All experimental protocols involving human participants in this study were reviewed and approved by the Medical Ethics Committee of Hubei University for Nationalities [Ethics Approval Number: (2025) 05]. Before the study commenced, the researchers explained the study objectives in detail to all participants or their legal guardians to ensure full understanding of the research. All participants voluntarily signed written informed consent forms. During the study, participants’ personal identifying information was replaced with Subject IDs instead of real names.

The dataset consists of 15 participants, including healthy individuals and patients with varying degrees of hand motor impairment, rated from 0 (no movement) to 4 (normal function) (Table 1). Motor ability was assessed based on observable voluntary control: 0 indicates a complete loss of voluntary movement with no observable joint motion; 1 indicates slight voluntary movement that can perform the intended action, but with poor coordination and noticeably slow execution; 2 represents the ability to perform simple movements with limited strength, amplitude, and precision; 3 denotes the ability to carry out most daily movements with near-normal function, though mild deficits may still appear in fine or rapid actions; and 4 corresponds to normal, well-coordinated movement without observable impairment.

Data were collected at the Affiliated Hospital of Hubei University for Nationalities and Laifeng County People’s Hospital using a six-channel sEMG sensor and a motion-capture glove. Participants performed six standardised gestures (0–5, Figure 1) while sEMG signals and finger-curvature data were recorded. Electrodes were placed over three target muscles—flexor pollicis longus (thumb), flexor digitorum superficialis (fingers 2–5) and flexor digitorum profundus (fingers 2–3)—following established placement guidelines (Ferrante et al., 2024). Pearson correlation analysis (Table 2) showed that channels 2, 3 and 6 had the strongest association with finger flexion, and were therefore used for further analysis. Electrode positions are shown in Figure 2.

All participants performed six predefined gestures after viewing the demonstration video. During execution, they maintained a relaxed posture with forearms resting naturally on the table or bed to minimise postural interference. Healthy subjects completed the movements independently based on visual cues, while testers provided verbal instructions and demonstrations for subjects with motor impairments to ensure correct understanding and maximise task completion.

Figure 3 shows the equipment used for this data collection. Figure 4 shows the data transmission process of the data acquisition device, which is based on the STM32F103VET6 microcontroller (Tong et al., 2024) and integrates the finger curvature sensor and the six-channel sEMG acquisition module (Sapsanis et al., 2013). The finger curvature sensor acquires finger movement data by detecting resistance changes (Yin et al., 2018), while the sEMG acquisition module records EMG signals from forearm muscles through six electrodes with a sampling rate of 225 Hz. All data are transmitted through the serial port and saved as CSV files for subsequent processing and analysis.

A preliminary cleaning step was applied to remove erroneous entries, missing values, and outliers, thereby reducing noise and improving data reliability. The cleaned data were then standardised and stored in CSV format to facilitate subsequent processing and analysis.

Following data cleaning, all signals were smoothed using the Savitzky–Golay filter. Each dataset file contains eight time-series channels, each corresponding to one signal stream. All channels underwent the same filtering procedure to retain key waveform characteristics while suppressing high-frequency noise. The mathematical expression of the Savitzky–Golay filter is shown in Equation 1 (Lei, 2014). y n = ∑ k = − M M c k ⋅ x } n + k (1)

In Equation 1, 2M+1 is the filter window length (optimised to 101 in this study, M = 50). c k denotes the coefficients obtained from least-squares fitting of a p-order polynomial (optimised to 4th order). Zero-padding was applied at both signal edges to prevent boundary artefacts and ensure valid indexing. This filter is a technique used for data smoothing that creates a smooth approximate version of the original data by performing a local polynomial regression on it (Cai and Tang, 2011). The Savitzky–Golay filter smooths signals by fitting local polynomial regressions, thereby preserving waveform characteristics while reducing noise (Cai and Tang, 2011).

Figure 5 presents a comparison of raw and filtered signals for gesture 0 from Subject 9. As shown in Figure 5, the filtered data exhibit visibly smoother trajectories with reduced high-frequency fluctuations, particularly near peak regions. These results, further supported by Figure 6, confirm that the Savitzky–Golay filter effectively suppresses noise while maintaining the essential temporal structure of both finger-curvature and sEMG channels.

As shown in Figure 6, subjects with motor function level 4 exhibit the strongest hand mobility, characterised by clear waveform fluctuations in both finger-curvature and sEMG channels, indicating active neuromuscular engagement. Individuals at level 3 retain relatively good coordination and movement amplitude, with pronounced activity in the thumb and index finger. In contrast, subjects with levels 1–2 show markedly reduced movement, reflected by weaker and less stable signal variations. The level 0 subject displays almost no detectable hand motion, with minimal fluctuations across all channels, indicating near-complete functional loss.

From these analysis results, it can be seen that the degree of impairment of hand function is closely related to the fluctuation of channel and finger data. The more frequent and larger the fluctuation, the better the motor function of the hand. In contrast, hand function may be limited or lost. These data can be used to help assess and track recovery from hand function in patients and provide a basis for rehabilitation treatment.

The Short-Time Fourier Transform (STFT) was applied to perform time-frequency analysis on the processed data. The STFT formula (Wang et al., 2020) is given by Equation 2. X m , k = ∑ n = 0 N w i n − 1 x m · N h o p + n · w n · e − j 2 π k n / N w i n (2)

In Equation 2, N w i n = 64 is the number of window samples, N h o p = 32 is the window step size (50% overlap), w n is a Hamming window (to reduce spectral leakage). X m , k is a complex coefficient, and the square of its magnitude is used for the spectrogram modelling.

Time–frequency features are widely used to reduce limb-position interference in sEMG recognition tasks (Khushaba et al., 2014). Since the raw samples vary in length, direct processing is cumbersome; therefore, STFT was applied to convert the signals into spectrograms with unified dimensions. Compared with the original waveform, spectrograms provide a clearer depiction of temporal frequency variations, reduce noise interference, and enhance model recognition performance (Xie et al., 2012). As shown in Figure 7, spectrograms were generated using matplotlib, with time and frequency as axes and colour intensity representing spectral energy, enabling intuitive observation of dynamic signal changes.

We evaluated the impact of the preprocessing parameters on classification performance. The effects of the STFT and Savitzky–Golay filter configurations are detailed in Tables 3, 4 respectively.

We used the KNN model to classify subjects with different levels of motor impairment, with the classification performance for each level shown in Figures 8, 9. Figure 8 displays the accuracy heatmaps corresponding to five impairment levels, while Figure 9 illustrates the results in line-chart form. As shown, the KNN model achieved the highest accuracy for subjects with motor function scores of 3 and 4.

However, clear differences in signal properties and model performance were observed across impairment levels. Subjects with a motor function score of 0 showed almost no voluntary movement and produced weak, unstable sEMG activity, resulting in highly similar patterns across gestures and poor classification accuracy. In contrast, subjects with higher function levels generated more distinguishable signals, leading to stronger model performance. Overall, these results indicate that sEMG remains effective for differentiating motor function states, even among individuals with notable impairment.

We used five machine learning models (KNN (Zhang and Li, 2021), SVM (Pisner et al., 2020), MLP (Taud et al., 2018), LightGBM (Sun et al., 2018), and Random Forest (Liu et al., 2012)) to test the classification of our self-built dataset.

Figure 10 shows the confusion matrices of the five classification models. While the models differ in classification performance, most exhibit relatively low misclassification rates, reflecting good overall recognition performance.

Table 5 presents the performance evaluation of five models—SVM, Random Forest, MLP, LightGBM, and KNN—on both multi-modal fusion sensor and single-sensor test sets. The F1-score, defined as the harmonic mean of precision and recall, reflects the overall effectiveness of a model in classification tasks (Lam et al., 2023). MCC takes into account true and false positives and negatives, making it particularly suitable for evaluating imbalanced datasets (Itaya et al., 2024). Cohen’s Kappa measures classification consistency by accounting for random agreement (McHugh, 2012). Recall evaluates the ability of the model to identify positive instances, that is, the proportion of actual positives correctly captured (Saito and Rehmsmeier, 2015).

As illustrated in Table 5, the multi-sensor fusion strategy achieves the highest performance in all models. Among these, SVM and MLP deliver the best classification results, demonstrating strong robustness, especially in handling imbalanced data. LightGBM follows closely with competitive performance, while KNN and Random Forest show moderate results. These findings underscore the importance of model selection in improving classification accuracy, particularly in tasks that require high recall and consistency, for which SVM and MLP emerge as preferable options.

Finger movement data and electromyography signals from different subjects performing specified actions were input into five machine learning models. Figure 11 illustrates the cross-validation accuracy and test accuracy of the five machine learning models under different sensor configurations. The cross-validation accuracy reflects the stability of the model in the training set, while the test accuracy evaluates its performance on unseen data. The results indicate that the multi-sensor fusion approach yields the best performance across all models. Specifically, SVM, LightGBM, and MLP achieve high performance in both cross-validation and test sets, each exceeding 95% accuracy. The KNN model shows a slight decrease in the test set, but still maintains high accuracy. In contrast, Random Forest performs relatively poorly in this task, which may be attributed to its sensitivity to high-dimensional feature spaces.

To evaluate performance on an external dataset, we conducted comparative experiments, as shown in Table 6. We used the Mendeley sEMG Gesture dataset (Ozdemir et al., 2022), selecting electromyography data corresponding to six standard gesture categories from 15 subjects as experimental samples. In model training and testing, for five machine learning classifiers, the same preprocessing pipeline as our self-collected dataset, with parameters adjusted according to the dataset characteristics. As seen from the tabular data, these classifiers performed well on the Mendeley dataset, achieving an average accuracy of 87.34%. This suggests that the adopted preprocessing and classification strategy remains effective when applied to another publicly available sEMG dataset.

The data acquisition glove used in this study is made of elastic knitted fabric, offering a comfortable fit for most adult users, though performance may decline in individuals with severe muscle atrophy. Durability tests indicate that the integrated flexible sensors withstand up to one million bending cycles before noticeable performance degradation occurs, suggesting suitability for regular household or community use. However, the sEMG electrode pads may loosen due to perspiration and require timely replacement to maintain data continuity. Tests performed outside controlled hospital environments also revealed that noise and interference increase under less stable conditions. Therefore, measurements are recommended in quiet settings with steady posture for optimal signal quality.

This study presents a self-built multimodal dataset consisting of sEMG and finger-movement signals, demonstrating its value in hand function assessment. Nonetheless, several limitations exist. First, the sample size is relatively small (15 subjects) with an uneven distribution of impairment types, and variations in neuromuscular patterns across different pathologies may affect model reliability. Therefore, larger and more diverse clinical cohorts are required to further validate the findings. Second, although the textile glove and adhesive electrodes provide good wearability, fit instability may occur in patients with muscle atrophy, and long-term use can reduce sensor sensitivity. Signal quality also fluctuates with sweat, posture, and other external factors. Future work will address these issues by expanding participant diversity, improving glove durability, and enhancing robustness against interference.

This work presents a device capable of synchronously acquiring sEMG and finger-movement signals, together with a dataset collected from subjects with different levels of hand motor function. The multimodal data provide a more complete representation of neuromuscular activity and physical movement, supporting applications in gesture recognition, rehabilitation assessment, and user identification. Benchmark experiments demonstrate that the dataset is reliable and suitable for multiple analytical tasks.

Future efforts will focus on increasing the participant pool, enriching impairment categories, and integrating additional sensing modalities. Improvements in glove durability and anti-interference performance will also be explored. These developments are expected to further advance research in motor-function assessment and contribute to the design of intelligent assistive rehabilitation systems.