Electrocardiogram Arrhythmia Classification with SE-ResNet: Effects of Signal Windowing and Hybrid Class Balancing

The overall framework of the proposed automated deep learning methodology was designed as a systematic, end-to-end pipeline to transform raw physiological signals into reliable classifications, as illustrated in Figure 1. This sequential process aimed to optimize both data quality and deep feature extraction capabilities.

The process initiated with the acquisition of the Raw ECG Dataset, specifically utilizing the widely benchmarked MIT-BIH Arrhythmia Database. Following acquisition, the continuous physiological recordings underwent Signal Segmentation. In this stage, the continuous time-series data was strategically sliced into static temporal windows (primarily the 360-sample window) to isolate individual cardiac cycles and fully capture the crucial P-QRS-T morphological timeframe [27, 28].

Once segmented, the data entered the Data Preprocessing phase. This stage involved standardizing the amplitude of the electrical signals and rigorously splitting the dataset into training, validation, and testing subsets to ensure unbiased model evaluation. Subsequently, to resolve the severe and inherent class imbalance present in real-world cardiology data, the Hybrid Data Balancing (Tomek + SMOTE) technique was strictly applied. This crucial step, which was applied exclusively to the training subset to prevent data leakage, cleaned overlapping inter-class boundaries and synthesized structurally representative minority class samples, establishing a perfectly balanced learning environment.

The pristine, balanced data was then fed into the core analytical engine: the proposed 1D-CNN architecture. Enhanced with SE attention mechanisms and Residual blocks, this deep learning model dynamically recalibrated its focus to extract the most discriminative spatial features from the ECG waveforms while mitigating vanishing gradients. The highly dimensional feature maps were then channeled into the Classification Layer, which utilized Global Average Pooling (GAP) and a Softmax activation function to compute categorical probabilities. Finally, the network generated the Output Prediction, classifying the input signal into one of the five standard Association for the Advancement of Medical Instrumentation (AAMI) arrhythmia classes (N, S, V, F, Q) with high diagnostic confidence [29, 30].

image003.png

Figure 1. The proposed methodology framework for automated electrocardiogram (ECG) arrhythmia classification

2.1 Dataset and windowing strategy

This study utilized the widely benchmarked MIT-BIH Arrhythmia Database [31, 32], originally obtained from the Beth Israel Hospital. The dataset is publicly accessible through the PhysioNet platform [33] (https://physionet.org/content/mitdb/1.0.0/) and is strictly distributed under the Open Data Commons Attribution License (ODC-By) v1.0. The database contains 48 half-hour excerpts of two-channel ambulatory ECG recordings from 47 subjects. To provide a comprehensive evaluation, the dataset includes 23 randomly chosen recordings from a mixed population and 25 intentionally selected recordings to capture less common but clinically significant arrhythmias. This deliberate inclusion of rare medical events inherently creates the severe class imbalance addressed in this study. The dataset was independently annotated by multiple cardiologists, yielding approximately 110,000 beat annotations, which were grouped into five standard AAMI classes: N, S, V, F, and Q.

image004.png

Figure 2. Standard electrocardiogram (ECG) waveform morphology and clinical intervals

To effectively classify these arrhythmias, it is essential to capture the underlying physiological signals accurately. As illustrated in Figure 2, a standard single cardiac cycle consists of several critical morphological landmarks, fundamentally comprising the P wave, the QRS complex, and the T wave. The P wave represents atrial depolarization, whereas the QRS complex strongly indicates ventricular depolarization, and the T wave corresponds to ventricular repolarization [34]. Furthermore, the clinical timing between these waves contains vital diagnostic information regarding the heart's electrical conduction system.

Crucially, the original MIT-BIH recordings were digitized at a sampling rate of 360 samples per second (360 Hz) with an 11-bit resolution over a 10 mV range. This technical specification strictly underpins the segmentation strategy implemented in this study. To analyze the effect of signal length representation on the model's performance, two static windowing scenarios were established:

•180-sample window (0.5 seconds of cardiac activity): Captures a narrower physiological timeframe. This length is frequently insufficient as it may truncate critical components, such as the distant T wave or the early P wave, thereby depriving the deep learning model of full contextual data.

•360-sample window (1.0 second of cardiac activity): Captures a broader timeframe. By aligning the window size perfectly with the dataset's original sampling rate, this length captures exactly one full second of the heart's electrical activity. This temporal span provides an adequate window to encompass the entirety of the P-QRS-T complex along with its surrounding baseline.

By utilizing the 360-sample window, the automated system ensures that the 1D-CNN has access to all necessary physiological features required for robust and discriminative feature extraction across all AAMI classes.

Before applying data balancing, the dataset was rigorously partitioned at the intra-patient (beat) level. To ensure class representativeness and computational reproducibility, a stratified splitting protocol was executed with a fixed random seed of 42. Specifically, 80% of the dataset (85,064 samples) was allocated for the primary training phase, while the remaining 20% (21,266 samples) was strictly isolated for the final testing evaluation. Furthermore, during the model training process, 20% of the primary training data was dynamically reserved as an internal validation set to monitor generalization and trigger the Early Stopping mechanism. It is critical to note that all data balancing operations (Tomek-Links, SMOTE, and ADASYN) were applied exclusively to the training subset to strictly prevent data leakage into the validation and testing environments.

2.2 Data balancing techniques and leakage prevention

To effectively address the severe class imbalance inherent in the MIT-BIH dataset—where the Normal (N) class heavily overshadows critical minority arrhythmias—a rigorous data balancing strategy was implemented. Crucially, to preserve the integrity of the evaluation metrics, all five balancing scenarios evaluated in this study were applied exclusively to the training dataset. This strict methodological isolation ensures that the validation and testing sets remain entirely representative of real-world, imbalanced clinical scenarios, thereby completely preventing the phenomenon of data leakage that often artificially inflates model performance [35].

The balancing scenarios explored in this study encompass three fundamental approaches: oversampling, undersampling, and hybrid combinations. For oversampling, the SMOTE and the ADASYN algorithm were utilized. Unlike naive random duplication, which leads to severe overfitting, SMOTE intelligently generates synthetic minority samples by interpolating between existing minority instances and their nearest neighbors. ADASYN further modifies this concept by adaptively shifting the importance of the class boundaries, generating more synthetic data for minority instances that are harder to learn.

However, relying solely on oversampling often exacerbates the overlapping of class boundaries in the feature space, introducing synthetic noise. To counteract this, an intelligent undersampling technique known as Tomek-Links was evaluated. Instead of randomly discarding majority instances, the Tomek-Links algorithm systematically identifies pairs of instances from opposite classes that are the closest neighbors to each other. By removing the instance that belongs to the majority class within these pairs, the algorithm effectively clears ambiguous borderline data, establishing a much sharper and more discernible decision boundary for the classifier.

Ultimately, to harness the strengths of both approaches while mitigating their individual weaknesses, hybrid methodologies—specifically Tomek-SMOTE and Tomek-ADASYN—were proposed as the primary solutions. In these hybrid scenarios, the synthetic generation algorithms (SMOTE or ADASYN) are first applied to populate the minority class space. Subsequently, the Tomek-Links algorithm is deployed as a post-generation data cleaning mechanism. This synergistic pipeline ensures that the 1D-CNN model receives a training dataset that is not only perfectly balanced in quantity but also structurally distinct and clearer from overlapping boundary noise.

2.3 Proposed One-Dimensional Convolutional Neural Network architecture

The core of the proposed classification model is a deep 1D-CNN meticulously designed for high-fidelity feature recalibration. Unlike standard CNNs, this architecture integrates SE blocks and residual connections to optimize pattern recognition in non-stationary ECG signals. The detailed flow of the tensor dimensions and the specific operations within the network are systematically illustrated in Figure 3.

As depicted in the schematic, the architecture processes the (360, 1) input tensor through three primary stages of spatial feature extraction:

•Initial and Intermediate Extractions (Stages 1 and 2): The network utilizes 32 and 64 filters with large kernel sizes to capture broad morphological patterns, reducing the spatial dimension down to (90, 64) through successive Convolutional, Batch Normalization, Rectified Linear Unit (ReLU) activation, and MaxPooling layers.

•Deep Feature Recalibration (Stage 3): This stage acts as the main analytical engine of the proposed method, utilizing 128 filters. It introduces a dual-mechanism approach to handle deep complexities. First, the SE block functions as an attention mechanism. It employs GAP followed by dense layers to explicitly model channel-wise interdependencies, prioritizing clinically significant ECG segments over background noise (denoted by the multiplication operation '×').

•Residual Skip Connection: Immediately following the attention mechanism, a Residual Block is employed. This mechanism provides a direct shortcut (skip connection) for the tensor to bypass intermediate convolutional layers. By fusing the original pre-block features with the processed features (denoted by the addition operation '+'), the network effectively mitigates the vanishing gradient problem, enabling the stable training of deep layers without losing the original ECG pattern context.

Finally, the recalibrated high-level features are flattened using a GAP layer and passed through a 128-unit dense layer with a 0.2 dropout rate to prevent overfitting. A final softmax classification layer maps the features into the five distinct AAMI arrhythmia classes.

To facilitate the full reproducibility of the proposed methodology, the comprehensive implementation details of the 1D-CNN (SE-ResNet) architecture are systematically summarized in Table 1. This table explicitly outlines the kernel sizes, strides, padding configurations, SE reduction ratios, and the progressive output tensor dimensions across each layer, based on the optimal 360-sample input window.

image005.png

Table 1. Comprehensive architectural details and hyperparameters of the proposed 1D-CNN (SE-ResNet) model

Note: 1D-CNN = One-Dimensional Convolutional Neural Networks, SE-ResNet = Squeeze-and-Excitation Residual Network, ReLU = Rectified Linear Unit.

2.4 Mathematical formulation of the hybrid balancing (Tomek-SMOTE)

The proposed hybrid balancing mechanism operates in two sequential mathematical phases: synthetic oversampling and boundary cleaning. First, the SMOTE generates synthetic instances for the minority classes by interpolating between existing minority samples. The new synthetic sample, $x_{syn}$, is mathematically formulated as shown in Eq. (1):

$x_{s y n}=x_i+\lambda \cdot\left(x_{k n n}-x_i\right)$   (1)

where,

$x_{syn}$ is the newly generated synthetic ECG sample.

$x_i$ represents the minority class instance under analysis.

$x_{k n n}$ denotes one of its k-nearest neighbor (k-NN) within the 360-dimensional feature space.

$\lambda$ is a random scalar uniformly distributed in the range [0,1] that determines the exact position of the synthetic sample along the line segment.

While SMOTE successfully expands the minority decision region, it often blindly synthesizes data into the overlapping zones of the majority class [36]. To resolve this, the Tomek-Links algorithm is applied as a rigorous undersampling filter to remove ambiguous borderline samples. Formally, the condition for a pair of instances to constitute a Tomek-Link is expressed:

$\nexists x_k: d\left(x_i, x_k\right)<d\left(x_i, x_j\right) \vee d\left(x_j, x_k\right)<d\left(x_i, x_j\right)$   (2)

where, $x_i$ and $x_j$ are a pair of instances belonging to different classes (one majority and one minority), $x_k$ represents any other arbitrary instance in the dataset, and $d(\cdot, \cdot)$ denotes the Euclidean distance between two instances in the 360-dimensional feature space.

If the condition in formula (2) is met, the pair $\left(x_i, x_j\right)$ is considered a Tomek-Link, indicating that they are each other's nearest neighbors despite belonging to opposite classes. Consequently, the instances are considered boundary noise, and the instance belonging to the majority class is systematically removed, thereby clarifying the spatial boundaries between the ECG classes.

2.5 Theoretical operations of the Squeeze-and-Excitation Residual Network architecture

The core learning capabilities of the proposed 1D-CNN rely on specific tensor operations. For a given 1D ECG sequence, the standard discrete convolution operation in the $l$-th layer to produce the output feature map $y$ at position $i$ is defined in Eq. (3):

$y_i=f\left(\sum_{j=0}^{k-1} W_j \cdot x_{i+j}+b\right)$    (3)

where, $y_i$ is the output feature value at position $i, x_{i+j}$ is the input ECG segment, $W_j$ represents the learnable kernel weights of size $k, b$ is the bias term, and $f(\cdot)$ denotes the non-linear ReLU activation function.

To mitigate the vanishing gradient problem inherent in deep networks, a Residual Block is integrated [37]. Instead of learning a direct unreferenced mapping, the network explicitly learns a residual function $\mathcal{F}(x)$. The output $H(x)$ of the residual block is defined by Eq. (4):

$H(x)=\mathcal{F}\left(x,\left\{W_i\right\}\right)+x$    (4)

where,

$H(x)$ is the final output tensor of the residual block.

$\mathcal{F}\left(x,\left\{W_i\right\}\right)$ represents the residual mapping to be learned by the convolutional layers parameterized by weights $W_i$.

$x$ is the original identity input tensor that bypasses these layers via the skip connection.

Furthermore, the architecture employs a SE block to perform dynamic channel-wise feature recalibration. The "Squeeze" operation compresses the spatial dimensions of the feature map $U$ (with length $L$) into a channel descriptor $z_c$ using GAP, as formulated in Eq. (5):

$z_c=\frac{1}{L} \sum_{i=1}^L u_c(i)$    (5)

where, $z_c$ is the squeezed scalar value for the $c$-th channel, $L$ is the spatial length of the input feature map, and $u_c(i)$ represents the $i$-th temporal element of the $c$-th channel in the original feature map $U$.

Subsequently, the "Excitation" operation models the channel-wise dependencies using a gating mechanism with two fully connected layers. The recalibration scalar $s$ is computed via Eq. (6):

$s=\sigma\left(W_2 \delta\left(W_1 z\right)\right)$   (6)

where, $s$ is the attention weight vector for all channels, $W_1$ and $W_2$ are the learnable weight matrices of the dimensionality-reduction and dimensionality-increasing dense layers respectively, $\delta$ denotes the ReLU activation function, and $\sigma$ represents the Sigmoid activation function.

Finally, the original feature map is scaled by these activations to produce the attention-weighted output $x_{\text {out }}$, as defined in Eq. (7):

$x_{\text {out }}=s_c \cdot u_c$   (7)

where, $x_{\text {out}}$ is the final recalibrated feature map, $s_c$ is the scalar attention weight assigned to the $c$-th channel (derived from vector $s$), and $u_c$ is the original pre-attention feature map of that specific channel.

This systematic mechanism ensures that the model dynamically enhances highly informative ECG features while actively suppressing irrelevant background noise through the operations defined in Eqs. (3)-(7).

2.6 Classification layer and loss function

Following the deep feature extraction process, the highly dimensional tensor is flattened and passed through a fully connected dense layer. To obtain the final categorical predictions for the five AAMI arrhythmia classes, a Softmax activation function is utilized at the output layer [38]. For a given input vector $z$, the Softmax function calculates the predicted probability $\widehat{y}_j$ for the $j$-th class as formulated in Eq. (8):

$\widehat{y_J}=\frac{e^{z_j}}{\sum_{k=1}^K e^{z_k}}$    (8)

where, $K=5$, representing the total number of target classes. During the training phase, the network optimizes its weights by minimizing the error between the predicted probabilities $\hat{y}$ and the true labels $y$. Due to the multi-class nature of this study, the Categorical Cross-Entropy loss function, denoted as $\mathcal{L}$, is employed, as defined in Eq. (9):

$\mathcal{L}=-\sum_{i=1}^N \sum_{k=1}^K y_{i, k} \log \left(\widehat{y_{l, k}}\right)$    (9)

where, $N$ is the number of samples in a batch, $y_{i, k}$ is the binary indicator (0 or 1) representing if class label c is the correct classification for observation $i$, and $\widehat{y_{l, k}}$ is the model's predicted probability.

To ensure optimal convergence and stable weight updates during the training phase, the network was trained utilizing the Adaptive Moment Estimation with Weight Decay (AdamW) optimizer. Unlike standard Adam, AdamW explicitly decouples the weight decay from the gradient update to improve regularization and model generalization. The parameter update rule at time step $t$ is mathematically formulated in Eq. (10):

$\theta_t=\theta_{t-1}-\eta \frac{\hat{m}_t}{\sqrt{\hat{v}_t}+\epsilon}-\eta \lambda \theta_{t-1}$    (10)

where, $\theta_t$ represents the model weights, $\eta$ is the learning rate, $\lambda$ is the weight decay coefficient, and $\hat{m}_t$ and $\hat{v}_t$ are the bias-corrected first and second raw moment estimates of the gradients, respectively. For this study, the initial learning rate $(\eta)$ was set to 0.0015 with a weight decay $(\lambda)$ of 0.0001 . The training process was executed for a maximum of 40 epochs utilizing a batch size of 64 . Furthermore, to prevent overfitting, a dynamic learning rate scheduler (ReduceLROnPlateau) was implemented alongside an Early Stopping mechanism with a patience of 6 epochs, which actively halted the training process when the validation loss ceased to improve.

2.7 Performance evaluation metrics

To rigorously evaluate the proposed model, especially under imbalanced conditions, relying solely on overall accuracy is insufficient. Therefore, a comprehensive set of evaluation metrics is computed using the True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) values extracted from the confusion matrix [39, 40]. The fundamental evaluation metrics are mathematically defined in Eqs. (11)–(14):

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$    (11)

Precision $=\frac{T P}{T P+F P}$   (12)

Recall $=\frac{T P}{T P+F N}$   (13)

F1-Score $=2 \times \frac{\text {Precision ×} \text { Recall}}{\text {Precision}+ \text {Recall}}$   (14)

Precision measures the exactness of the classifier (minimizing false alarms), while Recall (Sensitivity) measures the completeness (preventing missed critical arrhythmias). The F1-score provides the harmonic mean between Precision and Recall, serving as the most reliable indicator of the model's robustness against severe class imbalance.

4.1 Computational complexity and efficiency

The experimental results reveal a fundamental and unavoidable trade-off between diagnostic performance and computational duration. As highlighted in the evaluation metrics (Table 4), purely oversampling methods (like SMOTE and ADASYN) and their hybrid combinations incur substantially higher training times compared to simple undersampling techniques. For instance, the optimal hybrid method (360_Tomek-SMOTE) requires over 900 seconds to execute, whereas the standalone Tomek-Links method (360_Tomek) concludes in approximately 395 seconds.

image075.png

Figure 7. Performance versus computational time evaluation for the 360-sample window balancing strategies

To visually encapsulate this dynamic, Figure 7 presents a dual-axis comparative analysis of the 360-sample window scenarios. As illustrated by the bar chart, the predictive capabilities remain consistently exceptional across all balancing techniques, strictly hovering above 98%, with the hybrid Tomek-SMOTE methodology achieving the absolute peak performance at 98.76%. However, the overlaid line graph vividly exposes the latent computational cost required to achieve this marginal gain. The standalone Tomek-Links method sacrifices a mere 0.16% in F1-score while slashing the computational burden by more than half.

This significant variance in computational complexity is primarily driven by the underlying algorithmic mechanics of synthetic generation. SMOTE and ADASYN rely heavily on continuous k-NN spatial searches across high-dimensional feature spaces (especially within the expanded 360-sample window). Calculating the Euclidean distances between thousands of minority instances to generate linear interpolations is a mathematically intensive $O\left(N^2\right)$ operation.

However, from a practical deployment perspective, this computational burden is highly acceptable. The 921.96 seconds required for the 360_Tomek-SMOTE scenario is strictly an offline, one-time training cost. Once the 1D-CNN (SE-ResNet) model is fully trained, the inference time required to classify a single 360-sample ECG window takes only milliseconds. Therefore, sacrificing offline training time to achieve a 98.77% diagnostic accuracy is a highly justified trade-off for an automated deep learning classification model, where algorithmic precision and patient safety are the absolute priorities.

4.2 Structural coherence of synthetic electrocardiogram data

A critical and often overlooked concern in the application of medical automation is the structural coherence of synthetically generated signals. The human EGG is highly sensitive to subtle physiological conditions; even minor deviations in the QRS complex width or the ST segment elevation can denote entirely different pathological states. Standard linear interpolation, which serves as the foundation for both SMOTE and ADASYN, inherently risks generating artifactual waveforms that may mathematically belong to a minority class but do not logically exist in real-world cardiology [42, 43].

This risk of structural distortion is exponentially higher when dealing with noisy data or heavily overlapping inter-class boundaries. This phenomenon is perfectly illustrated by the substantial performance degradation of the 180_ADASYN scenario. ADASYN focuses on generating synthetic data for the minority instances that are hardest to learn (typically those residing deep within the majority class boundary). In a narrow 180-sample window lacking temporal context, ADASYN essentially amplifies boundary noise, forcing the CNN to learn clinically invalid, overlapping features, resulting in a severely degraded accuracy of 88.97%.

In stark contrast, the proposed Tomek-SMOTE framework effectively neutralizes this clinical risk. While Tomek-Links alone is computationally light and "safe" (because it only cleans existing data without synthesizing new, potentially invalid waveforms), it is insufficient for extreme imbalance. By strategically applying Tomek-Links before SMOTE, the framework actively deletes ambiguous and noisy borderline samples first. Consequently, when SMOTE performs its linear interpolation, it operates exclusively within a clean, high-confidence minority space. This ensures that the generated synthetic ECG segments maintain consistent morphological alignment with the original minority data, allowing the SE-ResNet architecture to learn robust, real-world diagnostic features rather than artificial noise.

4.3 Comparison with state-of-the-art methods

To validate the effectiveness and superiority of the proposed framework, a rigorous comparative analysis was conducted against several recent state-of-the-art methods published within the last five years. All selected benchmark studies utilized the MIT-BIH Arrhythmia Database, evaluating various advanced deep learning architectures and data handling strategies. The comparative results, focusing strictly on Accuracy, Precision, Recall (Sensitivity), and F1-score, are systematically ordered by publication year and summarized in Table 6.

Table 6. Performance comparison of the proposed method against state-of-the-art approaches on the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) dataset

Note: 1D-CNN = One-Dimensional Convolutional Neural Networks, SE-ResNet = Squeeze-and-Excitation Residual Network, SMOTE = Synthetic Minority Over-sampling Technique.

More recent and highly complex architectures also exhibit distinct limitations when handling severe class imbalance. For instance, both the CNN-Transformer hybrid by Wu et al. [47] and the advanced CNN-ET-LSTM model by Hasan et al. [48] attempted to resolve class imbalance by implementing standard SMOTE.

While Hasan et al. [48] achieved an impressive 97.62% F1-score with their hybrid architecture, the reliance on standard oversampling without prior boundary cleaning inherently leaves the model vulnerable to overlapping synthetic noise in the feature space.

In contrast to the existing literature, the proposed 1D-CNN (SE-ResNet) method achieves a highly effective balance of metrics, recording near-perfect scores of 98.77% in Accuracy and Recall, and 98.76% in F1-score. This superiority is driven by a deliberate, dual-optimization strategy. First, the integration of the hybrid Tomek-SMOTE technique effectively cleans the ambiguous decision boundaries before oversampling, ensuring the generation of structurally representative, non-overlapping synthetic data that actively resolves the boundary pitfalls encountered by previous SMOTE-reliant studies [47, 48]. Second, while previous methods rely on standard convolutions or computationally heavy Transformers, the integration of SE attention mechanisms combined with Residual blocks allows the network to dynamically focus on critical ECG segments, such as the QRS complex, while bypassing vanishing gradients. This architectural synergy allows the proposed deep learning classification model to achieve the highest overall metrics, demonstrating its exceptional reliability in detecting rare but life-threatening arrhythmias in real-world clinical datasets.

4.4 Limitations

Despite the high classification performance achieved, this study acknowledges several critical limitations that must be addressed before practical clinical implementation. First, regarding cross-dataset generalization, the proposed 1D-CNN (SE-ResNet) architecture and the Tomek-SMOTE framework were exclusively trained and evaluated on the MIT-BIH Arrhythmia Database. Although widely recognized as a foundational benchmark, real-world environments often present varying baseline wander and different ECG lead configurations, meaning the model's adaptability across independent datasets remains unverified.

Second, the dataset partitioning was conducted at the intra-patient (beat) level rather than utilizing a strict inter-patient scheme. While stratified beat-level splitting is widely employed in the literature, it inherently allows different ectopic beats from the same patient to appear in both the training and testing subsets. This characteristic may yield more optimistic classification metrics compared to strict subject-wise isolation.

Third, while the hybrid balancing technique successfully mitigates boundary noise, the reliance on synthetically generated ECG samples inherently carries minor risks. SMOTE relies on mathematical interpolations, which may not flawlessly replicate the unpredictable physiological complexities of true clinical arrhythmias.

Fourth, there is an absence of direct clinical validation. The framework has been validated strictly retrospectively using historical data, lacking prospective clinical trials involving active patient monitoring to evaluate its diagnostic impact in a real hospital setting. Finally, this study primarily optimizes offline classification metrics and lacks extensive real-time deployment experiments. Empirical testing on resource-constrained hardware, such as wearable edge devices, is required to definitively evaluate hardware latency, memory consumption, and energy efficiency. Future research must explicitly extend this methodology to conduct cross-dataset validations, inter-patient evaluations, and hardware deployment to ensure absolute clinical robustness.