MSCB-DualAttn Network for sEMG-Based Gesture Recognition in Transradial Amputees with Varying Residual Limb Lengths

Abstract

Transradial amputation severely impairs hand function, and myoelectric control offers a promising solution. However, due to the heterogeneity of neuromuscular compensation, existing methods often struggle to generalize across populations with varying residual limb lengths. In this study, we collected an EMG dataset from 41 participants, including healthy individuals and amputees with long, medium, and short residual limbs, performing 15 gestures, and proposed a Multi-Scale Convolutional Block with Dual Attention Network (MSCB-DualAttn) that integrates parallel multi-scale convolutions (kernel sizes 3, 5, and 7) with channel and temporal attention mechanisms. The model achieved recognition accuracies of 93.79%, 86.13%, and 78.10% in the healthy, long-stump, and middle-stump groups, respectively, and 62.69% in the short-stump group, representing a 13.83 percentage point improvement over the baseline model. Ablation studies and interpretability analyses confirmed the complementary roles of the multi-scale and attention modules, while the performance degradation in the short-stump group suggested a potential limitation of existing models when neuromuscular control shifts from fine synergy to coarse compensation. This study proposes a model framework that enhances gesture recognition accuracy in amputees and highlights the potential of physiology-driven architectures for myoelectric control.

1. Introduction

Over the past two decades, researchers have extensively applied pattern recognition techniques to the field of myoelectric control, particularly in individuals with transradial amputation. Within this control framework, the prosthetic control system acquires surface electromyographic signals from the residual limb as the amputee performs movements or gestures, and employs a series of algorithms to identify the user’s motion intention or corresponding gestures [1]. Therefore, the construction of recognition models has become a critical element in pattern recognition-based myoelectric control systems.

Classical machine learning and deep learning algorithms each have their own advantages and characteristics in the field of EMG signal gesture recognition [2,3]. As a more traditional approach, machine learning relies on feature engineering prior to input to select appropriate features. For example, Liao et al. [4] fused two time-domain features, wavelength and sample entropy, as a composite feature set, and performed motion classification based on KNN, achieving an accuracy of 91.05%.

Since the features used in machine learning training are manually selected, they may fail to capture the hidden patterns of sEMG signals across various gestures. In addition, when the number of features increases, machine learning models may suffer from overfitting and insufficient generalization capability. Currently, deep learning has effectively addressed the limitations of traditional machine learning, as deep learning models can automatically learn relevant features from raw signals, significantly reducing the need for extensive handcrafted feature design [5,6]. Convolutional neural network (CNN) dominates sEMG DL research due to its automated feature extraction capabilities [7]. Existing studies have utilized CNN and LSTM architectures and achieved recognition accuracy above 95% on multi-sensory gesture datasets [8]. Integrating a local spectral convolutional neural network (CNN) with a modified ResNet-18 encoder, this study proposes a dual-path deep neural network architecture named Embedded-STQNet, which achieves an accuracy of 98.48% based on high-density surface electromyography signals [9]. Another study [10] took residual limb length into consideration and found that deep learning is more adaptable than machine learning to different amputation conditions. However, this study included only four subjects and did not conduct an in-depth analysis or classification based on residual limb length. The study also highlighted that current datasets contain very few amputee subjects, resulting in limited data availability. A study achieved high accuracy in gesture recognition by considering sEMG data using CNN-based view pooling techniques [11]. Another study proposed a temporal-spectral cross-subject transfer learning framework based on multi-stream CNNs that processes each single sEMG channel individually, achieving promising performance in transradial amputee sEMG gesture recognition [12].

Attention mechanisms have been proven to enhance a model’s focus on key features [13]. The attention mechanism dynamically assigns weights to input features, allowing the model to focus on the most discriminative parts of EMG signals and improving its ability to capture key information. Zhang et al. [14] used attention mechanisms to focus on important sEMG channels for identifying muscle activation. Many studies have significantly improved geometric reasoning, multi-scale fusion, and spatial understanding by combining CNN with attention modules [15,16,17]. A study proposed a CNN architecture that incorporates channel, spatial, and temporal attention blocks after each convolutional layer, which led to significant performance improvements [18]. Shu et al. [19] converted EMG signals into images and combined a three-layer CNN (with 32, 64, and 128 kernels) with Vision Transformer (ViT), achieving 80.02% accuracy on the DB2 dataset. However, while such image-based recognition performs well in laboratory settings, it struggles to meet low-latency demands in practical applications and imposes high requirements on device performance. He et al. [20] leveraged physiological knowledge to construct an MSM module based on multi-scale convolutional kernels, combined with self-attention, for gesture prediction, achieving a recognition rate of 90.15% on the DB2 dataset. A study developed MSCANN for gesture recognition by combining multi-scale convolution with LSTM networks and self-attention mechanisms, achieving an accuracy of 93.80% on the DB2 dataset [21]. However, the network achieves multi-scale feature extraction by serially stacking convolutional layers with different kernel sizes, which prevents it from simultaneously preserving fine-grained details from the original scale. Some studies have developed lightweight attention modules such as sCBAM to suppress redundant and noisy HD-sEMG signals, improve gesture recognition accuracy, and enhance model interpretability [22].

Existing studies on sEMG gesture recognition have verified the effectiveness of combining multi-scale convolution and attention mechanisms. InRes-ACNet [23] constructs multi-scale receptive fields through an Inception-Attention parallel structure, and integrates the SE module and ACmix to achieve weighted fusion of convolutional and self-attention features. Evaluated on the NinaPro DB1 and NinaPro DB5 datasets, it achieves recognition accuracies of 87.94% and 87.04%, respectively. However, this method converts raw one-dimensional sEMG signals into two-dimensional image representations, which inevitably loses inherent temporal information. MS-CLSTM [24] adopts a dual-branch multi-scale convolutional framework embedded with residual CBAM, and employs Bi-LSTM to model long-range temporal dependencies. It reaches accuracies of 86.66% and 83.27% on Ninapro DB2 and DB4. Similarly, it requires two-dimensional signal transformation and introduces recurrent structures, leading to network redundancy and an increased risk of overfitting. TFFNAS [25], derived from TFF-Net, also achieves remarkable performance with an accuracy of 96.48% on the EMAHA dataset. It fuses short-term features from multi-scale convolution and long-range temporal patterns via a DRTCN branch, and only adopts a single temporal attention mechanism for weighting. Its multi-scale design is limited to two-dimensional spatial perception and suffers from redundant network branches. Overall, although these methods achieve favorable recognition accuracy by adopting multi-scale structures and attention designs, they still suffer from common drawbacks, including temporal information loss caused by two-dimensional signal transformation, redundant network architecture, and simplistic decoupling design of attention modules.

At present, several studies have made their datasets publicly available to facilitate further research and data integration by the research community, such as CapgMyo, CSL-HDEMG, Ninapro [26], and Hyser. Most existing studies adopt either public datasets or small-scale self-collected datasets. The public FORS-EMG dataset [27] contains sEMG gesture data from 19 healthy participants. Among the widely used Ninapro datasets, DB10 [28] has the largest number of amputee participants, with only 15 amputee subjects included. Sun et al. [29] recruited 7 healthy subjects from the DB2 dataset and 3 amputee subjects from the DB3 dataset. Furthermore, most studies based on this dataset rarely consider the clinical indicator of residual limb ratio. Moreover, most studies using Ninapro have rarely taken the clinical parameter of residual limb ratio into account. In the research of Niu et al. [30], sEMG gesture data were collected from 12 healthy individuals and 2 transradial amputees under both static and dynamic arm postures. Similarly, the sEMG gesture recognition study by Sharma et al. [31] recruited only 9 healthy participants and 2 transradial amputees. Overall, current sEMG datasets for amputees remain limited in sample size, and the factor of residual limb ratio is rarely considered in existing research.

Based on the above literature review, there is currently no transradial amputation gesture dataset classified according to different residual limb lengths. Meanwhile, the performance variation in existing recognition models when applied to subjects with different residual limb lengths remains unclear. It still needs to be clarified whether the recognition accuracy of the same model declines significantly for short residual limbs, or whether the model can learn fundamental muscle features and maintain stable performance with only marginal differences. In addition, it is worth exploring whether a unified model can be constructed to improve gesture recognition performance for amputee populations.

To address the above issues, this study first constructed an EMG gesture dataset covering amputees with varying residual limb lengths and healthy subjects. On this basis, the Multi-Scale Convolutional Block Dual Attention Network (MSCB-DualAttn) framework is proposed, which employs parallel multi-scale convolutional branches to simultaneously extract local details and global temporal features based on physiological time scales of EMG signals (3 ms, 5 ms, 7 ms). It further integrates channel attention and temporal attention mechanisms to adaptively focus on key muscle channels and critical temporal segments. This design aims to capture common EMG patterns across groups with different residual limb lengths, thereby building a unified recognition model capable of effectively identifying gestures across diverse subject populations. This study aims to fill the dataset gap in the related field by constructing an sEMG gesture dataset consisting of transradial amputees with different residual limb lengths and healthy subjects. Meanwhile, the proposed network framework is adopted to improve the gesture recognition performance for transradial amputees and to compare recognition differences among patients with varying residual limb lengths. This work is expected to provide a research foundation for the future development of recognition models and prosthetic devices targeted at transradial amputees.

2. Materials and Methods

This section details the experimental materials, datasets, experimental setup, preprocessing procedures, network models, and training configurations adopted in this study. The experiments are conducted on three sEMG gesture recognition datasets: a self-collected dataset containing 41 participants, including healthy subjects and individuals with transradial amputation of varying residual limb lengths, and two public benchmark datasets, Ninapro DB2 and DB8. The core recognition model proposed in this work is the Multi-Scale Convolutional Block Dual Attention Network (MSCB-DualAttn), with classic CNN models employed as baseline comparisons. All data processing, model construction, training, and evaluation are implemented based on Python 3.8.20 and the PyTorch 2.4.1 deep learning framework. The experimental hardware is a Windows 10 workstation equipped with an NVIDIA GeForce GTX 1050 Ti GPU, with CUDA 12.4 and cuDNN 90100 enabled to support accelerated model training.

2.1. Dataset Acquisition and Preprocessing

2.1.1. Different Residual Limb Length Amputee Dataset

In this study, participants were recruited in collaboration with the Shanghai Disabled Persons’ Federation to construct a surface electromyography (sEMG) dataset comprising individuals with transradial amputation and healthy subjects. This study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Shanghai YangZhi Rehabilitation Hospital (Shanghai Sunshine Rehabilitation Center) (Approval No. 2024121). The residual limb length percentage of the subjects was calculated as the ratio of the residual limb length to the length of the corresponding segment on the intact contralateral limb. Here, the residual limb length refers to the distance from the distal end of the stump to the radial styloid process, while the length of the corresponding segment on the intact side refers to the distance from the olecranon to the radial styloid process. The dataset was acquired using an eight-channel surface electromyography (sEMG) system (Portable Electromyograph, Model HMK-EMG01, Shanghai Haimeikang Intelligent Medical Technology Co., Ltd., Shanghai, China) with a sampling rate of 1000 Hz. Eight electrodes were placed on the upper arm and forearm of the subjects’ limbs, positioned as close as possible to the muscle bellies. The correspondence between each recording channel and the target muscles is illustrated in Figure 1b,d. For amputee subjects with a short residual limb, electrodes were positioned as distally as possible along the remaining segment while maintaining anatomical correspondence. A total of 41 subjects (

Table 1. Baseline Characteristics of the Participants.

	HS	LS	MS	SS
Sample size	21	5	7	8
Age	44.00 ± 14.86	44.80 ± 15.32	44.87 ± 6.59	62.13 ± 6.56
Residual limb length ratio (%)	/	82.38 ± 4.50	52.30 ± 4.07	28.61 ± 6.22
Gender (Male/Female)	12/9	3/2	7/0	3/5
Residual Limb/Dominant Side (Left/Right)	7/14	4/1	0/7	5/3

HS represents the healthy subjects group, LS represents the long stump group, MS represents the middle stump group, and SS represents the short-stump group.

) were included in the dataset and were divided into four groups according to residual limb length: 21 healthy subjects, 5 long stump subjects (residual limb length ratio > 75%), 7 middle stump subjects (residual limb length ratio between 40% and 75%), and 8 short-stump subjects (residual limb length ratio < 40%).

Fifteen types of gestures were collected in the experiment, as shown in Figure 1a. Each gesture was repeated five times, with each execution lasting four seconds. The experimental procedure was as follows: gestures were performed sequentially, with each gesture repeated five times. Each gesture execution lasted 4 s, followed by a 3 s rest interval between repetitions of the same gesture. A 5 s rest interval was provided between different gestures. The entire session lasted approximately ten minutes (Figure 1c). An instructional video was played in front of the participants, who performed each gesture in synchronization with the video cues. EMG signals were recorded from the dominant hand in able-bodied participants and from the amputated side in participants with amputation. All participants were instructed to perform the gestures bilaterally during execution to ensure the visualization of motor intent for amputee subjects.

2.1.2. Open-Access Dataset

Ninapro DB2 is one of the largest and most widely used datasets in the field of sEMG-based gesture recognition (Figure 2a). The dataset contains sEMG data from 40 healthy subjects (27 males, 13 females), all of whom are able-bodied, with an average age of 29.9 ± 3.9 years. DB2 includes 49 types of hand gestures, covering basic finger movements, wrist motions, grasp types, and functional gestures. Each gesture was repeated six times, with each repetition lasting 5 s. A 12-channel OttoBock EMG electrode system (OttoBock, Duderstadt, Germany) was used, with a sampling rate of 2000 Hz. The electrodes were placed circumferentially around the proximal forearm, covering the major flexor and extensor muscle groups. The system was equipped with a built-in band-pass filter (cutoff frequencies: 10–500 Hz) and a 50 Hz notch filter, and the signals were amplified and stored with 16-bit resolution. The 10–500 Hz band-pass filter is adopted to preserve the valid physiological frequency band of sEMG signals and suppress low-frequency motion artifacts, while the 50 Hz notch filter is used to eliminate power-frequency electromagnetic interference. This represents the most widely adopted raw signal preprocessing configuration for this dataset [32].

Ninapro DB8 is a dataset specifically designed for the amputee population (Figure 2b), containing data from both healthy subjects and individuals with transradial amputation, and is intended to evaluate the applicability of myoelectric control algorithms in real clinical scenarios. Subject composition: DB8 includes a total of 12 subjects, comprising 10 healthy subjects (mean age 45.5 ± 16.7 years) and 2 subjects with transradial amputation. DB8 contains nine types of hand gestures, all of which are common in daily life. Each gesture was repeated six times, with each repetition lasting 5 s. A 12-channel bipolar EMG electrode system (Delsys Trigno Wireless System) was used, with a sampling rate of 2000 Hz. Electrodes were placed over the muscle bellies of the remaining muscles on the residual limb. Signals were hardware-filtered (20–450 Hz band-pass) before storage.

2.1.3. Data Preprocessing

For the self-constructed dataset, the raw EMG signals were first subjected to a preprocessing pipeline. A band-pass filter (20–500 Hz) was applied to preserve the effective frequency band of the EMG signals while attenuating low-frequency motion artifacts and high-frequency noise. This was followed by a 50 Hz notch filter to suppress power-line interference. The onset and offset times of each hand gesture were extracted from synchronized video recordings to generate a motion-time annotation table. Each gesture segment was demarcated based on its recorded start and end times. The GestureLabel corresponds to the ID number (1–15) of each predefined gesture, while the RepetitionLabel indicates the execution order (1–5) of repeated trials of the same gesture. To mitigate inter-subject variability in motor initiation and termination, a focused analysis window was adopted. From each 4 s movement segment, the initial 0.5 s and the final 0.5 s were excluded, retaining the central 3 s execution window (3000 samples) for subsequent processing. To suppress amplitude variations caused by individual differences across subjects, Z-score normalization was performed on the EMG signals of each channel.

Since the public dataset had already undergone preprocessing, the processing method applied in this study was as follows: the original 2000 Hz signal was first downsampled to 1000 Hz, followed by Z-score normalization.

2.1.4. Sliding Window Setting

To meet the requirements of real-time gesture recognition and increase the number of segmented samples, a sliding-window approach was used. To avoid data leakage, sliding-window segmentation was applied separately to each subset after the raw signals were split into training and testing sets. The window size was set to 250 sampling points, corresponding to a duration of 250 ms at a sampling rate of 1000 Hz. This window length is sufficient to capture complete muscle activation patterns while satisfying the latency requirements of real-time recognition. The sliding step size was set to 100 sampling points, corresponding to 0.1 s. Adjacent windows had an overlap of 150 sampling points (overlap rate of 60%), effectively augmenting the number of samples while maintaining temporal resolution. For each gesture, the 3 s effective movement segment yielded approximately (3000 − 250)/100 + 1 ≈ 28 samples. Each gesture was repeated five times, generating approximately 28 × 5 = 140 samples per gesture. For the 15 gestures, this resulted in approximately 2100 samples per subject. Across 41 subjects, a total of approximately 86,000 samples were obtained for model training and evaluation. The same sliding window parameters were used to segment the samples in the public dataset.

2.1.5. Dataset Split

To mitigate data leakage caused by overlapping sliding-window segmentation, we split the raw sEMG recordings before generating any window samples. For each participant and each gesture class, the continuous recordings were divided based on the chronological order of repeated trials: each gesture was repeated five times in total, so the first four repetitions, making up roughly 80% of the original time-series data, were allocated to the training set, and the last repetition, around 20% of the data, was kept aside for testing. Following this repetition-level temporal split, sliding-window segmentation was carried out separately within the training and testing subsets only, with no windows crossing repetition boundaries. This means windows from the training set shared no raw signal samples with those in the testing set, and highly overlapping adjacent windows were never split across different subsets, avoiding random window-level partitioning and cutting down the risk of data leakage from overlapping windows. All gesture classes remained present in both the training and testing subsets as needed for standard supervised classification evaluation. Once window segmentation was complete, the final training set held roughly 68,800 window samples, while the testing set had about 17,200 window samples. For the public datasets, we followed the same rule of avoiding random window-level splitting, using the available repetition or temporal information to guide the partitioning process consistently.

The DB2 dataset contains six repetitions for each gesture. In this study, only the first five repetitions are used for evaluation on the public dataset, among which repetitions 1–4 are assigned to the training set and repetition 5 to the test set. For the DB8 dataset, each gesture of every subject includes 20 repetitions, following the same 80%/20% repetition-level split protocol. For both public datasets, repetition-level data partitioning is implemented first. Sliding window segmentation is then conducted independently within the training and test subsets, and no window is allowed to cross the boundary of different repetitions. In addition, to enable comparison with experimental protocols reported in existing literature, we conduct supplementary experiments on DB2 using a trial-based partitioning strategy adopted from previous studies [21]. Specifically, trials 1, 3, 4, and 6 are used for training, while trials 2 and 5 are reserved for testing, with a sliding window configuration of 150 ms window length and 50 ms step size.

2.2. Multi-Scale Convolutional Block Dual Attention Network (MSCB-DualAttn)

2.2.1. Complete Model Framework

This study proposes a Multi-Scale Convolutional Block Dual Attention Network (MSCB-DualAttn) for gesture recognition using EMG signals in individuals with transradial amputation and healthy subjects. The network fully considers the temporal characteristics of EMG signals, multi-scale muscle activation patterns, and the importance differences across channels and time domains, automatically extracting discriminative features from raw EMG signals through an end-to-end learning approach.

As shown in Figure 3, the MSCB-DualAttn network adopts a hierarchical feature extraction design, consisting of an input layer, a stem initial convolutional layer, two cascaded Multi-Scale Convolutional Blocks (MSCB), a dual attention mechanism module, a global average pooling layer, and a classification layer. The input data is a preprocessed 8-channel EMG signal window $\mathrm{X}\in {\mathbb{R}}^{8\times 250}$, corresponding to a 250 ms signal segment at a sampling rate of 1000 Hz. This window length is sufficient to capture complete muscle activation patterns while meeting the latency requirements of real-time recognition. The overall forward propagation process of the network can be formally expressed as:

$${\mathrm{F}}_0={\mathrm{f}}_{\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{m}}(\mathrm{X})\in {\mathbb{R}}^{64\times 250}$$

(1)

$${\mathrm{F}}_1={\mathrm{f}}_{\mathrm{M}\mathrm{S}\mathrm{C}\mathrm{B}1}({\mathrm{F}}_0)\in {\mathbb{R}}^{128\times 250}$$

(2)

$${\mathrm{F}}_1^{\mathrm{p}}={\mathrm{M}\mathrm{a}\mathrm{x}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}}_{\mathrm{k}=2}({\mathrm{F}}_1)\in {\mathbb{R}}^{128\times 125}$$

(3)

$${\mathrm{F}}_2={\mathrm{f}}_{\mathrm{M}\mathrm{S}\mathrm{C}\mathrm{B}2}({\mathrm{F}}_1^{\mathrm{p}})\in {\mathbb{R}}^{256\times 125}$$

(4)

$${\mathrm{F}}_{\mathrm{a}\mathrm{t}\mathrm{t}}={\mathrm{f}}_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}({\mathrm{F}}_2)\in {\mathbb{R}}^{256\times 125}$$

(5)

$${\mathrm{F}}_{\mathrm{g}\mathrm{a}\mathrm{p}}={\mathrm{A}\mathrm{d}\mathrm{a}\mathrm{p}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{A}\mathrm{v}\mathrm{g}\mathrm{P}\mathrm{o}\mathrm{o}\mathrm{l}}_{1\mathrm{d}(1)}({\mathrm{F}}_{\mathrm{a}\mathrm{t}\mathrm{t}})\in {\mathbb{R}}^{256}$$

(6)

$$\widehat{\mathrm{y}}=\mathrm{Softmax}(\mathrm{W}{\mathrm{F}}_{\mathrm{gap}}+\mathrm{b})\in {\mathbb{R}}^{15}$$

(7)

where ${\mathrm{f}}_{\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{m}},{\mathrm{f}}_{\mathrm{M}\mathrm{S}\mathrm{C}\mathrm{B}1},{\mathrm{f}}_{\mathrm{M}\mathrm{S}\mathrm{C}\mathrm{B}2} \mathrm{and} {\mathrm{f}}_{\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}$ represent the mapping functions of the stem layer, the two MSCB modules, and the dual attention mechanism, respectively. $\mathrm{W}\in {\mathbb{R}}^{15\times 256}$ and $\mathrm{b}\in {\mathbb{R}}^{15}$ denote the weights and bias of the fully connected classification layer.

2.2.2. Stem Layer

The Stem layer serves as the entry point of the network, responsible for preliminary feature extraction and channel transformation of the raw EMG signals. This layer employs a one-dimensional convolution with a kernel size of 3, mapping the 8-channel raw input to a 64-dimensional feature space. Padding is set to 1 to preserve the temporal dimension. Subsequently, the output passes through batch normalization and a ReLU activation function to obtain the initial feature map ${\mathrm{F}}_0$. The batch normalization layer standardizes the features of each channel, accelerating model convergence and alleviating overfitting. Its computational process is as follows:

$$\mathrm{B}\mathrm{N}(\mathrm{x})=\mathsf{\gamma }{\displaystyle \frac{\mathrm{x}-{\mathsf{\mu }}_{\mathrm{B}}}{\sqrt{{\mathrm{\sigma }}_{\mathrm{B}}^2+\mathsf{ϵ}}}}+\mathsf{\beta }$$

(8)

where ${\mathsf{\mu }}_{\mathcal{B}}$ and ${\mathrm{\sigma }}_{\mathcal{B}}^2$ are the mean and variance of the current batch, respectively; $\mathsf{\gamma }$ and $\mathsf{\beta }$ are learnable scaling and shifting parameters; and $\mathsf{ϵ}$ is a small constant to prevent division by zero. The design of the Stem layer enables the network to project the signal into a high-dimensional space that is more suitable for subsequent multi-scale processing while preserving the original temporal information.

2.2.3. MSCB

EMG signals contain muscle activation patterns across different frequency components, making it difficult for a single-scale convolutional kernel to simultaneously capture the transient firings of fast-twitch muscle fibers and the sustained activations of slow-twitch fibers. To address this, this study designs a Multi-Scale Convolutional Block (MSCB), which extracts multi-scale temporal features through parallel convolutional branches with different receptive fields. The structure of each MSCB is illustrated in Figure 4.

The three branches employ one-dimensional convolutions with kernel sizes of k = 3, 5, and 7, respectively.

A kernel size of 3 corresponds to 3 ms, which matches the duration of the rising phase of a single motor unit action potential (MUAP). In typical surface EMG signals, the rise time of a MUAP is approximately 2–4 ms [33]. The branch with a kernel size of 3 is capable of capturing the rapidly rising edge and peak of the EMG signal. It is sensitive to the onset timing and transient firing of motor unit action potentials, facilitating the precise temporal identification of muscle activation.

A kernel size of 5 corresponds to 5 ms, which covers the complete biphasic waveform cycle of a MUAP. The typical duration of a MUAP is approximately 5–10 ms [34]. The branch with a kernel size of 5 is able to capture the waveform morphology features of the MUAP, including peak-to-peak amplitude, waveform width, and zero-crossing points, which are closely related to the recruitment patterns and firing frequency of motor units.

A kernel size of 7 (corresponding to 7 ms) aligns with the synchronized discharge window of multiple motor units and the synchronous activation period of muscle fiber groups. Studies have shown that motor units exhibit clustered discharge patterns during sustained muscle contraction [35]. The branch with a kernel size of 7 can integrate the superimposed effects of multiple MUAPs, capturing the slow-varying components of the EMG signal associated with force output, muscle fatigue, and synergistic contraction, thereby reflecting a more macroscopic pattern of muscle activation.

The corresponding padding values are set to $\mathrm{p}=1,2,3$ respectively, to ensure that the output temporal dimension remains consistent with the input. Let the input feature map be ${\mathrm{F}}_{\mathrm{i}\mathrm{n}}\in {\mathbb{R}}^{\mathrm{C}\times \mathrm{T}}$. The number of output channels for each branch is set to ${\mathrm{C}}_{\mathrm{b}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{h}}=\lfloor {\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}/3\rfloor$, where ${\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$ is the target number of output channels for the MSCB. The operation of each branch can be expressed as:

$${\mathrm{F}}_{\mathrm{branch}}^{\mathrm{i}} = \mathrm{ReLU}(\mathrm{BN}({\mathrm{Conv}}_{{\mathrm{k}}_{\mathrm{i}},{\mathrm{p}}_{\mathrm{i}}}({\mathrm{F}}_{\mathrm{in}}))), \mathrm{i} = 1,2,3$$

(9)

where ki ∈ 3, 5, 7 and pi ∈ 1, 2, 3.

For the first MSCB (MSCB1), the input channels are $\mathrm{C}=64$ and the output channels are ${\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}=128$, so the number of channels for each branch is ${\mathrm{C}}_{\mathrm{b}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{h}}=42$. For the second MSCB (MSCB2), the input channels are $\mathrm{C}=128$ and the output channels are ${\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}=256$, so the number of channels for each branch is ${\mathrm{C}}_{\mathrm{b}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{h}}=85$ (rounded down to 85, with the total of three branches being 255, which is then adjusted to 256 through a subsequent 1 × 1 convolution).

Feature Fusion: The outputs of the three branches are concatenated along the channel dimension to obtain a multi-scale feature map ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}}\in {\mathbb{R}}^{3{\mathrm{C}}_{\mathrm{b}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{h}}\times \mathrm{T}}$. To fuse features from different scales and adjust them to the target number of channels, a 1 × 1 convolution is employed for cross-channel information interaction and dimension transformation:

$${\mathrm{F}}_{\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{e}\mathrm{d}}={\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}_{1\times 1}({\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}})\in {\mathbb{R}}^{{\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\times \mathrm{T}}$$

(10)

Between the two MSCB, a max pooling layer with a kernel size of 2 and a stride of 2 is inserted to downsample the temporal dimension of the feature map from 250 to 125. This operation, on one hand, expands the receptive field of subsequent convolutional layers, enabling the network to capture longer-term muscle synergy patterns; on the other hand, it reduces the resolution of the feature map, thereby decreasing computational cost and the number of parameters, which helps prevent overfitting.

2.2.4. Dual Attention Block

Different channels of EMG signals correspond to different forearm muscles, and the activation intensity of each muscle varies significantly across gestures. Meanwhile, different phases of gesture execution contribute differently to classification. To enable the network to adaptively focus on important channels and critical time steps, a Dual Attention Mechanism Module is introduced after the second MSCB, consisting of two submodules: Channel Attention and Temporal Attention (Figure 5).

Channel Attention Module: This module aims to learn the importance weights of each channel, enhancing the responses of useful channels while suppressing the influence of noisy or irrelevant ones. Given the input feature map F ∈ ℝ256 × 125, global average pooling is first applied to aggregate temporal information for each channel, yielding a channel descriptor $\mathrm{z}\in {\mathbb{R}}^{256}$:

$${\mathrm{z}}_{\mathrm{e}}={\displaystyle \frac{1}{\mathrm{T}}}{\displaystyle \sum _{\mathrm{t}=1}^{\mathrm{T}}}{\mathrm{F}}_{\mathrm{c},\mathrm{t}}$$

(11)

To capture the nonlinear dependencies among channels, a gating mechanism consisting of two 1 × 1 convolutional layers is employed. The first convolutional layer reduces the number of channels to $\mathrm{C}/\mathrm{r}$ (with reduction ratio $\mathrm{r}=8$), followed by a ReLU activation. The second convolutional layer restores the channel dimension to $\mathrm{C}$. Finally, a Sigmoid function maps the output to the range $[0, 1]$, producing the channel attention weights ${\mathrm{w}}_{\mathrm{c}}\in {\mathbb{R}}^{256}$:

$${\mathrm{w}}_{\mathrm{c}}=\mathrm{\sigma }(\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}(\mathsf{\delta }(\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}(\mathrm{z}))))$$

(12)

where $\mathrm{\sigma }$ denotes the Sigmoid function and $\mathsf{\delta }$ denotes the ReLU function. The attention weights are then multiplied element-wise with the original feature map to achieve feature recalibration:

$${\mathrm{F}}_{\mathrm{c}}=\mathrm{F}\odot {\mathrm{w}}_{\mathrm{c}}$$

(13)

where $\odot$ denotes element-wise multiplication along the channel dimension, with ${\mathrm{w}}_{\mathrm{c}}$ being broadcast across the temporal dimension. This mechanism enables the network to dynamically adjust the response intensity of each channel based on global information, highlighting the muscle channels relevant to the current gesture.

Temporal Attention Module: Based on the channel-weighted features, the temporal attention module further learns the importance of each time step. A 1 × 1 convolutional layer is used to compress the 256 channels into a single channel, and the output is passed through a Sigmoid activation to generate the temporal attention weight vector ${\mathrm{w}}_{\mathrm{t}}\in {\mathbb{R}}^{125}$:

$${\mathrm{w}}_{\mathrm{t}}=\mathrm{\sigma }(\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}({\mathrm{F}}_{\mathrm{c}}))$$

(14)

This weight vector reflects the contribution of each time step to the classification task, with key phases of gesture execution receiving higher weights. The temporal weights are then multiplied element-wise with the channel-weighted feature map along the time dimension:

$${\mathrm{F}}_{\mathrm{a}\mathrm{t}\mathrm{t}}={\mathrm{F}}_{\mathrm{c}}\odot {\mathrm{w}}_{\mathrm{t}}$$

(15)

where ${\mathrm{w}}_{\mathrm{t}}$ is broadcast across the channel dimension.

This cascaded design of the dual attention mechanism enables the network to adaptively focus on discriminative information from both the channel and temporal dimensions, effectively enhancing feature representation and classification performance.

2.2.5. Classifier and Loss Function Design

The feature map ${\mathrm{F}}_{\mathrm{a}\mathrm{t}\mathrm{t}}\in {\mathbb{R}}^{256\times 125}$ enhanced by the dual attention mechanism is passed through a global average pooling layer to compress the temporal dimension into a single value, yielding a global feature vector $\mathrm{v}\in {\mathbb{R}}^{256}$:

$${\mathrm{v}}_{\mathrm{c}}={\displaystyle \frac{1}{\mathrm{T}}}{\displaystyle \sum _{\mathrm{t}=1}^{\mathrm{T}}}{\mathrm{F}}_{\mathrm{a}\mathrm{t}\mathrm{t}}^{\mathrm{c},\mathrm{t}}$$

(16)

This operation preserves channel-wise features while eliminating temporal positional information, making the model robust to time shifts in the input signal. The resulting feature vector is then fed into a fully connected classification layer to map it to logits for the 15 gesture classes:

$$\mathrm{o}=\mathrm{W}\mathrm{v}+\mathrm{b},\hspace{1em}\mathrm{W}\in {\mathbb{R}}^{15\times 256},\mathrm{b}\in {\mathbb{R}}^{15}$$

(17)

During the training phase, the cross-entropy loss is adopted to measure the discrepancy between the predicted probability distribution and the ground truth labels:

$$\mathcal{L}=-{\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}}{\displaystyle \sum _{\mathrm{j}=1}^{15}}{\mathrm{y}}_{\mathrm{i},\mathrm{j}}\log ({\widehat{\mathrm{y}}}_{\mathrm{i},\mathrm{j}})$$

(18)

$$\widehat{\mathrm{y}}=\arg \underset{\mathrm{j}}{\max }{\displaystyle \frac{\exp ({\mathrm{o}}_{\mathrm{j}})}{{\displaystyle {\sum }_{\mathrm{k}=1}^{15}}\exp ({\mathrm{o}}_{\mathrm{k}})}}$$

(19)

where N is the batch size, ${\mathrm{y}}_{\mathrm{i},\mathrm{j}}$ denotes the one-hot encoding of the ground truth label for the i-th sample, and ${\widehat{\mathrm{y}}}_{\mathrm{i},\mathrm{j}}=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}({\mathrm{o}}_{\mathrm{i}}{)}_{\mathrm{j}}$ represents the predicted probability. During the inference stage, the logits are converted into a probability distribution via the Softmax function, and the class with the highest probability is selected as the final prediction.

2.3. Model Evaluation Metrics and Interpretability Analysis

2.3.1. Model Evaluation Metrics

In this study, five metrics were adopted to comprehensively evaluate the performance of the model: Accuracy, F1-Score, Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE).

Accuracy measures the proportion of correctly predicted samples among all samples and is defined as:

$$\mathrm{Accuracy}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}$$

(20)

where TP, TN, FP, and FN denote the numbers of true positives, true negatives, false positives, and false negatives, respectively.

The F1-Score is the harmonic mean of Precision and Recall, balancing the relationship between the two and being particularly suitable for imbalanced class scenarios. For each class i, the F1-Score is defined as:

$$\mathrm{F}{1}_{\mathrm{i}}=2\times {\displaystyle \frac{\mathrm{Precisio}{\mathrm{n}}_{\mathrm{i}}\times \mathrm{Recal}{\mathrm{l}}_{\mathrm{i}}}{\mathrm{Precisio}{\mathrm{n}}_{\mathrm{i}}+\mathrm{Recal}{\mathrm{l}}_{\mathrm{i}}}}$$

(21)

In this paper, the Macro F1-Score is adopted, which is the arithmetic mean of the F1 scores across all classes, to evaluate the overall performance of the model across different classes.

Meanwhile, MSE, RMSE and MAE were further utilized to quantify the feature distribution deviation and myoelectric pattern fitting error among groups with different residual limb lengths, providing quantitative evidence for analyzing neuromuscular compensation differences and model performance degradation. Given the one-hot encoding ${\mathrm{y}}_{\mathrm{i}}$ of the ground truth labels and the predicted probabilities ${\mathrm{p}}_{\mathrm{i}}$ output by the model, these metrics are defined as follows:

$$\mathrm{MSE}={\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}}({\mathrm{y}}_{\mathrm{i}}-{\mathrm{p}}_{\mathrm{i}}{)}^2$$

(22)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}}({\mathrm{y}}_{\mathrm{i}}-{\mathrm{p}}_{\mathrm{i}}{)}^2}$$

(23)

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{N}}}|{\mathrm{y}}_{\mathrm{i}}-{\mathrm{p}}_{\mathrm{i}}|$$

(24)

where N is the total number of samples. These three metrics reflect the confidence and calibration of the model’s predictions, with smaller values indicating that the predicted probability distribution is closer to the ground truth labels.

To intuitively illustrate the classification details of the model in the multi-class gesture recognition task, a confusion matrix is employed to visualize the prediction results.

2.3.2. Ablation Study

To validate the effectiveness of the proposed modules, two comparison models were constructed in this study.

The Baseline Convolutional Neural Network (BaselineCNN) adopts a traditional serial convolution-pooling structure without multi-scale branches or attention mechanisms. Its specific architecture is as follows: The first convolutional layer maps the input from 8 to 64 channels with a kernel size of 3, followed by batch normalization, ReLU activation, and a max pooling layer with a kernel size of 2, resulting in an output feature map of size 64 × 125. The second convolutional layer increases the number of channels from 64 to 128 with a kernel size of 3, also followed by batch normalization, ReLU, and max pooling, yielding an output size of 128 × 62. The third convolutional layer increases the number of channels from 128 to 256 with a kernel size of 3, followed by batch normalization, ReLU, and global average pooling, producing a 256-dimensional feature vector. Finally, a fully connected layer maps the features to 15-class outputs.

The MSCB-NoAttn Model was constructed to separately evaluate the contribution of the multi-scale convolutional module to EMG signal feature extraction. This model retains the basic network framework but replaces the single-scale convolutional layers in the original BaselineCNN with MSCB, without introducing any attention mechanisms.

2.3.3. t-SNE Visualization

To intuitively evaluate the discriminability of the features extracted by the model, t-Distributed Stochastic Neighbor Embedding (t-SNE) was employed for dimensionality reduction and visualization of high-dimensional features. t-SNE is a nonlinear dimensionality reduction technique that maps sample points from a high-dimensional space into a two- or three-dimensional space while preserving the local similarity structure among samples. Specifically, the feature map $\mathrm{F} \in { \mathrm{R}}^{256\times 125}$ before the global average pooling layer was extracted from the trained MSCB-DualAttn network, and global average pooling was applied along the temporal dimension to obtain a 256-dimensional feature vector as the input to t-SNE. The perplexity was set to 30, the learning rate to 200, and the optimization was performed for 1000 iterations. By comparing the feature distributions of different models, differences in intra-class compactness and inter-class separability can be intuitively assessed.

2.4. Experimental Training Settings and Hyperparameters

As shown in

Table 2. Experimental Training Settings and Hyperparameters.

Configuration Item	MSCB-DualAttn & MSCB-NoAttn	Baseline CNN
Optimizer	Adam	Adam
Initial Learning Rate	0.001	0.001
Batch Size	128	128
Training Epochs	50	30
Loss Function	Cross-Entropy Loss	Cross-Entropy Loss
Learning Rate Scheduler	ReduceLROnPlateau (patience = 5, monitored on val accuracy)	Not used
Random Seed	42	42

, the network models in this study were all built and trained based on the PyTorch framework. Consistent training hyperparameters were adopted across models: the Adam optimizer was used with an initial learning rate of 0.001, and cross-entropy loss was taken as the classification loss function. For the proposed MSCB-DualAttn and its ablation variant, the batch size was set to 128 with 50 training epochs. A ReduceLROnPlateau learning rate scheduler was employed, which adjusted the learning rate dynamically according to validation accuracy with a patience of 5 epochs. The baseline CNN model adopted the same optimizer, learning rate and batch size, and was trained for 30 epochs without a learning rate scheduler. All experiments were implemented on a Windows 10 workstation with an Intel processor and an NVIDIA GeForce GTX 1050 Ti GPU. The deep learning framework was PyTorch 2.4.1 with CUDA 12.4 and cuDNN 90100 support. Python version 3.8.20 was used, together with NumPy 1.24.4, Pandas 2.0.3, scikit-learn 1.3.0, SciPy 1.10.1, and Matplotlib 3.7.2 for data processing, model evaluation, and visualization.

3. Results

3.1. Overall Performance Evaluation Across Different Groups

Table 3. Group-wise Performance.

Category	Metric	Mean ± Std	95% Confidence Interval
HS	Accuracy	0.9379 ± 0.0348	[0.9220, 0.9537]
	F1	0.9349 ± 0.0384	[0.9175, 0.9524]
	MSE	0.0066 ± 0.0038	[0.0049, 0.0084]
	RMSE	0.0814 ± 0.0230	[0.0678, 0.0888]
	MAE	0.0092 ± 0.0044	[0.0072, 0.0112]
LS	Accuracy	0.8614 ± 0.0677	[0.7773, 0.9454]
	F1	0.8541 ± 0.0695	[0.7678, 0.9404]
	MSE	0.0148 ± 0.0069	[0.0062, 0.0233]
	RMSE	0.1216 ± 0.0320	[0.0786, 0.1579]
	MAE	0.0200 ± 0.0095	[0.0082, 0.0319]
MS	Accuracy	0.7809 ± 0.0718	[0.7146, 0.8473]
	F1	0.7660 ± 0.0821	[0.6901, 0.8421]
	MSE	0.0242 ± 0.0083	[0.0165, 0.0319]
	RMSE	0.1555 ± 0.0259	[0.1297, 0.1776]
	MAE	0.0315 ± 0.0093	[0.0229, 0.0401]
SS	Accuracy	0.6269 ± 0.0896	[0.5520, 0.7019]
	F1	0.6066 ± 0.0943	[0.5277, 0.6855]
	MSE	0.0380 ± 0.0096	[0.0300, 0.0460]
	RMSE	0.1950 ± 0.0247	[0.1730, 0.2143]
	MAE	0.0537 ± 0.0117	[0.0439, 0.0634]

HS represents the healthy subjects group, LS represents the long stump group, MS represents the middle stump group, and SS represents the short-stump group.

summarizes the overall recognition and regression performance of the MSCB-Attention network across the healthy group (HS), long residual limb group (LS), medium residual limb group (MS), and short residual limb group (SS). Clear performance disparities can be observed among the four groups, following a consistent pattern that model accuracy and F1 score gradually decline as residual limb length decreases, while regression errors rise steadily.

The healthy group yields the best overall performance. It achieves a mean accuracy of 93.79% and a mean F1-score of 0.9349. Accuracy presents low standard deviations (±0.0348) and narrow 95% confidence intervals of [0.9220, 0.9537], while F1-score has a standard deviation of ±0.0384 with a 95% confidence interval of [0.9175, 0.9524]. The corresponding MSE, RMSE and MAE are 0.0066, 0.0814 and 0.0092, respectively, also accompanied by small fluctuations and tight confidence intervals. Such results demonstrate that healthy participants have complete forearm muscle structure and stable sEMG signal quality, enabling the model to extract highly distinguishable gesture features with reliable and stable prediction performance. The long residual limb group ranks second in performance, with a mean accuracy of 86.14% and a mean F1-score of 0.8541. Its accuracy standard deviation increases to ±0.0677, and the 95% confidence interval expands to [0.7773, 0.9454]; the F1-score shows a standard deviation of ±0.0695 and a 95% confidence interval of [0.7678, 0.9404]. All error metrics are slightly higher than those of the healthy group but remain at a relatively low level. This indicates that patients with long residual limbs still preserve considerable muscle activation and synergy information, and the proposed model can effectively capture effective sEMG characteristics for gesture classification. For the medium residual limb group, the mean accuracy and F1-score drop to 78.09% and 0.7660, with a higher standard deviation of ±0.0718 for accuracy and ±0.0821 for F1-score, and wider corresponding 95% confidence intervals of [0.7146, 0.8473] and [0.6901, 0.8421] respectively. Noticeable increases in MSE, RMSE and MAE are also observed. The degraded recognition accuracy and enlarged performance fluctuation reflect the evident decline in sEMG signal completeness and stability, which substantially weakens the model’s feature representation capability. The short residual limb group obtains the lowest performance across all metrics, with mean accuracy and F1-score only reaching 62.69% and 0.6066. It presents the largest standard deviation (±0.0896 for accuracy, ±0.0943 for F1-score) and the broadest 95% confidence intervals [0.5520, 0.7019] and [0.5277, 0.6855], indicating poor stability of recognition results. Meanwhile, MSE, RMSE and MAE rise sharply compared with the healthy group. Short residual limbs lead to severe loss of forearm muscle tissue, fewer available signal acquisition sites, and weaker and more unstable sEMG signals. These factors greatly increase the difficulty of gesture recognition and bring large performance variability across individuals.

To evaluate the gesture classification performance of the MSCB-DualAttn model across populations with different residual limb lengths, confusion matrices were plotted for the four groups—healthy subjects, long stump, middle stump, and short stump—as shown in Figure 6. For the healthy group, characterized by standardized muscle synergy patterns and optimal signal quality, the model achieved the best classification performance. The confusion matrix shows high and densely distributed diagonal elements, with very few misclassified samples across categories, and only occasional confusion among a few similar gestures. This indicates that the model accurately captures standard muscle activation features and exhibits strong gesture recognition stability. In the long stump group, where limb loss is relatively mild and muscle compensation is moderate, the model performed well. The confusion matrix reveals high diagonal values, with most gesture categories accurately recognized. Only a small number of samples were misclassified among boundary categories, and the overall classification accuracy approached that of the healthy group, validating the model’s ability to adapt to mild compensatory differences. For the middle stump group, limb loss and muscle compensation are more pronounced, increasing recognition difficulty. In this group, the confusion matrix still shows relatively high diagonal values, with a slight increase in the number of misclassified samples. However, category boundaries remain clear, and no widespread class confusion occurs, indicating that the model can effectively extract compensation-related features in middle stump patients and maintain stable recognition performance. The short-stump group, characterized by the most severe limb loss, complex muscle compensation patterns, and poor signal quality, represents the most challenging group for recognition. Although the confusion matrix for this group contains a relatively larger number of misclassified samples, the diagonal elements still dominate, and the model does not exhibit complete confusion. This demonstrates that even in extreme signal-deficient scenarios, the model can still extract discriminative features and adapt to highly individualized compensatory differences.

3.2. Ablation Study Results

To validate the effectiveness of each component in the MSCB-DualAttn network, this study observed changes in model performance by gradually adding or removing key modules. The experimental results are shown in Figure 7 and Figure 8, with all experiments conducted under the same dataset and training conditions.

Figure 7 shows that for the healthy group, characterized by standardized muscle synergy patterns and optimal signal quality, all three models achieved high recognition accuracy. Among them, the complete MSCB-DualAttn model achieved an accuracy of 93.79%, representing improvements of 2.91 and 5.03 percentage points over MSCB without attention (90.88%) and the baseline CNN (88.76%), respectively, indicating that multi-scale convolution combined with attention mechanisms can further explore discriminative features for gestures. In the long stump group, the complete MSCB-DualAttn model achieved an accuracy of 86.13%, representing improvements of 10.71 and 9.46 percentage points over MSCB without attention (75.42%) and the baseline CNN (76.67%), respectively. Notably, in this group, MSCB without attention performed slightly worse than the baseline CNN, suggesting that multi-scale convolution alone may introduce redundant features, while the attention mechanism effectively recalibrates channel weights, filters out interfering information, and achieves a significant performance leap. In the middle stump group, the complete model achieved an accuracy of 78.10%, representing improvements of 7.62 and 8.54 percentage points over MSCB without attention (70.48%) and the baseline CNN (69.56%), respectively. Meanwhile, error metrics such as MSE and RMSE were significantly reduced, indicating higher prediction confidence. The short-stump group, characterized by the most severe limb loss and the most complex muscle compensation, represents the most challenging recognition scenario. The complete MSCB-DualAttn model achieved an accuracy of 62.69% in this group. Although the absolute value is relatively low, it represents improvements of 6.94 and 13.83 percentage points over MSCB without attention (55.75%) and the baseline CNN (48.86%), respectively, validating the model’s ability to extract effective features even in extreme signal-deficient scenarios.

As shown in the overall confusion matrices in Figure 8a, the BaselineCNN model exhibited substantial inter-class confusion. The MSCB (without attention) model improved feature discriminability through multi-scale convolution, while the MSCB-DualAttn model achieved the highest diagonal values and the fewest misclassified samples, demonstrating the best overall classification performance. The ROC curve comparison for the short-stump group (Figure 8b) further indicates that BaselineCNN had a relatively weak ability to discriminate gestures under compensatory patterns, with an average AUC of only 0.78. The MSCB (without attention) model showed improved performance, achieving an average AUC of 0.85. In contrast, the ROC curves for various gestures in the MSCB-DualAttn model were all significantly farther from the diagonal line, with the average AUC increasing to 0.91. This demonstrates the model’s enhanced capability to capture individualized muscle compensation patterns in short-stump patients, exhibiting stronger robustness and generalization ability in challenging gesture recognition tasks.

3.3. t-SNE Analysis Results

The results are shown in Figure 9. The feature distribution of the BaselineCNN model exhibits severe inter-class overlap, with point clouds of different colors intermingled and ambiguous boundaries across a large number of gesture categories, indicating significant cross-class overlap. Sample points within the same gesture category are loosely distributed and fail to form compact clusters. Features of some gesture categories (e.g., labels 0, 2, and 4) are scattered throughout the space, making it difficult for the model to effectively distinguish between different gesture patterns. This suggests that the BaselineCNN model can only extract basic temporal features and cannot fully capture fine-grained differences across gestures and populations in EMG signals, resulting in limited feature discriminability. Compared with the other two models, the feature distribution of the MSCB-DualAttn model shows improved intra-class compactness, with sample points within the same gesture category forming clearer and more independent clusters (e.g., categories such as labels 0, 2, 4, and 14 all form tight clusters). Inter-class separation is also enhanced, with clearer boundaries between clusters of different gesture categories, thereby improving feature discriminability. The overall feature distribution exhibits a more structured clustering pattern. This result indicates that through the synergistic effect of multi-scale convolutional blocks and dual attention mechanisms, the MSCB-DualAttn network can effectively extract discriminative features for different gestures and populations from EMG signals, providing a reliable feature foundation for high-precision gesture recognition.

3.4. Comparison of Saliency Across MSCB Branches

As shown in Figure 10, the saliency distributions of the three branches exhibit distinctly complementary characteristics. The saliency of the small kernel branch (Branch 1, k = 3) appears as sharp peaks only at a few specific time steps (e.g., around steps 20 and 80), while the importance at other time steps approaches zero, indicating that this branch primarily focuses on local abrupt changes and is sensitive to transient activation patterns in EMG signals. The saliency distribution of the medium kernel branch (Branch 2, k = 5) is relatively uniform, with the highest number of peaks covering almost the entire temporal range, suggesting that this branch balances local details with medium-range temporal dependencies and is well-suited for capturing periodic fluctuations and sustained activation patterns. The saliency of the large kernel branch (Branch 3, k = 7) exhibits fewer peaks but with higher individual peak magnitudes, mainly concentrated around steps 30 and 90, indicating that this branch pays more attention to global patterns over longer temporal scales and is appropriate for capturing slow changes or long-range dependencies in muscle synergies.

3.5. Gradient Importance Saliency

As shown in Figure 11, the importance curve exhibits significant fluctuations across the entire time axis rather than remaining at a stable low value, indicating that the model does not ignore any time period but continuously encodes and processes the temporal dynamics of the EMG signal. Multiple prominent response peaks are distributed along the curve, with particularly prominent high-value regions near time steps 100, 150, and 200, where the peak around step 200 reaches the highest value. These regions represent the most critical temporal segments in the model’s decision-making process, suggesting that the model captures decisive muscle activation patterns, event-related changes, or transient features at these positions. In addition to the main peaks, numerous secondary peaks are densely distributed and cover a wide range, indicating that the model not only focuses on single key moments but also effectively captures local details and dynamic evolution characteristics within the temporal data. This multi-peak response pattern reflects the model’s ability to finely model multi-stage muscle synergy activities in EMG signals.

3.6. Channel Attention Weight Distribution

As shown in Figure 12, in terms of the average channel attention weights across all samples, the MSCB-DualAttn model does not exhibit an extreme preference for any single channel. The overall weight distribution ranges from 0.47 to 0.55, demonstrating that the model comprehensively utilizes multi-channel information. Among all channels, ch3 (FDS) achieves the highest average weight of 0.54, followed by ch5 (ECU) with 0.53. This indicates that the flexor digitorum superficialis and extensor carpi ulnaris provide the most discriminative features for overall gesture classification.

In the grouped analysis, muscle contribution patterns show obvious individual specificity across different subjects. Subjects in the LS rely most heavily on ch3 (FDS), with a weight as high as 0.59, which is significantly higher than that of other channels; meanwhile, the weight of ch1 (B) reaches 0.56, suggesting that this subject mainly depends on the coordinated activation of the flexor digitorum superficialis and biceps brachii. The MS shows a relatively balanced weight distribution but presents a local peak at ch4 (FCU) with a weight of 0.54. The HS also attaches great importance to ch3 (FDS) with a weight of 0.58, and the weight of ch8 (MB) rises to 0.56, reflecting the vital role of upper arm muscles in the activation pattern of this group. Differently, the SS displays a unique dependence on distal muscles: its ch8 (MB) weight is the highest among the four groups at 0.55, while ch1 (B) has the lowest weight of only 0.42.

Although FDS (ch3) serves as the dominant channel in most cases, the model can adaptively capture differentiated features on channels such as BB (ch1), FCU (ch4), and MD (ch8) across different individuals. Furthermore, the SS shows an inconsistent dominant muscle activation pattern compared with the other groups.

3.7. Comparison with Other Studies

As shown in

Table 4. Comparison with Other Studies.

Model	Input	Sliding Window Parameters	Partition Scheme	F1-Score	Accuracy (%)	Recall (Macro)	Data
MSCB-DualAttn	sEMG raw data	Window 250 ms, step 100 ms	Temporally ordered 80%/20% split	93.89	95.81	96.21	DB8
MSCB-DualAttn	sEMG raw data	Window 250 ms, step 100 ms	Temporally ordered 80%/20% split	85.94	86.42	87.12	DB2
MSCB-DualAttn	sEMG raw data	Window 150 ms, step 50 ms	Trials 1/3/4/6 Train, 2/5 Test	86.94	88.79	88.93	DB2
CviT [19]	Time domain And Frequency domain Data	Window 200 ms, step 100 ms	Trials 1/3/4/6 Train, 2/5 Test	-	80.02	-	DB2
MSDS-FusionNet [20]	sEMG raw data	Window 200 ms, overlap 5 ms	Trials 1/3/4/6 Train, 2/5 Test	-	90.15	89.34	DB2
MSCANN [21]	sEMG raw data	Window 150 ms, step 50 ms	Trials 1/3/4/6 Train, 2/5 Test	-	93.80	-	DB2
RF (spectro-temporal feature) [36]	Spectro-temporal feature	Window 250 ms, overlap ratio 10%	10-fold cross-validation	-	98.16	-	DB8
CNN-MSTINet [38]	sEMG raw data	Window 200 ms	Trials 1/3/4/6 Train, 2/5 Test	-	85.77	-	DB2
LST-EMG-Net [39]	Signal augmentation based on sEMG window	Window 300 ms, step 10 ms	Trials 1/3/4/6 Train, 2/5 Test	/	81.47	-	DB2
MCMP-Net [40]	Rms calculation	Window 20 points, sliding RMS	Trials 1/3/4/6 Train, 2/5 Test	86.4	84.50	88.3	DB2
STMS-Net [41]	Rms calculation	Window 20 points, sliding RMS	Trials 1/3/4/6 Train, 2/5 Test	90.1	90.00	91.8	DB2
SVM [37]	Spectral Features	Window 400 points, step 20 points	1st set Train, 2nd set Test	-	85.50	-	DB8 (5 subjects)
CNN [37]	sEMG raw data	Window 400 points, step 20 points	1st set Train, 2nd set Test	-	83.25	-	DB8 (5 subjects)
LSTM [37]	sEMG raw data	Window 400 points, step 20 points	1st set Train, 2nd set Test	-	95.10	-	DB8 (5 subjects)
FDM [37]	sEMG raw data	Window 400 points, step 20 points	1st set Train, 2nd set Test	-	93.53	-	DB8 (5 subjects)
MFFCNN-LSTM [37]	sEMG raw data	Window 400 points, step 20 points	1st set Train, 2nd set Test	-	98.50	-	DB8 (5 subjects)
MS-CLSTM [24]	sEMG grayscale image	Window 200 ms, step 50 ms	1st set Train, 2nd set Test	-	86.66	-	DB2
MSCNN [42]	sEMG raw data	Window 100 ms, overlap ratio 75%	Train: 2/4/6; Val:1/5; Test:3	77.0	82.4	-	DB2

, on the full DB8 dataset, the proposed model yields an accuracy of 95.81%, an F1-score of 93.89%, and a Macro Recall of 96.21%. Compared with the spectro-temporal feature-based RF method [36] (98.16%) and the MFFCNN-LSTM model [37] (98.50%) reported in previous studies, our accuracy is slightly lower by 2.35 and 2.69 percentage points, respectively. Nevertheless, the RF method relies on complex manually designed spectro-temporal feature extraction based on VMD and SVD, which heavily depends on prior expert knowledge. In contrast, the proposed model adopts an end-to-end paradigm that directly learns discriminative representations from raw sEMG data, avoiding cumbersome manual feature engineering. Notably, the comparative MFFCNN-LSTM [37] only adopts a subset of DB8 with merely five healthy subjects, while our evaluation employs the complete DB8 dataset containing both healthy and amputee participants, which undoubtedly increases the recognition difficulty and better verifies model generalization capability. Under the same five-subject DB8 subset setting, the proposed method also prominently surpasses conventional SVM (85.50%), CNN (83.25%), LSTM (95.10%), and FDM (93.53%) in accuracy, and maintains superior comprehensive performance in F1-score and Macro Recall.

On the DB2 dataset with unified trial-based partitioning and sliding window settings, the proposed method reaches an accuracy of 86.42%, an F1-score of 85.94%, and a Macro Recall of 87.12%. In terms of comprehensive metric performance, it outperforms several mainstream baseline models, including CNN-MSTINet [38] (85.77%), LST-EMG-Net [39] (81.47%), and MCMP-Net [40] (84.50%) across Accuracy and F1-score. The ViT-based CviT method [19] only obtains an accuracy of 80.02%, and its image conversion strategy introduces high computational latency, limiting practical wearable deployment. Although MSDS-FusionNet [20] (90.15%) and MSCANN [21] (93.80%) achieve higher accuracy on DB2, they adopt serial multi-scale stacking structures that easily lose fine-grained sEMG temporal details and lack stable performance in F1-score and Macro Recall. Additionally, compared with STMS-Net [41] (90.00%), our method is slightly lower in accuracy by 3.58 percentage points but achieves comparable F1 and recall metrics, demonstrating acceptable competitive performance.

3.8. Analysis of Model Parameters

To further verify the real-time deployment capability for gesture recognition, we evaluate the computational complexity and inference efficiency of the proposed MSCB-DualAttn model. As listed in

Table 5. Model Parameters.

Evaluation Metric	MSCB-DualAttn
Total Parameters	308,967
Model Size (MB)	1.178616
Single-sample Inference Time (ms)	3.358563
Throughput (Samples per Second)	297.75

, the model has only 308,967 parameters with a memory size of about 1.18 MB. The single-sample inference latency is about 3.36 ms, and the throughput reaches 297.75 samples per second. The results demonstrate that the proposed model is lightweight, low-latency, and fully satisfies the requirement of real-time myoelectric gesture recognition.

4. Discussion

This study focuses on the problem of EMG-based gesture recognition in amputees with different residual limb lengths and healthy individuals, proposing a Multi-Scale Convolutional Block with Dual Attention Network (MSCB-DualAttn). Through systematic experimental validation and interpretability analysis, the model demonstrates significant advantages in both recognition performance and physiological interpretability.

From the perspective of group-wise performance, the recognition accuracy of the MSCB-DualAttn network exhibits a clear monotonic decreasing trend as residual limb length shortens: the healthy subject group (HS) achieves the highest accuracy of 93.79% ± 3.48% with a narrow 95% confidence interval of [0.9220, 0.9537], followed by the long stump group (LS) at 86.14% ± 6.77%, the middle stump group (MS) at 78.09% ± 7.18%, and the short-stump group (SS) obtaining the lowest accuracy of 62.69% ± 8.96%. This gradient-like degradation does not simply reflect changes in signal quality, but rather confirms, from a neural control perspective, the fundamental impact of residual limb length on muscle synergy structure and signal discriminability. The F1-score was slightly lower than accuracy across all groups but maintained a highly consistent numerical trend, indicating that the proposed model possesses excellent class balance and stable generalization capability. The model accurately extracts standard EMG features from healthy individuals, while the long stump group retains a certain degree of synergistic information, allowing the model to maintain robust recognition performance. For the middle stump group, muscle compensation increases and recognition difficulty rises, yet the model still effectively captures compensation-related features. In the short-stump group, where signals are severely deficient and compensatory patterns are complex, although the absolute recognition accuracy is relatively low, it still achieves a significant improvement of 13.83 percentage points over the baseline model, indicating that the model retains a certain capacity for feature extraction even in signal-deficient scenarios. In terms of regression error metrics, MSE, RMSE, and MAE show a monotonic upward trend with the reduction in residual limb length. The healthy group presents the minimum error (MSE = 0.0066 ± 0.0038, RMSE = 0.0814 ± 0.0230, MAE = 0.0092 ± 0.0044), while the short-stump group yields the highest error (MSE = 0.0380 ± 0.0096, RMSE = 0.1950 ± 0.0247, MAE = 0.0537 ± 0.0117). The gradually enlarged standard deviation and wider confidence interval further demonstrate that signal instability and individual variability increase significantly for amputees with shorter residual limbs.

Ren et al. [43] revealed a clear physiological correspondence to the hierarchical mechanism of neuromuscular control. Their study confirmed that the neural regulation of forearm flexors and extensors during hand gestures follows a multi-band and multi-scale temporal organization. The low-frequency delta band reflects global common inputs and integrated muscle synergies, whereas medium and high-frequency bands correspond to local transient activations and independent neural drives. In addition, Huesler et al. [44] reported significantly stronger motor unit synchrony in extrinsic hand muscles than in intrinsic ones. Datta et al. [45] further demonstrated that motor unit synchrony in the first dorsal interosseous was nearly twice as high as that in the tibialis anterior and approximately four times higher than that in the medial gastrocnemius. These findings collectively imply that a multi-scale convolutional architecture enables the capture of richer neuromuscular information. Within the MSCB module, the small-kernel branch specializes in extracting transient, local, and rapid muscle activations, corresponding to high-frequency independent neural control. The medium-kernel branch provides balanced coverage over the entire temporal dimension, matching stable and periodic activation patterns. The large-kernel branch focuses on slowly varying global trends and long-range muscle synergies, consistent with low-frequency common neural inputs. This consistency indicates that the multi-scale structure of the proposed model is not purely data-driven, but aligns with the inherent multi-scale characteristics of human neuromuscular control. Such physiological interpretability enhances the reliability and rationality of the model in myoelectric gesture recognition. Collectively, the three branches conform to the hierarchical organization of EMG signals from motor unit firing to macroscopic muscle synergies, providing a neurophysiological foundation for the model design. In the full network, the multi-peak structure of the temporal importance curve confirms that the model emphasizes critical moments while retaining local details and dynamic evolutionary characteristics. As demonstrated by Ren et al., hand gestures rely on the coordinated and independent neural control of forearm flexors and extensors, among which the flexor digitorum superficialis (FDS) and extensor muscle groups serve as key contributors to finger movements. In the proposed model, channels corresponding to FDS (ch3) and ECU (ch5) receive the highest attention weights, which is highly consistent with their physiological importance during gesture generation. Meanwhile, the adaptive shifts in muscle weighting across groups with different residual limb lengths—especially the reliance on distal muscles in the short-stump group and proximal muscles in the long-stump and healthy groups—reflect the flexible strategy of neuromuscular reorganization following amputation. Some studies have evoked wrist movements by electrically stimulating four key muscles: FDS, FCR, ECU, and ECR [32]. This indicates that only a small set of core muscles is sufficient to compose and execute predefined hand gestures [46]. These results also verify the synergistic value of the symmetric dual-attention mechanism in effective channel selection and fine-grained temporal focusing. In the short-stump group, the complete model outperforms the baseline CNN by 13.83 percentage points, the largest improvement among all groups. This suggests that the attention mechanism becomes increasingly critical under more severe muscle compensation and more disordered signal conditions. The two core components of MSCB-DualAttn work collaboratively: the multi-scale convolutional block captures diverse activation patterns across multiple temporal scales, and the dual-attention module further extracts the most discriminative components from these features.

However, the recognition accuracy of 62.69% for the short-stump group still leaves substantial room for improvement. As the most challenging subgroup for gesture recognition, the short-stump group suffers from the most severe limb deficiency, complex muscle compensation patterns, and severely degraded EMG signal quality, yielding the lowest recognition accuracy among the four groups, accompanied by a notable rise in error metrics. Specifically, its MSE and MAE are approximately 5.7 and 5.8 times those of healthy subjects, respectively. Nevertheless, the diagonal elements of its confusion matrix remain dominant without complete category aliasing. Moreover, the proposed model achieves a prominent accuracy improvement of 13.83 percentage points over the baseline CNN model, which is the largest increment across all four subgroups. This fully demonstrates that the MSCB-DualAttn model retains powerful capability for discriminative feature extraction for short-stump amputees and can adapt to highly individualized differences in muscle compensation. From the perspective of module ablation, the advantage of the MSCB multi-scale module is diminished in the short-stump group. This indicates that once the hierarchical structure of EMG signals is disrupted, multi-scale convolutional features no longer carry explicit physiological hierarchical implications and merely function to expand feature dimensionality. In contrast, the attention mechanism still plays a vital role under short-stump conditions: it boosts the accuracy from 55.75% in the MSCB_NoAttn variant to 62.69% in the complete model. This verifies that adaptive weighting along channel and temporal dimensions remains a robust feature enhancement strategy even under severe signal degradation, where the attention mechanism primarily captures stable latent patterns within muscle activation representations reconstructed by compensatory motor strategies. Further analysis of the confusion matrix reveals that gestures such as fist clenching maintain a high correct classification rate across healthy individuals and all amputee subgroups, whereas fine manipulation gestures, including three-finger pinch and palm external rotation, present relatively higher recognition difficulty. This phenomenon is closely associated with the inherent complexity of muscle activation patterns underlying different gesture categories, suggesting that future research can design targeted feature enhancement strategies for highly confusing and difficult-to-recognize gestures. Meanwhile, the learned channel weights further imply that the core activated muscle groups of short-stump amputees have undergone fundamental changes. Models trained solely on data from other subgroups exhibit limited generalization when applied to the short-stump population. Therefore, future studies should interpret the differences in muscle activation patterns of forearm amputees with short residual limbs from a neuromuscular perspective, thereby embedding neuroscientific prior knowledge to guide model learning. In addition, the model shows restricted performance in recognizing fine motor movements for the short-stump group. Capturing subtle variations in neuromuscular activation during fine motor tasks via physiological analysis will also be an important research direction for subsequent work.

Comparative experiments with existing studies further highlight the superiority and practicality of the proposed MSCB-DualAttn model. On the DB8 dataset, our method achieves a recognition accuracy of 95.81%. Although slightly lower than the method proposed by Shafiq et al. [36], which relies on sophisticated time-frequency handcrafted feature extraction, our approach eliminates the need for manually designed features. It learns discriminative representations directly from raw sEMG signals in an end-to-end manner, reducing dependence on manual feature engineering and offering greater practical applicability. Compared with the work of Li et al. [37], this study adopts the complete DB8 dataset consisting of both healthy participants and amputees, which presents a more challenging recognition task. Even so, our method remarkably outperforms the SVM, CNN, and LSTM models reported in their work, and is only marginally inferior to their MFFCNN-LSTM. Notably, the MFFCNN-LSTM was evaluated merely on a small subset of five subjects, limiting its generalizability. On the DB2 dataset, the proposed method reaches an accuracy of 86.42%, outperforming CNN-MSTINet from Ma et al. [38], LST-EMG-Net from Zhang et al. [39], and MCMP-Net from Mian et al. [40]. While its accuracy is somewhat lower than that of STMS-Net proposed by Wang et al. [41], the performance remains within a comparable range. The structurally similar MSMS-CLSTM2 [24] adopts a dual-branch CNN architecture with large-scale and small-scale streams. Each branch integrates a multi-scale module and a residual attention module (ResCBAM), followed by a Bi-LSTM layer to capture temporal dependencies. Unlike the hybrid attention design of conventional CBAM adopted by MSMS-CLSTM2, the proposed model decouples attention into a serial arrangement of independent channel attention and temporal attention modules. Meanwhile, our network architecture is more lightweight and concise without redundant structures. Under the identical partition scheme, the recognition accuracy of our model on the DB2 dataset is 1.83 percentage points higher than that of MSMS-CLSTM2. In addition, the MSCNN transfer learning model [42] achieves an F1-score of only 77.0 on the DB2 dataset, while the proposed method obtains a much better F1-score performance, further demonstrating its superior comprehensive classification capability. Moreover, our model exhibits superior adaptability to amputee participants with different residual limb lengths, further verifying its strong generalization capability and practical application value.

Based on the above analysis, we argue that the limitation of the MSCB-DualAttn model in recognizing short-stump gestures is not merely a technical issue, but rather reflects a fundamental challenge faced by existing deep learning paradigms when confronted with the structural reorganization of the neuromuscular control system. Motor control in short-stump patients has shifted from fine-grained synergy to coarse compensatory strategies, representing a qualitative change in the generation mechanism of EMG signals. The current model design is still predicated on the implicit assumption that the signal structure remains largely intact and only requires adaptation to parameter differences. When this assumption no longer holds, the model’s performance becomes constrained by the inherent priors embedded in its architecture. This finding suggests that future research should not remain confined to improving model structures but should delve deeper into how the neuromuscular compensatory patterns revealed by muscle synergy analysis can be translated into prior knowledge for the model, enabling it to dynamically select different feature extraction pathways based on varying residual limb lengths.

Limitations of the Study and Future Work

This study still has several limitations. A total of 20 transradial amputees were enrolled in our dataset, which is already a relatively large sample size compared with most existing related studies. However, these amputee participants were further divided into subgroups according to residual limb length, resulting in a limited number of samples within each subgroup. Restricted by practical conditions, it is difficult to recruit additional matched amputee subjects. Nevertheless, we will continue to enroll more participants in future work to expand and enrich the dataset.

Although the proposed model achieves substantial performance improvements over baseline methods, its recognition accuracy for the short residual limb group is still insufficient for practical clinical application. We observe large individual differences in recognition performance among short residual limb subjects, with several participants showing rather low accuracy. This phenomenon may be attributed to differences in amputation duration and other clinical factors.

In addition, the network architecture of this study was designed to balance recognition performance and model lightweightness. Accordingly, there remains a performance gap compared with state-of-the-art complex models. In follow-up research, we will further explore effective strategies to boost model accuracy while maintaining a lightweight network structure. Furthermore, we will conduct neuromuscular investigations on muscle activation and compensatory patterns of transradial amputees with different residual limb lengths, further analyzing the muscle activation characteristics of individuals with short residual limbs. We will also explore approaches to more directly embed muscle synergy prior knowledge into network architectures. Meanwhile, transfer learning, meta-learning, and generative data augmentation techniques will be integrated to further enhance the model’s generalization ability across transradial amputees with varying residual limb lengths.

5. Conclusions

In summary, the MSCB-DualAttn network achieves excellent recognition performance in healthy individuals, long stump, and middle stump groups through the synergistic design of multi-scale parallel convolution and dual attention mechanisms, validating the effectiveness of combining physiology-driven multi-scale feature extraction with dual-channel attention selection.

However, the performance degradation observed in the short-stump group reveals an inherent bottleneck of existing paradigms when confronted with structural reorganization of the neuromuscular control system: when the generation mechanism of the signal undergoes a qualitative change, the synergy structure disintegrates, channel functional differentiation disappears, and any model predicated on structural preservation will face challenges. We argue that future research should pursue a deeper integration of neuromuscular control theory with model design, enabling the model not only to learn features but also to recognize the type of control strategy, thereby dynamically switching between feature extraction and decision-making mechanisms under different compensatory patterns, ultimately achieving universal adaptability across populations with varying residual limb lengths.