fatigue was quantified using a fatigue index defined as the relative decline in power output during the 30-second sprint, calculated as the percentage decrease from each participant’s peak power output. This continuous fatigue index served as the target variable for subsequent predictive modeling.

To visualize temporal activation patterns throughout the task, sEMG envelope signals of the four recorded lower limb muscles were obtained by full-wave rectifying the band-pass filtered sEMG signals, followed by low-pass filtering using a fourth-order Butterworth filter with a cutoff frequency of 5 Hz to smooth the signal and extract amplitude modulation. These envelope signals were used exclusively for descriptive visualization and were not used as inputs to any predictive model.

To construct input samples, a sliding window approach was adopted. Each window contained 3000 consecutive data points (corresponding to 3 s) from four muscle channels, forming a 3000 × 4 matrix. The stride between consecutive windows was set to 250 points (0.25 s), resulting in overlapping windows. For each 3-second window, the corresponding fatigue index label was assigned as the mean value of the interpolated fatigue index across all time points within that window. This averaging approach provides a stable target value for each input sample.

For the deep learning model, the raw sEMG signals within each window were directly used as input. A global z-score normalization was applied to each muscle channel based on the statistics of the training data [21].

For traditional models, the raw sEMG signals were first band-pass filtered (5–495 Hz) and a 50 Hz notch filter. From each filtered window, six handcrafted features were extracted for each muscle, yielding a 24-dimensional feature vector per sample. The extracted features included the root mean square (RMS), median frequency (MF), and mean power frequency (MPF) based on Fourier transform, as well as Lempel–Ziv complexity (LZC), fractal dimension (FD), and sample entropy (SE). RMS, MF, and MPF were calculated following the procedures described by Cifrek et al. [8], while LZC, FD, and SE were computed according to the methods proposed by Kaspar et al. [22], Boccia et al. [23], and Richman and Moorman [24], respectively.

For both modeling approaches, the input (either raw signals or feature vectors) was paired with the interpolated fatigue index to form the final dataset for training and evaluation.

Model construction

A deep neural network integrating channel attention, convolutional layers, and a bidirectional long short-term memory (Bi-LSTM) module was constructed to estimate the fatigue index from sEMG signals.

First, a channel attention mechanism was applied to adaptively reweight the four sEMG channels. Global average pooling was performed along the temporal dimension to obtain channel descriptors, which were passed through two fully connected layers with ReLU and sigmoid activations to generate attention weights. The weights were then multiplied with the original input to produce channel-refined signals.

The weighted signals were processed by two one-dimensional convolutional layers (128 filters, kernel size = 5) with batch normalization, followed by max-pooling and dropout (rate = 0.2). The extracted features were then fed into a Bi-LSTM layer with 64 hidden units, followed by layer normalization and dropout. Global average pooling was applied to aggregate the temporal features.

Finally, the pooled feature vector was passed through two fully connected layers (128 and 64 units) with ReLU activation and L2 regularization, followed by a sigmoid output neuron that predicted the normalized fatigue index.

Model optimization and training strategy

To optimize the regression model for fatigue index prediction, the mean squared error (MSE) was adopted as the loss function.

After extensive hyperparameter tuning, the following training strategy was adopted for optimal stability and precision:

• Optimizer: Adam with an initial learning rate of 0.0005

• EarlyStopping: Patience of 10 epochs, monitoring validation loss (val_loss), with restoration of the best weights.

• ReduceLROnPlateau: Learning rate decay factor of 0.5, patience of 5 epochs, and a minimum learning rate of 0.000001

• Batch size: 16.

• Maximum epochs: 1000.

During training, model weights were updated based on the training set, while the val_loss was monitored to prevent overfitting and guide learning rate adjustment.

Model assessment

Baseline machine learning models

To comprehensively evaluate the effectiveness of the proposed deep learning model, several conventional regression algorithms were implemented as baseline methods. All baseline models were trained using the handcrafted 24-dimensional feature vectors described in section  "Data processing and feature extraction".

The following regression algorithms were considered:

• Linear Regression, Ridge Regression (Ridge) [25], Lasso Regression (Lasso) [26], Support Vector Regression (SVR) [27], Random Forest [28], Gradient Boosting Regression (GBR) [29], Extreme Gradient Boosting (XGBoost) [30] and Time-only Regression.

All baseline models were trained under the same data split as the proposed deep learning model to ensure fair comparison. Hyperparameters were selected using cross-validation where applicable. Default settings were adopted when no significant improvement was observed after tuning.

Ablation study design

To quantify the contribution of each architectural component in the proposed deep learning framework, a series of ablation experiments were conducted. The following model variants were evaluated under identical experimental conditions as the full model:

• Without Channel Attention (w/o Attention): the attention layer was removed while keeping the convolutional and Bi-LSTM layers unchanged.

• Without CNN (w/o CNN): convolutional layers were removed, and the raw attention-weighted signals were directly fed into the Bi-LSTM.

• Without Bi-LSTM (w/o Bi-LSTM): the recurrent layer was removed, and convolutional features were directly connected to the output layer.

Performance evaluation

A Leave-One-Subject-Out (LOSO) cross-validation with subject‑specific calibration/fine‑tuning was adopted. For each fold, data from a target subject were held out for fine-tuning and testing, while data from the remaining subjects were used for initial training.

Of note, this evaluation protocol is best characterized as a semi-personalized or calibrated LOSO scheme, rather than a strictly unseen-subject prediction setting. Specifically, for each target subject, a small portion (20%) of chronologically early windows was used for fine-tuning the pre-trained model, while the remaining 80% served as the test set. This design reflects a realistic application scenario where a generic model can be rapidly calibrated to a specific athlete using a brief initial recording, thereby improving cross-subject generalization without requiring full retraining. For each training subject, the first 80% of windows were used for training and the last 20% for validation, which guided early stopping and learning rate adjustment during initial training, following the same hyperparameters described in section "Model optimization and training strategy".

The model was first trained on the combined training set (with validation monitoring) according to the training strategy detailed in section "Model optimization and training strategy". After initial training, the model was fine-tuned on the fine-tuning set for a few epochs with a reduced learning rate.

Finally, the model was evaluated on the test set using three metrics: mean absolute error (MAE), coefficient of determination (R²), and Pearson correlation coefficient (Pearson r). The procedure was repeated for all target subjects, and the mean±standard deviation of each metric was reported.

The same evaluation procedure (LOSO cross-validation with temporal splitting, global normalization, and subject‑specific calibration/fine-tuning) was applied to all models to ensure fair comparison.

Visualization analysis

Scatter plots of predicted versus actual fatigue index values were generated for all models to visualize prediction accuracy and bias (Supplementary Figure S1). For the deep learning model, MAE across cross-validation folds and training–validation loss curves were plotted to illustrate performance variability and convergence behavior.

Statistical analysis

Spearman correlation analysis was performed to assess the correlation among time, power output and fatigue index. Wilcoxon signed-rank tests were conducted to assess the differences in MAE and Pearson r between the deep learning model and each baseline model. Results are presented as mean ± standard deviation, along with 95% confidence intervals (CIs) for MAE. Statistical significance was set at p < 0.05. All statistical analyses were performed using SPSS 29.0 (IBM Corp., Armonk, NY, USA).

Results

Temporal changes in power output and fatigue index during exercise

Figure 2 illustrates the time-dependent changes in power output and fatigue index during the 30-second all-out cycling task. Spearman correlation analysis revealed a significant monotonic decrease in power output over time (r = − 0.943, p < 0.001), alongside a significant monotonic increase in the fatigue index (r = 0.944, p < 0.001) as exercise progressed.

Fig. 2.

Time-course curves of power output and fatigue index during the 30-second all-out cycling task (A. power output during the 30-second all-out cycling task; B. fatigue index during the 30-second all-out cycling task)

Characteristics of sEMG changes during task

Figure 3 shows the sEMG envelopes of four lower limb muscles during a 30-second all-out cycling task for a representative participant. It can be observed that the envelopes of the RF and GL exhibit a clear gradual decline over time. In contrast, the envelopes of the BF and TA show no obvious changes.

Fig. 3.

sEMG envelopes of four lower limb muscles over time for a representative participant

Deep learning model evaluation

Figure 4 illustrates the prediction performance and training process of the deep learning model:

Fig. 4.

Deep learning model evaluation. A: Scatter plot of true vs. predicted fatigue values; B: Subject-wise MAE; C: Training and validation loss curves

Figure 4A presents a scatter plot of predicted versus true fatigue index values across all test windows and subjects. Most points are tightly clustered around the ideal diagonal line, indicating good overall agreement between predictions and actual fatigue levels. Some dispersion is observed at low fatigue levels, where predictions exhibit slightly larger deviations, likely reflecting the limited discriminability of sEMG patterns during the early phase of the sprint when fatigue has not yet substantially developed. No systematic over- or under-estimation bias is evident across the full range of fatigue values.

Figure 4B displays the subject-wise MAE distribution, revealing notable inter-individual variability in prediction accuracy. Most subjects achieved MAE below 0.06, while a few exhibited higher errors (MAE > 0.07). This variability may be attributed to differences in individual sEMG signal characteristics, electrode placement consistency, or the dynamic range of fatigue progression across subjects. Nevertheless, the majority of subjects show clinically acceptable prediction errors, supporting the model’s general applicability.

Figure 4C shows the training and validation loss curves over epochs for a representative subject. Both curves decrease rapidly during the initial epochs and stabilize thereafter, with no widening gap between training and validation loss, indicating stable convergence without obvious overfitting. Early stopping was triggered at epoch 45, suggesting that the model reached optimal generalization before the maximum epoch limit. The learning rate adjustments (triggered by ReduceLROnPlateau) are visible as stepwise drops in the loss trajectory, contributing to the final convergence.

Comparison of model performance across fatigue assessment models

Table 1 summarizes the performance metrics of each model along with the statistical comparison results against the Deep Learning model. Figure 5 presents the boxplots of MAE and Pearson’s correlation coefficient for all evaluated models.

Table 1.

The performance metrics of each model along with the statistical comparison results against the Deep Learning model

Model	MAE			Pearson r			R²
	Mean ± SD	95%CI	Effect size r	Mean ± SD	95%CI	Effect size r	Mean ± SD	95%CI
Deep Learning	0.048 ± 0.019	[0.037, 0.058]	—	0.822 ± 0.123	[0.751, 0.893]	—	0.493 ± 0.249	[0.349, 0.637]
Linear	0.107 ± 0.056**	[0.075, 0.139]	0.81	0.728 ± 0.246*	[0.585, 0.871]	0.56	-2.160 ± 3.361	[-4.100, -0.220]
Lasso	0.101 ± 0.065**	[0.064, 0.139]	0.71	0.743 ± 0.232	[0.609, 0.877]	0.21	-2.461 ± 4.153	[-4.859, -0.063]
Ridge	0.107 ± 0.056**	[0.075, 0.140]	0.81	0.730 ± 0.242*	[0.590, 0.870]	0.56	-2.123 ± 3.305	[-4.031, -0.215]
SVR	0.087 ± 0.035**	[0.067, 0.108]	0.75	0.727 ± 0.193	[0.616, 0.838]	0.41	-1.337 ± 2.753	[-2.926, 0.252]
Random Forest	0.105 ± 0.065*	[0.068, 0.142]	0.66	0.503 ± 0.349**	[0.301, 0.705]	0.86	-2.627 ± 5.321	[-5.699, 0.445]
GBR	0.096 ± 0.065*	[0.059, 0.134]	0.61	0.597 ± 0.422**	[0.353, 0.841]	0.78	-2.986 ± 7.460	[-7.292, 1.320]
XGBoost	0.097 ± 0.058*	[0.064, 0.131]	0.65	0.543 ± 0.354***	[0.339, 0.747]	0.88	-2.760 ± 5.954	[-6.197, 0.677]
Time-only Regression	0.074 ± 0.042*	[0.050, 0.099]	0.55	0.903 ± 0.188	[0.794, 0.999]	0.48	-2.900 ± 8.671	[-7.905, 2.105]

P-values and effect sizes (r) are derived from Wilcoxon signed-rank tests comparing each model with the Deep Learning Model for both MAE and Pearson correlation coefficient (Pearson r)

Significance levels are indicated as *p < 0.05, **p < 0.01, ***p < 0.001

Fig. 5.

The boxplots of MAE and Pearson’s correlation coefficient for all evaluated models (A. The boxplots of MAE for all evaluated models; B. The boxplots of Pearson’s correlation coefficient for all evaluated models)

As shown in Table 1, the Deep Learning model achieved the lowest MAE (0.048 ± 0.019). Compared with the Deep Learning model, all other models exhibited significantly higher MAE (Linear Regression: p = 0.002, r = 0.81; Lasso: p = 0.008, r = 0.71; Ridge: p = 0.002, r = 0.81; SVR: p = 0.005, r = 0.75; Random Forest: p = 0.013, r = 0.66; GBR: p = 0.022, r = 0.61; XGBoost: p = 0.016, r = 0.65 and Time-only Regression: p = 0.038, r = 0.55).

Regarding Pearson’s correlation with actual fatigue levels, the Deep Learning model showed significantly higher r compared to Linear Regression (p = 0.035, effect size r = 0.562), Ridge Regression (p = 0.035, r = 0.562), Random Forest (p = 0.001, r = 0.864), GBR (p = 0.003, r = 0.797), and XGBoost (p < 0.001, r = 0.881). No significant differences were observed between the Deep Learning model and Lasso (p = 0.433, r = 0.21), SVR (p = 0.124, r = 0.411), or Time‑only Regression (p = 0.074, r = 0.478). Among all models, Time‑only Regression achieved the highest Pearson’s r (0.903 ± 0.188), but this value was not statistically different from that of the Deep Learning model.

Notably, the Deep Learning model was the only one to achieve a positive R² value (0.493 ± 0.249), whereas all other models exhibited negative R² values, indicating poor explanatory power.

Ablation study of model components

As shown in Table 2, the Full Model achieved the best overall performance (MAE = 0.048 ± 0.019; R² = 0.493 ± 0.249; Pearson r = 0.822 ± 0.123), confirming the effectiveness of the integrated CNN-Bi-LSTM-attention architecture for continuous fatigue estimation. When the CNN module was removed (w/o CNN), performance declined (MAE = 0.053 ± 0.022; R² = 0.236 ± 0.339; Pearson r = 0.563 ± 0.412), indicating that local feature extraction plays an important role in capturing spatial patterns across multi-muscle inputs.

Table 2.

The ablation results

Model	MAE			Pearson r			R²
	Mean ± SD	95% CI	Effect size r	Mean ± SD	95% CI	Effect size r	Mean ± SD	95% CI
Full Model	0.048 ± 0.019	[0.037, 0.058]	—	0.822 ± 0.123	[0.751, 0.893]	—	0.493 ± 0.249	[0.349, 0.637]
w/o Attention	0.049 ± 0.018	[0.038, 0.059]	0.09	0.800 ± 0.127	[0.727, 0.873]	0.5	0.338 ± 0.328	[0.148, 0.528]
w/o CNN	0.053 ± 0.022	[0.040, 0.066]	0.23	0.563 ± 0.412**	[0.325, 0.801]	0.75	0.236 ± 0.339	[0.040, 0.432]
w/o Bi-LSTM	0.073 ± 0.047	[0.046, 0.100]	0.5	0.761 ± 0.206	[0.642, 0.880]	0.42	-0.778 ± 2.404	[-2.166, 0.610]

P-values and effect sizes (r) are derived from Wilcoxon signed-rank tests comparing each model with the Full Model for both MAE and Pearson correlation coefficient (Pearson r)

Significance levels are indicated as **p < 0.01

Removing the attention mechanism (w/o Attention) led to slight degradation in R² and Pearson correlation (MAE = 0.049 ± 0.018; R² = 0.338 ± 0.328; Pearson r = 0.800 ± 0.127).

The most substantial decline was observed when the Bi-LSTM module was removed (w/o Bi-LSTM). In this case, MAE increased to 0.073 ± 0.047, while R² dropped dramatically to -0.778 ± 2.404, indicating a substantial loss of explanatory power.

Muscle-specific attention weight analysis

The average attention weights of the four sEMG channels are shown in Table 3. Overall, the weights were comparable across muscles, with the antagonist muscles (BF and TA) showing slightly higher mean weights than the agonist muscles (RF and GL); however, these differences were not statistically significant.

Table 3.

The muscle-specific attention weights

Muscles	Attention weights (mean ± SD)
Rectus femoris	0.491 ± 0.004
Biceps femoris	0.512 ± 0.010
Tibialis anterior	0.510 ± 0.008
Lateral gastrocnemius	0.488 ± 0.014

Discussion

This study presents a novel approach for continuous fatigue assessment during a 30-second all-out cycling sprint using a CNN–Bi-LSTM–Attention model applied to raw sEMG signals. The model achieved the lowest MAE (0.048 ± 0.019) and was the only one to yield positive R² (0.493 ± 0.249) among all evaluated models, demonstrating its capability to provide meaningful subject-specific fatigue predictions after rapid calibration. While the time-only regression baseline showed higher Pearson correlation (r = 0.903 vs. 0.822), this high correlation primarily reflects the strong monotonic relationship between elapsed time and fatigue progression rather than accurate magnitude estimation. Critically, the time-only model exhibited substantially larger prediction errors and negative R², indicating that time alone cannot reliably estimate fatigue magnitude and that sEMG-derived physiological information provides additional predictive value beyond elapsed time.

Previous studies have typically relied on handcrafted sEMG features combined with traditional statistical models for fatigue assessment. These approaches are simple, interpretable, and computationally efficient, and therefore have been widely adopted in early fatigue research. Coorevits et al. [31] evaluated muscle fatigue during a standardized static back extension task and found that linear regression was applicable in static and simple tasks; however, its effectiveness decreased when the fatigue mechanisms became more complex or the task dynamics increased. Golabchi et al. [32] reported that regression coefficients indicated that features such as RMS and envelope amplitude variability were informative for estimating dynamic activity levels, but these models may not be sufficient to handle the complex changes that occur during high-intensity or real-time training. In this study, all traditional regression approaches, including Linear Regression, Lasso, Ridge, SVR, Random Forest, GBR, and XGBoost, demonstrated markedly lower performance compared to the proposed CNN‑Bi‑LSTM‑Attention model. This performance degradation highlights the limitations of conventional methods in capturing nonlinear, non-stationary temporal patterns inherent in sEMG signals during short-duration, high-intensity exercise.

Building on these findings, more advanced approaches such as deep learning are required to capture the complex and dynamic characteristics of neuromuscular fatigue. Previous studies have highlighted the promise of advanced machine learning and deep learning models for muscle fatigue assessment. In recent years, several studies have shown that deep learning and machine learning approaches can effectively leverage sEMG and multimodal signals to capture the nonlinear and temporal features of muscle fatigue, enabling accurate recognition and short-term prediction of fatigue states, and demonstrating the feasibility of real-time monitoring [17, 18, 33]. However, these studies have primarily focused on classification tasks under static or relatively controlled conditions, which do not capture the continuous nature of fatigue progression.

Recently, Garofalo et al. [34] proposed a deep learning framework for continuous regression of proximity to muscular failure during isometric biceps brachii holds, using spectrogram representations of sEMG and an LSTM architecture with LOSO cross-validation. Their results showed that continuous fatigue estimation from sEMG is feasible in static, single-joint isometric tasks where muscle activation patterns are relatively stable and motion artifacts are minimal. Building on this concept, the present study addresses a more dynamic context: a 30‑second all-out cycling sprint involving multi-joint, explosive contractions with rapid joint angle changes and non‑stationary sEMG signals. Unlike elapsed time used as a proxy for fatigue in Garofalo et al., our fatigue index is derived directly from power output decline, providing an objective, performance-based measure relevant to athletic contexts. To our knowledge, this is the first study to validate an end-to-end deep learning model for continuous regression of a task-level fatigue index during a high-intensity, ecologically valid sprint, demonstrating the feasibility of sEMG-based continuous fatigue estimation in dynamic, multi-joint athletic tasks.

To further contextualize our contribution, Table 4 summarizes recent studies on sEMG-based fatigue assessment using machine learning and deep learning approaches. Compared to existing work, which has predominantly focused on static or isometric tasks and classification problems, the present study addresses a continuous regression task in a dynamic, high-intensity cycling protocol using raw multi-channel sEMG signals with a CNN–Bi-LSTM–attention architecture. This comparison highlights the unique challenges of estimating fatigue in ecologically valid sports settings and underscores the methodological advancements offered by our approach.

Table 4.

Comparison of the proposed method with recent sEMG-based fatigue assessment studies

Study	Exercise / Task	Deep Learning Model	Signal Input	Validation Strategy	Performance
Mu et al. (2022) [17]	Static isometric biceps contraction	TCN	sEMG + ECG	Cross-validation	Accuracy ≈ 90% (3-class)
Moniri et al. (2021) [18]	Static trunk task	ML-based forecasting	sEMG features	Time-series forecasting	Feature prediction performance
Liu et al. (2023) [33]	Dynamic elbow flexion–extension	CNN–LSTM–Transformer	sEMG (ApEn + PE)	Cross-validation	Classification accuracy
Hwang et al. (2025) [19]	Static and dynamic tasks	CNN + LSTM + Attention	sEMG + IMU	Cross-validation	F1-score
Garofalo et al. (2026) [34]	Isometric hold	LSTM	Spectrogram	Leave-one-subject-out (LOSO)	MSE = 49.44 ± 18.34; MAE = 4.68% (R² not reported)
Ours	Dynamic sprint cycling	CNN–Bi-LSTM–Attention	Raw sEMG (4 channels)	LOSO + subject-specific fine-tuning	MAE = 0.048; R² = 0.493; r = 0.822

The ablation results (Table 2) quantify the contribution of each architectural component in the proposed CNN–Bi-LSTM–Attention framework. Removing the CNN module (w/o CNN) resulted in a marked performance decline, with Pearson r dropping from 0.822 to 0.563 (effect size r = 0.75) and R² decreasing from 0.493 to 0.236, confirming that convolutional layers are essential for extracting local spatial patterns and rapid amplitude fluctuations from raw sEMG signals without manual feature engineering [35, 36]. Removing the Bi-LSTM module (w/o Bi-LSTM) caused the largest increase in MAE (from 0.048 to 0.073, effect size r = 0.50) and rendered R² negative (− 0.778), highlighting that temporal modeling of progressive fatigue across the 30‑second sprint is indispensable; without capturing long‑range dependencies, the network fails to track the sequential evolution of muscle activation. Removing the attention mechanism caused minimal changes in MAE (0.048 vs. 0.049) and a slight decrease in Pearson r (0.822 vs. 0.800), suggesting its primary contribution lies in physiological interpretability rather than predictive performance. This interpretation is consistent with the observation that muscle-specific attention weights were comparable across channels and not statistically significant (Table 3).

Together, the CNN–Bi-LSTM–attention architecture integrates multiscale spatial features, long-term temporal dependencies, and context-aware salience, making it highly effective for real-time, continuous fatigue monitoring in elite cycling [37, 38].

The proposed model achieved a mean R² of 0.493 ± 0.249 under the LOSO cross-validation protocol with subject-specific fine-tuning. Although this indicates that about half of the variance in fatigue indices was explained, the result should be interpreted in the context of the well-known difficulty of cross-subject generalization in sEMG analysis. Variability in R² may partly arise from the heterogeneity of fatigue label distributions across subjects, as some individuals exhibited a limited dynamic range in the fatigue index derived from power decline, reducing the total variance available for explanation. In addition, substantial inter-individual variability in sEMG signals, caused by differences in muscle properties, motor unit recruitment, and electrode placement, introduces distribution shifts that challenge cross-subject modeling [39, 40]. Notably, continuous fatigue regression using sEMG signals has rarely been explored. For example, Garofalo et al. [34] reported a deep learning approach for fatigue estimation during isometric elbow flexion and achieved a strong MAE of 4.68%, but R² values were not reported, limiting direct comparison. Importantly, the proposed deep learning model was the only model among all baseline methods that achieved a positive R², whereas all conventional machine learning models yielded R² ≤ 0 under LOSO validation. Despite the moderate R², the model achieved low MAE (0.048 ± 0.019) and high correlation (r = 0.822 ± 0.123), indicating that it can reliably track fatigue progression and provide meaningful subject-specific fatigue predictions.

From an applied perspective, the combination of low MAE and strong correlation supports the practical utility of the proposed framework. In athletic training contexts, the ability to track fatigue progression within 5% error and maintain high rank-order consistency with true fatigue levels is sufficient for informing real-time training adjustments, fatigue monitoring, and recovery optimization. The model’s demonstrated generalizability across 14 elite athletes, combined with its use of raw sEMG signals without extensive preprocessing, positions it as a promising candidate for integration into wearable devices for field‑based fatigue monitoring, although field‑based validation is still required. Future work could explore further improvements through larger and more diverse training datasets or advanced domain generalization techniques, but the current results already establish a solid foundation for future development in elite sports settings.

Despite promising results, limitations include the small sample size (N = 14), absence of external validation, indirect fatigue index derived from power decline, and potential confounding among time, power, and fatigue progression. Moreover, although the sliding-window segmentation substantially increased the number of training samples, these windows are derived from a limited number of subjects and should not be considered statistically independent observations; the effective sample size remains at the subject level. Future work should include larger and more diverse cohorts, field-based validation, and direct physiological measures to further confirm robustness and applicability.

Conclusion

This study demonstrates that a deep learning model integrating CNN–Bi-LSTM–attention architecture can accurately estimate continuous fatigue indices directly from raw sEMG recorded during 30-second all-out cycling in elite athletes. In comparison to machine learning models based on handcrafted features, the proposed deep learning framework demonstrates superior performance in both accuracy and consistency, offering a more robust and reliable approach for real-time muscle fatigue monitoring in all-out cycling sprint.

Abbreviations

BF

Biceps femoris

Bi-LSTM

Bidirectional long short-term memory

CI

Confidence interval

CNN

Convolutional neural network

FD

Fractal dimension

GBR

Gradient boosting regression

GL

Lateral gastrocnemius

LOSO

Leave-one-subject-out

LSTM

Long short-term memory

LZC

Lempel-Ziv complexity

MAE

Mean absolute error

MF

Median frequency

ML

Machine learning

MPF

Mean power frequency

MSE

Mean squared error

RF

Rectus femoris

RMS

Root mean square

SE

Sample entropy

sEMG

Surface electromyography

SVR

Support vector regression

TA

Tibialis anterior

XGBoost

Extreme gradient boosting

Authors’ contributions

RZ: Data curation, Wrote the main manuscript, Formal analysis, Investigation, Methodology. GM: Investigation, Writing – review & editing. JQ: Investigation, Writing – review & editing. MG: Investigation, Writing – review & editing. WN: Investigation, Writing – review & editing. LW: Methodology, Writing – review & editing. All authors reviewed the manuscript.

Funding

Not applicable.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Tongji University, Shanghai, China (Approval No. 2020tjdx006) and was performed in accordance with ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Written informed consent to participate in this study was provided by the participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.