Temporal convolutional network architectures: a novel simultaneous spatio-temporal model for comparative analysis

Conventional temporal-based deep learning models often fail to extract inter- channel information from electromyographic (EMG) signals. Existing spatio-temporal approaches typically sequentially combine spatial and temporal networks, but this strategy increases model complexity and parameter count. We introduce a simultaneous spatio-temporal convolutional deep network, which integrates spatial and temporal feature extraction connections within a single, explainable deep network. To evaluate the new architecture through a comprehensive comparative analysis, we compared its performance and model size with three other established decoding methods. We used two internal and two publicly available EMG databases. We report that the application of convolutional filters in both spatial and temporal directions simultaneously enhances myoelectric decoding accuracy. Finally, we explain the proposed model using the saliency maps method. The findings indicate that the proposed simultaneous spatio-temporal configuration offers reliable classification performance and is well-suited for real-time on-board deployment. The proposed model explains how simultaneous spatio-temporal convolution enhances the contribution of both temporal and spatial components of EMG activity, resulting in improved classification performance.

Temporal convolutional networks (TCNs) are a subclass of convolutional neural networks that utilize multiple layers of one-dimensional dilated causal convolutions, making them well suited for time-series processing and sequence modeling tasks (Bai et al., 2018; Cote-Allard et al., 2021; Van Den Oord et al., 2016a; Van Den Oord et al., 2016b). Using parallel processing, TCNs can capture long-range temporal information from past sequences with minimal computational costs (Zhou et al., 2022; Lauritsen et al., 2020; Nakagome et al., 2020; Kastaniotis et al., 2020; Jarrett et al., 2019; Lu et al., 2024; Musallam et al., 2021; Wang et al., 2022; Fernando et al., 2022; Khalili and Asl, 2021; Zhang et al., 2023). With advantages such as stability, speed, flexible receptive fields, and reduced memory demands compared to many state-of-the-art deep learning models, TCNs have emerged as strong candidates for myoelectric control applications (Cote-Allard et al., 2021; Li et al., 2021; Tsinganos et al., 2019; Betthauser et al., 2019b; Xie et al., 2024; Lin et al., 2024; Zhang et al., 2024; Betthauser et al., 2019a).

Despite this growing popularity, TCNs may not capture spatially distributed information. Any inter-channel information is only incorporated at the final classification stage, where features from different channels are concatenated into a single vector. Figure 1A provides a schematic representation of a vanilla temporal convolutional network, where the output of each 1-dimensional convolution layer is generated by applying convolutional operations over time points t to t − n independently across all channels. Enhancing TCNs to account for both temporal and spatial components can improve classification performance and reliability (Betthauser et al., 2019b; Jabbari and Nazarpour, 2024; López et al., 2024; Huang and Chen, 2019; Karnam et al., 2022; Li and Langari, 2022).

A naive approach to capture spatial information is to combine inputs from all channels before passing them into a 1-dimensional convolutional layer, as illustrated in Figure 1B. Here, convolutional connections link inputs across all channels directly to the output layer nodes. Because spatial connections are applied solely at the input level, we refer to this architecture as Spatial-in-Input TCN (Betthauser et al., 2019b). A common approach for capturing both spatial and temporal aspects of EMG signals involves first applying spatial convolution through standard one- or two-dimensional convolutional layers, then feeding the spatial block’s output into a temporal block, e.g., a long short-term memory network (Jabbari and Nazarpour, 2024; López et al., 2024; Huang and Chen, 2019; Karnam et al., 2022; Li and Langari, 2022). Figure 1C depicts this configuration. We refer to this architecture as the cascaded spatio-temporal TCN, or Cascade for brevity.

We propose that applying spatial and temporal convolutions simultaneously enhances decoding performance in an explainable manner and with a reduced number of parameters, that is, reduced model size. Figure 1D shows the proposed configuration, in which the output of each layer at time t is influenced not only by connections between features from channels m and q (spatial connections) at that time, but also by connections within each channel from time t and t − n, (temporal connections). We refer to this architecture as the simultaneous spatio-temporal configuration.

We report results across five datasets: a low-density public dataset (Ninapro Database 5), a medium-density dataset recorded in-house with a 21-channel system, a simulated medium-density dataset (Hyper, 32 channels, public) and a high-density EMG setup (Hyper, 64 channels, public). In total, data from 68 participants were analyzed.

Furthermore, another contribution of this study is the comparison of the proposed simultaneous spatio-temporal structure with other configurations under equal model size conditions. This analysis provides deep insight into designing a model capable of real-time on-board implementation.

The present study extends our previous work (Aghchehli et al., 2025), which introduced the foundational framework and reported preliminary results on one dataset. Here, we enhance the analysis by evaluating the model on four additional datasets and performing a comprehensive model-size comparison. We assume that a model capturing simultaneous spatio-temporal structure can provide reliable results within a model size range compatible with real-time on-board applications.

1-Dimensional dilated convolution

A one-dimensional convolutional layer has two key features: 1) dilation and 2) causal connection (Van Den Oord et al., 2016b). For an input sequence x ∈ R n and a kernel of size k denoted as f: 0,1, …, k − 1 → R, the causal dilated convolution on the element s of the input is: where X, d, and s − d ⋅ i represent the input sequence, dilation factor, and direction of the past, respectively. The dilation factor represents a fixed step between two adjacent kernel taps. Selecting a larger dilation leads to an expanded receptive field and includes data from the distant past.

Vanilla temporal convolutional network

Figure 1A shows the structure of the vanilla temporal convolution network concept. In this approach, there is no combination of features in the spatial direction at either the input level or the higher layers. The input space can be defined as x ∈ R c h × n f, where c h is the number of channels and n f the number of extracted features per channel. In this scenario, the dimension of the output of the convolutional layer will be y 1 (s), y 2 (s), …, y c h (s) with the direction of the kernel aligned on the time axis. A schematic block diagram of the vanilla structure used in this study is shown in Figure A1. After feature extraction, five dilated temporal layers were used to extract temporal components, followed by batch normalization and dropout layers. A max-pooling layer was used to reduce the risk of overfitting, and the classification decision was made by a softmax.

Spatial-in-input TCN

A simple way to combine spatio-temporal information is to place features extracted from all channels into a single 1-dimensional array at the input level before feeding them into a temporal block. As shown in Figure 1B, in this configuration the input space is considered as x ∈ R i n. The dimension i n is defined as f 1 c h 1, f 2 c h 1, …, f z c h q where f z c h q is defined as the feature z extracted from channel q. The output is calculated by convolutional connections from all features and all channels at times t and t − n. We fed the input to a 5-layer dilated temporal convolutional block, followed by batch normalization and dropout layers and a softmax, to build the desired spatial-in-input TCN structure.

Cascade spatio-temporal convolutional network

A conventional approach to combining spatial and temporal components is to first perform spatial convolution and then feed the output into the temporal block. Previously, it has been shown that combining convolutional neural network (CNN) models with Long Short-Term Memory (LSTM) neural networks achieves promising results in EMG-based hand movement classification. Inspired by these types of combined architectures, we used a dilated TCN block instead of LSTM and implemented a cascade framework of spatial-temporal convolutional networks. In this approach, as concept shown in Figure 1C, the spatial convolution outputs are obtained by spatial connections between features in the channels m and q at times t − n and t, independently. After extracting spatial features, the output of each time step is fed to a vanilla TCN. The block diagram of the proposed cascade configuration is depicted in Figure A1.

Simultaneous spatio-temporal convolutional network

Our objective was to develop a structure that applies spatial and temporal convolutions in a parallel and simultaneous structure. Figure 1D represents this configuration, in which both temporal and spatial outputs are achieved through simultaneous spatial-temporal connections. In this approach, the output of temporal and spatial convolutions is calculated simultaneously. To summarize briefly and as an example, there are two temporal and two spatial output nodes. The first temporal output is obtained by connecting the features of channel m at time t and t − n, while the second temporal output is achieved by applying convolution between the features of channel q at time t and t − n. In calculating the temporal output, there are no connections within the channel. For the first and second spatial outputs, the convolution is obtained by convolving features from channels m and q at time t and t − n, respectively. The difference is that convolutional connections are realized simultaneously, not in a spatial-in-input or cascade structure. For example, the feature set in channel q at time t is not only connected to the feature set of the same channel at time t − n, but is also connected to the feature set from channel m at time t.

As in previous models, each one-dimensional convolution layer was followed by batch normalization and dropout layers to reduce the risk of overfitting. Classification was achieved with softmax. Finally, it is worth mentioning that in all implemented structures, spatial, temporal, and simultaneous spatio-temporal convolutions operate on the feature topology of the electrodes, extracted from the raw voltage topology, rather than on the raw signals.

Model size equalization

To ensure a fair comparison between models, we considered the number of filters and the use of a max-pooling layer before the softmax layer. Furthermore, to investigate the performance of the proposed spatio-temporal configuration with varying the number of parameters, we aimed to compare it with the second-best model from the comprehensive analysis of four models on five datasets, which is the spatial-in-input structure. By adjusting the number of filters and, if necessary, pool size, we sought to control the number of parameters in the utilized models and make them equal. For the in-house low-density and medium-density data, as the best performance was achieved with a lower number of parameters compared to other datasets, it was straightforward to make the models equal in terms of the number of parameters by varying only the number of filters for each convolutional layer. For instance, for in-house low-density data, to create spatio-temporal and spatial-in-input models with a total number of parameters of about 3.9 K, the number of filters is set to 3 and 8, respectively. The temporal kernel size for both models is set to 5, and the spatial kernel size for the spatio-temporal model is set to 4. For the DB5 and Hyser medium and high-density data, as the best performance was achieved with a higher number of parameters, it was not possible to equalize the number of parameters by varying only the number of filters. Therefore, in addition to adjusting the number of filters, we deployed a max-pooling layer for the spatial-in-input model to make its number of parameters equal to that of the spatio-temporal model. For example, for DB5 data, to develop spatio-temporal and spatial-in-input models with a total number of parameters of about 30 K, the number of filters is set to 3 and 8, respectively, and the pool size for the spatial-in-input model was set to 6. It is worth mentioning that when we refer to the number of parameters, it is an approximate value rather than an exact count. For example, the number of parameters in the 30 K range in DB5 for spatio-temporal and spatial-in-input configurations is 30,647 and 30,853, respectively. It is worth mentioning that across all datasets, the number of filters per convolutional layer was adjusted per dataset with the goal of approximately equalizing the total number of trainable parameters between the spatio-temporal and spatial-in-input models.

Explaining the results using saliency maps

To provide a visualized, explainable way of showing how the proposed spatio-temporal model contributes more spatio-temporal components from the input feature, we implemented the saliency maps method. The primary motivation to develop this method was to highlight the most important pixels of an image input in computer vision when the model is classifying images (Meng et al., 2023). In this algorithm, to visualize the amount of contribution of an input pixel to the model’s final output, an importance score is assigned to it. In addition to image classification, saliency maps can be implemented for time series classification (Crabbé and Van Der Schaar, 2021; Ismail et al., 2020; Yan et al., 2021). In our scenario, features from each EMG channel at every time are considered as a spatio-temporal pixel. We deployed the saliency map method using the keras.vis library as described in Simonyan (2013). To calculate the contribution of each channel-time component when a specific input I 0 is used to decode class c, a class score function S c (I 0) is defined. This contribution is based on its influence on the score S c (I 0). Suppose a linear score for the class c, S c (I), as: where i, w, and b c are the vectorized version of I, the weight vector and the bias of the model. The magnitude of elements of w describes the importance of the corresponding spatio-temporal component of input I for the class c. Considering a temporal-channel as input for the decoding model, the class score S c (I) is a non-linear function of I. However, this non-linearity can be approximated by a first-order Taylor expansion as

To achieve the score of each spatio-temporal component from a given input, the gradient of the classification score S c with respect to the input I at the spatio-temporal component of I 0 is calculated.

EMG data

We employed four EMG datasets, including the publicly available DB5 dataset (Pizzolato et al., 2017) from the Ninapro website (http://ninaweb.hevs.ch), the Hyser dataset (Jiang et al., 2021), and two datasets previously recorded in-house datasets (Jabbari and Nazarpour, 2024; Aghchehli et al., 2024). In Table 1, we provide detailed information about these datasets. For the Hyser dataset, we considered 32 and 64 channels as medium-density and high-density, respectively.

EMG features

In this study to train the models, we implemented the six features, referred to as spatio-temporal-based feature sets (Samuel et al., 2019). In our previous study we showed that on-board implementation of these features, which firstly have been proposed in Samuel et al. (2019) as input for a one-dimensional spatial CNN model can obtain an accuracy about 94 % for a real-time classification task (Jabbari et al., 2024). Furthermore, it has been shown that the average time required to extract this feature set on a randomly generated dataset is approximately 0.6691 ms, which represents a moderate computational cost and offers the possibility of real-time implementation (Khushaba and Nazarpour, 2021). Therefore an important motivation of using these feature sets was their compatibility for on-board implementation. In addition to their low computational cost, making them suitable candidates for real-time purposes, they have also shown acceptable off-line performance in previous studies (Samuel et al., 2019; Khushaba and Nazarpour, 2021; Jabbari et al., 2021). We believe that utilizing six STFS feature sets for future on-board implementation of the proposed simultaneous spatio-temporal structure will be critical. The first feature is the integral square descriptor of the signal. This feature is calculated by summing the squared magnitudes of the signal values over a given window and represents the power of the signal. The second and third features were derived from the difference between the integral square descriptor and the normalized root-square coefficients of the first and second differential derivatives. The fourth feature estimates muscle contraction force for a specific gesture. This is achieved using a variant of the non-linearized form of the log-detector presented in Tkach et al. (2010). For the fifth feature, the mean value of the square root of the given window as a temporal representation of the muscle activity is calculated (Samuel et al., 2018). Finally, mean derivatives of the higher order moments were adopted from Al-Timemy et al. (2015). Table 2 presents the equations used to extract these features. In the table, x represents a window of the signal, n denotes the window length, and D D x 1 and D D x 2 correspond to the first and second differential derivatives of the signal, respectively. Furthermore, R S C x represents root-square coefficient of the signal x. We used an overlapping windowing scheme to extract the mentioned features for all the utilized datasets. For the in-house low-density, in-house medium-density, DB5, and Hyser datasets, we used 200 ms, 281.16 ms, 150 ms, and 97.65 ms windows with step sizes of 20 ms, 9.37 ms, 50 ms, and 24.41 ms, respectively, to extract features. We selected different window lengths for different datasets was to allow for a broader comparison and to evaluate the models across a range of window lengths. Since all the window lengths are within an acceptable range for real-time purposes (300 ms), these variations can be helpful for better evaluating the models under different window length conditions, as the window length is an important parameter in EMG-based hand gesture classification frameworks (Khushaba and Nazarpour, 2021).

Train-test split approach

For the in-house low-density and in-house medium-density datasets, we used the first 50 % of the data to train the models and the second half to test them. In DB5, repetitions 1, 2, and 3 chosen for training, while repetitions 4, 5, and 6 were used to test the models. For this dataset, we used all gestures from all three exercises. For the Hyser data, we utilized the first half of each session (sessions 1 and 2 and dynamic task type) for training and the second half for testing the models. In this study, a subject-dependent training approach was considered to train and evaluate the models for each dataset. The model trained from the data from one participant is evaluated on the data from the same participant.

Statistical analysis

The performance of the methods in gesture classification was evaluated using classification accuracy. The Wilcoxon signed-rank test was applied to assess the statistical significance of the results, with a p -value less than 0.05 and 0.01 considered statistically significant, shown with ∗ and ∗ ∗, respectively.

Training configuration

We used the Adam optimizer with a learning rate of 0.001 to train all models. All the networks were trained for up to 300 epochs using an early stopping strategy by monitoring accuracy, with a batch size of 32. All preprocessing steps, including filtering and segmentation, were conducted in Python using the NumPy and SciPy libraries. We utilized TensorFlow to implement one-dimensional CNN-based classification models, and all analyses were performed using an NVIDIA GeForce RTX 3090 GPU.

Comparison with conventional TCN-based models

As one of the main contributions of this study, we evaluated the proposed spatio-temporal model alongside three conventional TCN-based models, including vanilla, spatial-in-input, and Cascade configurations. To ensure a comprehensive comparison, we utilized low, mid, and high-density EMG datasets to cover a wide range of scenarios.

Figure 2 illustrates the average classification accuracy values of all conventional TCN-based models, as well as the proposed simultaneous spatio-temporal method, for in-house low-density, DB5, in-house medium-density, Hyser medium-density and Hyser high-density datasets in A, B, C, D and E, respectively. For all datasets, the proposed spatio-temporal model significantly outperformed vanilla, spatial-in-input, and cascade configurations. For the DB5 dataset, where the number of gestures is significantly higher than the in-house data, increasing to 53, the proposed spatio-temporal method achieved an accuracy of 84.86 % significantly outperforming the spatial-in-input, which achieved 80.90 % as the second best method (p < 0.01 for all comparisons). For the in-house low density data, the simultaneous spatio-temporal model achieved an average classification accuracy of 97.02 %, demonstrating highly accurate performance for EMG-based hand gesture classification. This outperformed the second best configuration, the cascade configuration, which achieved an accuracy of 95.45 %. The same scenario, showing the superiority of the proposed spatio-temporal, was achieved for the Hyser dataset. The results for the Hyser medium-density dataset show that the spatio-temporal model achieves the highest average classification accuracy rate of 86.28 ± 7.36 %, with a significant improvement compared to vanilla TCN, spatial-in-input and cascade structures (p < 0.01 for all comparisons). It is interesting to note that increasing the number of electrodes from 32 to 64 for the same dataset results in a more accurate classification accuracy rate rate for the spatio-temporal model, with an average of 93.70 ± 4.83 % with p < 0.01 for all comparisons.

Performance of spatio-temporal configuration with varying sizes

A crucial aspect in developing a model for EMG-based hand gesture classification is considering its size and computational cost, as it is one of the main challenges in real-time implementation of deep learning on board. The size of the model depends directly on the number of parameters in the model, which results from factors such as the number of filters, kernel size, dilation factor, etc. Most deep learning-based models used in EMG classification tasks suffer from a large number of parameters, which poses a significant challenge for implementation on embedded systems. To evaluate the proposed spatio-temporal model with different number of parameter, we conducted a comprehensive analysis on the utilized five datasets: in-house low density, DB5, in-house medium-density, Hyser medium-density, and Hyser high density. We selected three ranges for the number of parameters for each dataset, and since the previous analysis showed that spatial-in-input was the second-best model, we included only this configuration for comparison with the proposed spatio-temporal structure. The average classification accuracy rates for spatio-temporal and spatial-in-input models for five datasets are illustrated in Figure 3. For the in-house low-density data, across all parameter ranges (1.3 K, 2.6 K, and 3.9 K), the proposed spatio-temporal architecture demonstrated superior performance, achieving average classification accuracy rates of 97.00 %, 97.02 %, and 96.81 %, respectively, while spatial-in-input achieved 94.55 %, 95.45 %, and 94.26 %. It is worth mentioning that with decreasing the number of parameters from 3.9 K to 1.3 K, statistical test proves that the proposed spatio-temporal model can significantly outperform the spatial-in-input configuration (p < 0.01). The same pattern of results is observed for DB5, proving that the spatio-temporal model is more accurate than the spatial-in-input structure, achieving 83.82 %, 84.86 %, and 84.29 % compared to 77.83 %, 80.90 %, and 81.24 % for the three parameter ranges, respectively (p < 0.01 for all comparisons). For the in-house medium-density data, when the number of parameters is approximately 3.8 K, the spatio-temporal model significantly outperforms the spatial-in-input configuration, achieving 92.92 % compared to 82.65 % (p < 0.01). However, with increasing number of parameters to about 22 K, the difference between the accuracy of classification deploying spatio-temporal and spatial-in-input models is less than one percent, still with better accuracy for spatio-temporal configuration. For the Hyser medium-density data, across all ranges of number of parameters, including 31 K, 70 K, and 122 K, the spatio-temporal structure outperformed the spatial-in-input configuration. However, with increasing number of parameters, the difference between the two models decreased (p < 0.05). Finally, for the Hyser high-density data, despite a significant difference between spatio-temporal and spatial-in-input configurations with 17 K parameters (p < 0.01), as the number of parameters is increased to 136 K, the difference is about 0.98 % with p > 0.05, with still better performance for the spatio-temporal model.

Explainability

One of the primary aims of this study was to explain why the proposed spatio-temporal model exhibits significantly better performance compared to conventional configurations. The main motivation behind developing our method was to capture both spatial and temporal patterns from the input. Consequently, significant activation of the time-channel pixels is expected to be observed in the explainable maps. In the following, we demonstrate the use of saliency maps as a tool to visualize which regions of the input contribute more in both spatio-temporal and vanilla TCN. We evaluated the models on two DB5, and Hyser datasets.

In Figures 4A,B (left), we present the heat map resulted from the saliency map method for DB5 and Hyser high-density datasets, respectively. It is clearly evident that the main motivation for proposing a spatio-temporal model is realized by the activation of most temporal-spatial (time-channel) components through the input of the model when using proposed spatio-temporal structure, compared to vanilla TCN. It shows that for both medium- and high-density datasets, the proposed model successfully increases the contribution of all channels simultaneously at most times, effectively addressing the spatio-temporal aspects of the EMG signals.

In the DB5 dataset shown in Figure 4A (left), two Myo armbands were located around the forearm to record sEMG signals. Channels 0-7 are were placed at the height of the radio humeral joint, and channels 8-15 represent the second Myo, tilted by 22.5° clockwise. Therefore, each channel captures a mixture of activity from multiple underlying muscles. For the spatio-temporal model, both proximal (channel indexes 0–7) and distal (channel indexes 8–15) channels contribute, with slightly stronger emphasis in distal channels near the wrist at the most recent time steps. This indicates the proposed spatio-temporal model improves the contribution of the coordinated forearm muscle group activity, rather than a single localized source; consistent with multi-muscle involvement in hand gestures. The Hyser dataset provides four spatially organized 8 * 8 HD-sEMG arrays covering the forearm flexor and extensor muscle groups. In this study, we uniformly subsampled the 256 channels by selecting every fourth electrode, resulting in 64 channels distributed across all arrays. This selection preserves global forearm in both extensor and flexor muscles. Therefore, saliency analysis is interpreted at the level of distributed forearm regions covered by 64 channels. As illustrated in Figure 4B (left), channels 0–15, 16–31, 32–47, and 48–63 correspond to distal extensor, proximal extensor, distal flexor, and proximal flexor muscle groups, respectively. Since channels are spatially distributed across the forearm, the model is exploiting coordinated activation of multiple forearm regions, and this reflects synergistic recruitment of flexor and extensor muscle groups rather than reliance on a single localized source. However, the vanilla structure fails to capture meaningful spatio-temporal dependencies.

Furthermore, we calculated the average gradient of the output with respect to each input channel across all subjects, as illustrated in Figures 4A,B (right) for the in-house medium-density and Hyser high-density datasets, respectively. The average values calculated for each channel can be considered as an indication of the importance of the electrode. Therefore, the comparison between spatio-temporal and vanilla TCN shows that the importance representation of each channel is significantly higher in the proposed model, providing clear evidence of the model’s capability to highlight more spatially distributed temporal information.

This study provides a comprehensive overview of using one-dimensional convolutional based models in EMG-based pattern recognition framework for myoelectric applications, including the development of a novel model, a thorough performance evaluation, the feasibility of real-time implementation in terms of model size, and the potential of the proposed model. Key contributions of this study include: proposing a novel one-dimensional based convolutional spatio-temporal structure capable of applying simultaneous convolutions in both time and channel directions, providing a comprehensive comparison between three conventional TCN-based configurations and the proposed method, emphasizing the explainability of the TCN model, and conducting a parameter-dependent evaluation of the method for future real-time implementation studies.

In EMG-based hand movement decoding, addressing both spatial and temporal components can lead to more reliable and generalized intuitive control of a prosthetic hand (López et al., 2024; Karnam et al., 2022; Li and Langari, 2022; Li and Langari, 2023; Bao et al., 2020; Chen et al., 2020). Although the combination of CNN and LSTM has been shown to effectively capture underlying relationships, the main challenge with these methods is the large number of parameters they require. This challenge complicates real-time implementation. However, leveraging TCN’s ability to extract informative components through dilated, larger receptive fields, the proposed spatio-temporal model demonstrates that even with a small number of parameters (1 K to 2 K), as discussed in Results, Section B, reliable results can still be achieved.

Recently, TCN-based models have been widely taken into account for processing EMG pattern recognition in a temporal-sequential way. However, most of these models are purely TCN, processing data only in the temporal direction, and some do not provide sufficient details about the input shape, leaving it unclear whether they are pure TCN or spatial-in-input TCN. In this work, we demonstrated that there is a significant difference in performance between these configurations. This clearly highlights the importance of incorporating spatial information, even in the spatial-in-input approach, which significantly outperformed the vanilla TCN. In the comparison between spatial-in-input and cascade configurations, it can be observed that when the number of channels is smaller (as in the in-house low-density and DB5 datasets), or the number of classes is low (as in the in-house medium-density dataset), there is no significant difference between configurations (refer to Figure 2, p > 0.05). However, when the number of electrodes significantly increases (as in the Hyser medium and high-density datasets), spatial-in-input demonstrates better accuracy. Therefore, more evaluation should be done on these two types of configurations.

An interesting finding is that with decreasing the number of parameters, for all datasets, the proposed spatio-temporal structure can significantly outperform the spatial-in-input configuration with p < 0.01, while increasing the number of parameters increases the p -value for the in-house low-density, in-house medium-density, and Hyser high-density datasets (p > 0.05). This shows that the proposed spatio-temporal method will be appropriate for real-time on-board implementations, as they need a lower number of parameters to be implemented (Jabbari et al., 2024).

Beyond off-line analysis, TCN-based methods have shown promising results in real-time experiments in various biological and health-related tasks. In particular, they boast significantly lower computational costs and memory requirements compared to other leading deep learning algorithms. Their straightforward structure also simplifies the training process. These advantages are crucial for enabling real-time control of sEMG-based prosthetic hands and make TCNs well-suited for embedded systems at the edge. In addition, TCNs have shown superiority over other recurrent neural networks, including spatio-temporal, while maintaining favorable computational efficiency. Using these properties, a real-time embedded system was implemented using a multi-core IoT processor, specifically the GAP8 system proposed in Zanghieri et al. (2019). Inspired by one-dimensional spatial CNN approach for EMG pattern recognition a spatial CNN based model implemented on Arduino Nano 33 BLE with TinyML framework, weighing only 5 g, is proposed in Jabbari et al. (2024). The model utilized in this study contains about 7 K parameters. Our parameter-dependent results prove that with even fewer parameters than 7 K, high accuracy performance, more than 97 % can be achieved. Furthermore, in terms of compatibility with hardware and frameworks, all layers and operations implemented can be deployed in TinyML to enable real-time tasks. Therefore, future work could focus on deploying a real-time, on-board implementation of the proposed spatio-temporal model to perform EMG-based hand gesture classification tasks.

An explainable machine learning approach for EMG-based movement recognition is crucial, as it helps uncover patterns and behaviors in the data that generalize well to unseen cases, ultimately contributing to the development of more robust models (Gehlot et al., 2024; Rani et al., 2024; Gozzi et al., 2022; Lee et al., 2022). In the case of spatio-temporal models, it is important to determine whether performance improvements are driven by the data itself or are the result of specific patterns that the model can learn and extract. Saliency map analysis clearly demonstrated that the developed spatio-temporal model effectively enhances the contribution of spatially distributed information over time, in contrast to the vanilla TCN, which was expected to emphasize fewer spatial components. Since all parameters and the amount of data in each training batch were identical for both vanilla and proposed spatio-temporal configurations, the saliency maps indicate that the accuracy improvement with the proposed structure is due to the patterns captured by the model itself, rather than the quantity of data.

Despite the increasing trend of using deep learning models in hand gesture classification, most of them focus solely on accuracy and rely on a black-box approach. The proposed methods, including TCN-based models, care only about the performance by only looking at the output of the model as a result. Furthermore, it is conventional to look at the input type of models in a feature extraction framework to develop optimal methods to have a trade-off between computational cost and accuracy. This is a blind-to-explainability way of performing EMG-based hand gesture classification. However, explaining and exposing the underlying relationships within the internal layers and connections of the model reveal transparency and confidence in decision making during the classification process and provide the opportunity to refine the internal structure of the model rather than implementing hand-crafted features to achieve more accurate performance (Jiang et al., 2023). Explainable deep learning results in a more reliable model and pattern recognition process, which can bridge the gap between research and commercial applications by improving generalization.

Although this study focuses on one-dimensional TCN-based models, which can be considered neural-inspired architectures for EMG decoding, we acknowledge that biologically plausible spiking and pulse-coupled models have demonstrated promising results for EMG signal classification (Sava et al., 2023; Baracat et al., 2025a, Baracat et al., 2025b; Donati et al., 2019). Spiking Neural Networks (SNNs), often referred to as the third generation of neural networks, employ discrete spike events to model temporal information through the precise timing of spikes (Yamazaki et al., 2022). SNNs can also be utilized to capture spatio-temporal components of the biological signals with sparse and event-driven computation, which results in a low computational cost approach (Xu et al., 2024; Yu et al., 2022; Kugele et al., 2020). Possessing biological plausibility and great potential to process temporal information with lower computational costs, SNNs have shown promising performance in EMG-based hand gesture classification (Yang et al., 2023; Ke et al., 2020; Xu et al., 2023; Xing et al., 2020). Furthermore, Pulse-Coupled Neural Networks (PCNNs), as a specific type of SNN models and equipped with a more complex biologically plausible neuron model, have been widely applied to image-like classification tasks (Mohammed et al., 2015; Liu et al., 2022; Johnson and Padgett, 1999). An important advantage of the PCNNs is their ability to apply local coupling with complex intra- and inter-channel mechanisms, resulting in more robust spatial representations. This capability makes PCNN models a suitable candidate for extracting spatio-temporal components from grid-based sensor arrays, including high-density EMG datasets and hand gesture classification. A compact (light) SNN structure is proposed in Montazerin et al. (2023a) to perform hand gesture classification from high density EMG recording by capturing spatio-temporal features of the signals. By narrowing the search interval and designing a pretrained subnet SNN with more hidden layers, a rapid SNN model proposed in Cheng et al. (2019) could achieve an accuracy of about 97.4 % for an EMG-based gesture classification. It has also been shown that a neuromorphic implementation of an SNN model can be used to extract spatio-temporal features from EMG, improving hand gesture classification performance with significantly low computational cost (Donati et al., 2019). Therefore, recent approaches proposing the conversion of artificial neural networks to SNNs, pseudo-gradient training methods, and neuromorphic hardware implementations demonstrate that spiking models can support low-power, on-chip real-time systems, making them potential candidates for future on-board pattern recognition–based hand gesture classification (Bu et al., 2025; Rathi et al., 2020; Pawlak and Howard, 2025; Renner et al., 2024; Syed et al., 2021).

It is also worth mentioning that attention-based and transformers architectures were not included in our comparison due to the computational cost. In this study, focusing on only one-dimensional based convolutional architectures, one of the main motivations is to consider the future possibility of real-time on-board implementation with the lowest latency and memory constraints. Although we acknowledge that transformer-based models have demonstrated excellent performance in EMG-based hand gesture classification, previously shown in Montazerin et al. (2022); Mahboob et al. (2023); Montazerin et al. (2023b); Lin et al. (2024); Dere and Lee (2024), they were not included in the comparison due to their generally higher computational cost. They suffer from quadratic attention cost (O (n 2)), imposing substantial computational burden (Wang et al., 2020; Yun et al., 2020). In other words, self-attention mechanisms scale quadratically with sequence length, which makes the training process slower.

When developing deep models, it is critical to consider the number of parameters, as most require a large number to work effectively, which presents a challenge in real-time EMG-based classification. Comparisons with previously developed methods should be made in a fair manner in terms of model size. This is an important challenge that previous studies have struggled to address, with most focusing only on highlighting the performance of the proposed model.

Finally, our results demonstrate that the proposed simultaneous spatio-temporal configuration achieves reliable classification accuracy within a model size range compatible with real-time on-board deployment (Jabbari et al., 2024).