Hybrid Deep Learning Framework for Cardiac Rhythm Detection Using Vision Transformer and EfficientNetB7

Detection of cardiac rhythm is an obligatory instrument in diagnosing and managing a wide variety of cardiovascular diseases that are still the most common causes of mortality and morbidity in the whole world [1, 2]. Detection of arrhythmia and abnormal cardiac rhythm has important clinical implications; the timeliness of arrhythmia alerts can play an important role in early treatment and intervention of patients [3]. Traditional diagnostic approaches, such as clinician-dependent manual interpretation of ECGs, are time-consuming and suffer from interobserver variability, which causes the demand for automatic, reliable, and time-saving systems [4]. Recent developments in deep learning have illuminated the exciting future in medical image analysis, which includes cardiac imaging and ECG interpretation [5].

Recent developments in artificial intelligence, in particular deep learning, offer a potential advantage for improving ECG interpretation by automatically learning relevant features and discriminating different types of complex cardiac rhythms [6]. A statistical summary is also another type of representation widely used in ECG classification, as it allows approximate invariances to be estimated, making easier the extraction of a compact representation of the input signal [7]. Nevertheless, despite their remarkable success, CNNs are inherently poor at dealing with long-range temporal dependencies and global context for subtle arrhythmic patterns, especially over prolonged time windows. There are a few works trying to mitigate these limitations, among which are the transformer-based models that were first proposed for natural language processing and proved to be powerful in modeling long-distance dependencies on designed sequential data and have recently been adopted in certain computer vision tasks such as medical imaging.

In particular, the convolutional neural networks (CNNs) are widely used to learn spatial patterns in medical images but can hardly handle long-range dependence and global context. To tackle the aforementioned challenges, a recent trend in research has prompted scholars to embrace transformer-based architectures, which have transformed natural language processing and computer vision tasks [3]. The deep learning libraries that were employed on a number of the best-performing Vision Transformer (ViT) models that are known to excel in capturing global relationships in images using the self-attention mechanism [8], these types of models seem appropriate to solve complex pattern recognition tasks within health care. In contrast, EfficientNetB7 [9] is an efficient CNN model architecture because of the SOTA accuracy while utilizing computational resources with the help of compound scaling to scale the depth, width, and resolution during training [10].

This paper introduces a novel hybrid deep learning architecture that integrates the strengths of Vision Transformer (ViT) and EfficientNetB7. This innovative framework demonstrates the ability to attain high accuracy and robustness in cardiac rhythm detection. The suggested method has a two-stage architecture. EfficientNetB7 is the feature extractor that captures fine-grained local representations of ECG images, and the Vision Transformer component model captures the global dependencies and context between them. The current framework suggests using each of these architectures to fill in the gaps and improve our understanding of heart rhythms. The hybrid model is trained and tested with the public benchmark ECG data set so that the results are clear and can be repeated.

Some strict preprocessing methods, including signal denoising, normalization, and image augmentation, were applied to improve the data quality and model generalization. Further, we incorporate explainability techniques (GradCAM) to visualize the important regions in input images on which the model focuses more while predicting results and enhance clinical interpretability and trust. Comparisons with state-of-the-art methods are made to prove that the performance of our method can achieve better accuracy, sensitivity, specificity, and F1 score coming soon in clinical practice. This research not only contributes to the development of state-of-the-art AI-powered diagnostic tools but also to counteracting immediate needs for scalable and trustworthy solutions in resource-constrained health. The value of this hybrid deep learning system may be to enable automatic, accurate, and efficient detection of cardiac rhythm abnormalities that could lead to timely diagnosis by clinicians and ultimately reduce the impact of diagnostic failures and improve patient outcomes. We hope that our work will both push the frontier with respect to medical deep learning and show that a fusion of Vision Transformer and EfficientNetB7 could be combined effectively for merging two approaches, opening new doors for future automatic cardiac diagnostics as well as beyond in the context of healthcare applications.

The research is organized into sections: Introduction, Related Works, Proposed Method (which describes data preprocessing and model architecture), Training Procedure, Evaluation Framework, Results (which include a comparative analysis, an ablation study, computational efficiency, generalization, error analysis, explainability, and clinical relevance), and References that support its strong, understandable, and usable cardiac rhythm detection system.

To achieve this, we proposed a hybrid deep learning model integrating EfficientNetB7 and Vision Transformer for cardiac rhythm monitoring. First of all, the ECG signals are pretreated with noise filtering, normalization, and image augmentation techniques to obtain high-quality ECG images. The attention mechanism is designed based on EfficientNetB7 for extracting fine-grained local features and uses Vision Transformer instead to estimate global context relationships. After that, these extracted features are fused and propagated across the fully connected layers to proceed with classification. The model is trained with a cross-entropy loss followed by the Adam optimizer learning schedule to ensure convergence. The performance is measured in terms of accuracy, sensitivity, specificity, F1 score, and ROC AUC. Explainability is brought to the fore by GradCAM for visualization of decision regions in order to maintain clinical interpretability. Comparison with baseline CNN and stand-alone ViT models demonstrates the superiority of our proposed action in enabling stable and interpretable automation for automated cardiac rhythm diagnostics, as shown in Figure 1.

Figure 1. Hybrid deep learning framework for cardiac rhythm detection

3.1 Data collection and dataset description

Three publicly available ECG datasets are exploited in the proposed approach to achieve robust and generalizable cardiac rhythm detection performance. The first dataset is the MIT-BIH Arrhythmia Database; it consists of more than 48,000 ECG beats annotated from 47 subjects and provides diverse normal and abnormal rhythm samples. The second dataset, China Physiological Signal Challenge 2018 (CPSC2018), provides 6,877 12-lead ECG recordings associated with nine rhythm classes and serves to multiclass classification in a clinical context. The third data set is the PTB Diagnostic Database, comprised of 549 ECG recordings of 290 subjects with actual clinical diagnoses, e.g., myocardial infarction and other diseases. Each dataset needs to be preprocessed, and that includes getting rid of noise, scaling the data, and converting the ECGs from waveforms into an image representation ready for deep learning. Variability in the lead maps, sampling rates, and patient populations between these datasets enhances the models' generalized ability to be employed in real-world clinical applications. Through integration of these complementary resources, we present a method to train and evaluate a hybrid EfficientNetB7 and Vision Transformer model with a quantifiable better Trimodal AUC (Z = 9.195, p < 0.0001) profile that can likewise differentiate between normal and abnormal cardiac rhythms more precisely with better interpretability and clinical utility.

3.2 Data preprocessing

Data processing is an important preprocessing procedure to offer high-quality inputs for the hybrid deep learning model. ECG signals from three raw databases are initially filtered by wavelet denoising and baseline wander removing, which can help reduce noise and improve the clarity of the signal. After the removal of noise in ECG signals, they are normalized into the same range to alleviate intersubject differences and stabilize model learning. The ECG templates are also transformed into 2D images by plotting the traces in grayscale as ECG images that preserve both temporal and morphological characteristics beneficial for diagnosis. To avoid overfitting, rotation, horizontal flip, scaling, and brightness of the profiles as data augmentation are considered to augment more artificial training data. To guarantee fair comparison as well as equitable assessment, training/validating/test splits are made so that the ratio (or at least proportion) of different classes remains identical and no patient appears in multiple sets. In addition, all images are resized and normalized to the same size as what the EfficientNetB7 model would take. They go on to use batch normalization and random noise injection for improved model stability. With well-curated data processed by the aforementioned steps, the model can be fed high-resolution inputs specifically designed to capture localized and global cardiac rhythm patterns for both the EfficientNetB7 and Vision Transformer subnetworks.

All datasets were preprocessed with identical quantification parameters. A 0.5 Hz high-pass Butterworth filter removed baseline wander, and wavelet denoising with a Daubechies-6 mother wavelet with a soft threshold of 3σ removed both muscle and powerline noise. We made the signals normal by applying z-score normalization (with μ = 0 and σ = 1) to each recording. To construct the image, a 2D plot was drawn over ECG segments with an image resolution of 224 × 224 pixels (equal to that of input for EfficientNetB7), using a sampling window of 2.5 seconds. Data augmentation comprised random Gaussian noise (σ = 0.01), rotations (± 10°), and changes in the brightness (+/−15%). These values ensure that input quality remains constant and the model is as strong as possible.

We also included class imbalance effectively in the training so that no bias would be given to majority rhythm classes. We employed a combination of (i) class-balanced sampling to ensure equal probability was given for each rhythm category during minibatch formation and weighted categorical cross-entropy, where the class weights were computed as reciprocal 1:class frequency, and (ii) targeted augmentation, which was applied only on minority classes with the goal to enforce distribution diversity of representation without further expanding majority distributions. These approaches minimized the misclassification of rare arrhythmias, particularly supraventricular and ventricular ectopic beats. They also contributed to enhancing the sensitivity and F1 score of minority classes.

3.3 Model architecture

The hybrid model is made up of the EfficientNetB7 and the Vision Transformer (ViT), which is intended to extract rich local fine-grained information as well as global context in ECG images. The architecture is started with the EfficientNetB7 feature extractor (a CNN that uses compound scaling, a way to uniformly scale network depth, width, and resolution). The formulation of the compound scaling is characterized by the following:

depth $=\alpha^\phi$, width $=\beta^\phi$, resolution $=\gamma^\phi$               (1)

Subject to the constraint α⋅β²⋅γ²≈2 where ϕ is the user-defined scaling coefficient and α,β,γ are constants determined through grid search. EfficientNetB7 is a strong feature extractor and constructs the feature maps, which contain local morphological descriptions of the ECG traces in terms of QRS complex shapes, levels, and ST-segment hours. The extracted features are passed to the Vision Transformer, where the feature map is partitioned into a set of fixed-size patches and each of the patches is embedded into a token vector. The patch embedding process can be summarized as:

$z_0=\left[x_{\text {class }} ; E x_p\right]+E_{\text {pos }}$              (2)

where, x_p are the flattened image patches, E is the learned embedding matrix, and E_pos is the positional encoding matrix. The embedded tokens undergo a series of transformer encoder blocks comprising multi-head self-attention layers and feedforward networks. In the suggested framework, a two-stage mechanism is used to combine features. First, a linear layer flattens and projects the final convolutional feature map from EfficientNetB7 (size: 8 × 8 × 2560) to match the embedding dimension of the ViT output (768 units). By pooling class tokens, the Vision Transformer creates a sequence-level embedding, which results in a 1 × 768 global representation. The fused representation is a 1 × 1536 feature vector made by putting together the projected EfficientNetB7 vector and the ViT embedding. Then, this vector goes through a fusion MLP with two fully connected layers (1536 → 1024 → 512) and GELU activation and dropout (p = 0.2). We tried attention-based fusion, but we didn't use it because it was too expensive to run and didn't show any performance improvements. The classification head uses the final fused embedding as input. The self-attention mechanism is formalized as

$\operatorname{Attention}(Q, K, V)=\operatorname{softmax}\left(\frac{Q K^T}{\sqrt{d_k}}\right) V$                (3)

where, Q, K, and V are the query, key, and value matrices formed by input tokens, and dk is the dimensionality of the key vectors. The features from EfficientNetB7 and the Vision Transformer are eventually combined via concatenation or attention-based fusion layers, which ensure that local and global representations are aligned. The final layers in the concatenated vector are fully connected and have a softmax activation function classifying the rhythms into more than one class. This model is fine-tuned with cross-entropy loss.

$\mathcal{L}=-\sum_{i=1}^N y_i \log \left(\hat{y}_i\right)$             (4)

where, $y_i$ and $\hat{y}_i$ are the true and predicted class probabilities. The design ensures complementary learning, enabling robust, interpretable detection of diverse cardiac rhythms.

Our feature extractor is EfficientNetB7. It is a compound scaled fine-tuned convolutional neural network. The hybrid architecture uniformly adjusts depth (layers), width (channels), and resolution for power-of-two scaling, striking the right balance between efficiency and accuracy. We initialize EfficientNetB7 with pretrained ImageNet weights and adapt it to process ECG image inputs in order to discover domain-specific features and avoid overfitting. Whilst generating patches in the context of this and the embedding stage, the Vision Transformer module divides extracted features into patches of fixed size. Then it flattens and linearly projects each patch onto a high-dimensional embedding with learnable positional encodings to understand where in space things are. There are global interdependence and temporal linkages among ECG waveforms. The multi-head self-attention mechanism extracts these by computing attention weights across all patches. Positional encoding injects information about the positions of things that’s otherwise missing from pure attention developments, such as sine and cosine functions or learned embeddings. This is useful in allowing the model to distinguish between features that are consecutive. The hybrid integration fuses local spatial EfficientNetB7 features with global contextual embeddings of ViT through concatenation or direct or attention-based fusion between feature vectors, based on learned weights that emphasize relevant locations. Adding more representation layers, fused representations are in good shape for processing when they are concatenated with linear projection layers or adaptive pooling. This is what the dense layers look like for a classification task. The Vision Transformer module is developed from the ViT-Base configuration but performed with ECG image patches. We used a model that consists of 12 transformer encoder layers with 12 attention heads; the hidden dimension is 768. The MLP ratio is chosen to be 4, and thus the feed-forward layers are wide with 3072 pixels. Each image has 196 patches, and each patch is of size 16 × 16 pixels. The positional encoding is learned rather than being sinusoidal so that it can more easily cope with variation in the ECG shape. Layer normalization (LN) is employed just before the attention and MLP blocks, and residual connections are used throughout. A dropout of 0.1 is employed on the fully connected block and patch embeddings. These hyperparameters were chosen based on initial tuning through the three benchmark data sets. The combined approach uses convolutional neural networks (CNNs) to extract detailed information about the shape of things and transformers to model long-range relationships in order to identify heart rhythms accurately and meaningfully for clinical use.

3.4 Training procedure

The hybrid EfficientNetB7 & Vision Transformer-based model is well trained for robust convergence and generalization over ECG datasets. The most common loss function employed is the categorical cross-entropy loss (eq. Both apply to the true and the predicted class probability and optionally run experiments with focal loss to solve the classes' imbalance through weight modulating factors applied to the loss (emphasizing* very hard) examples. The preferred optimization method is Adam, which couples adaptive learning rates with momentum to accelerate convergence given the parameter updates involving estimates of first and second moments, while for comparison stochastic gradient descent (SGD) is used as a baseline alternative. Within the architecture, a learning rate scheduling mechanism is adopted, which includes dropping the learning when validation loss fails to improve further, thereby increasing the possibility of fine convergence and reducing the likelihood of local minima. Regularization methods, such as dropout layers that stochastically deactivate neurons during training to prevent co-adaptation, as well as the use of weight decay that discourages high weight values with L2 regularization. Advocate for simpler models that generalize more. Early stopping observes the validation performance and terminates training when no improvement is observed over a number of epochs (avoids overfitting and leads to model efficiency). The entire training procedure is performed in a high-performance computing system using NVIDIA GPUs for parallelization and tensor operations. There are various libraries in systems like TensorFlow [4] and PyTorch [5] that come with significantly large packages of mixed-precision training, automatic differentiation, and efficient fast data loading. These careful choices of loss functions, optimization techniques, learning rate control schedules (learning rate), regularization, and early stopping strategies will guarantee that the hybrid model achieves a satisfactory trade-off between accuracy (performance) and stability as well as being computationally affordable/cost-effective and reproducible to support clinical practice in terms of automatic cardiac rhythm detection.

All hyperparameters employed for training are explicitly described to ensure reproducibility. The hybrid model was trained for 60 iterations per batch (32 items). The initial learning rate was 1 × 10⁻⁴, and it was halved when the validation loss did not decrease any further (patience = 5). The Adam optimizer is configured with β₁ = 0.9, β₂ = 0.999, and ε = 1 × 10⁻⁸. We applied a weight decay (L2 regularization) of 1 × 10⁻⁵ to all trainable parameters. The dropout rates for the fully connected layers of EfficientNetB7 were set as 0.2, and all Vision Transformer encoder blocks utilized a dropout rate of 0.1. To ensure that the behavior was consistent across all folds, we ran all experiments with a fixed random seed (42). Mixed precision training (FP16) was enabled to increase training efficiency with no loss in accuracy. These explicit hyperparameter settings allow other researchers to replicate how the model operates, as well as what it does.

3.5 Evaluation and experimental validation framework

To show the efficacy and clinical importance of the proposed hybrid EfficientNetB7 and Vision Transformer model for detecting cardiac rhythm, we designed an evaluation and testing framework that extensively examines multiple quantitative criteria, including accuracy, sensitivity, specificity, precision, and F1 score, as well as AUC (area under receiver operating characteristic curve) for various balanced or imbalanced classes. This confusion matrix analysis is shown to provide a more detailed understanding about the true positive, false positive, false negative, and true negative rates in diagnostic strengths and proud_-failure modes. Explainability and visualization are the core of this framework, where the GradCAM heatmaps highlight the most salient ECG regions responsible for the decision of models, and the attention map interpretation of the Vision Transformer reveals processes underlying global context modeling for better clinical interpretability. Documenting the experimental setup, for example, the hardware and software version numbers of libraries used, allows for the replication of the experiment. Random seeds and configuration files are saved to reproduce the training and testing. We perform comparisons to baselines from not only CNNs but also pretrained Vision Transformers and state-of-the-art models in order to provide insight into the strength and limitations of our model. The selection of baselines is justified by their architectural significance, relevance to the dataset, and documented prevalence in the literature. By comprehensive metric-driven analysis, explainable & visualizable interpretability, replicable experiment protocols, and transparent comparative studies, we demonstrate that the model is accurate and clinically useful to an extent that it's robust and ready for practice launch, leading to safer and more reliable automated cardiac rhythm diagnosis.

A consistent experimental protocol was performed to enforce fair comparison over the MITBIH, CPSC2018, and PTB Diagnostic benchmarks. The exact same normalization, segmentation, and image conversion were done to every dataset. In order to avoid data leakage, a strong patient-level separation as well as the 70/15/15 train-validation-test split ratio was followed (the script and data are available at Koenka/hepdc_khoroshilova_2020). All models, including baselines, were trained with identical hyperparameters, batch sizes, and learning-rate schedules to permit fair comparisons. Furthermore, the formulas for all evaluation metrics (accuracy, sensitivity, specificity, F1 score, and AUC) were used. This regular experimental pipeline ensures methodological coherence and makes relevant comparisons (cross-dataset) trustable and meaningful.