Deep Learning-based diagnosis of cardiovascular disease from ECG signals involves two major steps: pre-processing (Safdar et al., 2024) and classification (Wu and Guo, 2025). The pre-processing step generally involves resizing and filtering of signals. The classification step involves feature extraction and signal categorization. Popular DL architectures like CNN and its variants have shown significant results in the medical domain, especially with ECG. In the case of ECG as a one-dimensional signal, 1D-CNNs are deployed. These models are enhanced and contributed in many ways: like combining with Leaky-ReLU (Lakhdari and Saeed, 2022) activation function, an enhanced model for extracting signals from paper-based ECG data (Nguyen et al., 2022), with real and noise-attenuated ECG signals (Ahmed et al., 2023), for robust classification 1-D Convolutional deep residual neural networks (Khan et al., 2023) utilized, and even explored in authentication (Yuniarti et al., 2024). Apart from different scenarios, ECG signals can be combined with different advanced CNN variants such as Deep-CNN (Li et al., 2021), SE-ResNet152 (Xu et al., 2021), MobileNetV2 (Cordos et al., 2021), ResNet152V2, DenseNet169, COV-ECGNET (Rahman et al., 2022), MobileNetV2 combination with BiLSTM (Shin et al., 2022), and InceptionV3 (Bhosale and Patnaik, 2023) and showed promising results. In these architectures, ECG signals are subjected to a sequence of convolution and pooling layers for feature extraction and then to a dense neural network layer for decision-making (i.e., classification). In most of the ECG classifications, scalogram images are utilized from the signal by CWT approach, for extracting non-linear and non-stationary features (Gupta et al., 2021), image classification by representing image over the time-frequency domain (Kim, 2021), extracting R peak and RR interval features (Wang et al., 2021), statistical parameters (Alharbey et al., 2022), classification with various CNN variants for better results (Dessai and Virani, 2023), with RGB image classification from scalogram (Kumar and Ramachandran, 2023), for collecting multi-spectral information (Mewada, 2023), and for classification with transformers (Qiu et al., 2024). In one of the cases, the binarized version of the scalogram image is used for classification (Naz et al., 2021).

Owing to the fact that each version of the image contains some unique features, in this work, we propose an approach termed “alpha blending' which blends the features extracted from the scalogram and its binarized version through a sequence of convolution and pooling layers. The blended feature map is subjected to a dense neural network layer (as in traditional CNN) that classifies the image. This proposed architecture is termed as BlendNet. The blending step is flexible as it involves a parameter α. For experiments, a total of 162 ECG recordings from the PhysioNet database were used. There were three categories of patients: Congestive Heart Failure (CHF), Cardiac Arrhythmias (ARR), and Normal Sinus Rhythms (NSR). There are 36 recordings from people with NSR, 30 with people with CHF, and 96 with ARR. The objective of the proposed approach is to classify ARR, CHF, and NSR. The experimental results shows that the BlendNet architecture achieves the best performance α = 0.7, and it outperforms non-blending approaches. The proposed BlendNet is flexible in terms of dense layer settings, as it can accommodate any complicated dense layers, for example, the dense layers in InceptionV3, ResNet152V2, DenseNet169, or MobileNetV2. In applications where execution speed is of utmost concern, the dense layer can be replaced with faster alternatives (i.e., ML algorithms) for faster convergence.

Notable works in literature relevant to this problem statement are the DL framework in Prusty et al. (2024) which utilizes Scale Invariant Feature Transform (SIFT) based features for detecting heart failures and the framework in Saeed and Yousif (2021) which utilizes the slantlet based statistical features. Both these approaches extract features from the PhysioNet ECG data and apply a DL architecture. Although they have exhibited good classification performance, they depend on a single paradigm for feature extraction. On the other hand, the BlendNet framework proposed in this manuscript deploys a blending framework to form a composite feature map, which is then subjected to a classification architecture/algorithm.

The contributions of this manuscript are as follows:

The remaining sections are structured as follows. Section 2 presents a survey of several ML and DL architectures associated with diverse ECG data. Section 3 provides a detailed explanation of the proposed BlendNet and presents an analysis of its computational complexity. Section 4 presents the experimental validation of the proposed architecture and establishes a comparison with the state-of-the-art. Section 5 concludes the paper with recommendations for future research.

Recent research has shown substantial progress in identifying irregular heart rhythms using CNNs. Ahmed et al. (2023) employed a 1-D CNN to classify four distinct categories in the MIT-BIH dataset and achieved a remarkable accuracy of 99%. In another work (Lakhdari and Saeed, 2022), including the LeakyReLU activation function in 1-D CNN architectures on the same dataset results in accuracies ranging from 97% to 99%. A modified version of the 1-D CNN called SEResNet18 was used in a dataset of ECG images containing data from cardiac and COVID-19 patients. The model achieved accuracies of 98.42% to distinguish COVID-19 cases from normal cases and 98.50% to distinguish COVID-19 cases from other classes. In particular, the model successfully extracted signals from the scanned ECG records (Khan et al., 2021; Nguyen et al., 2022). There is another progress where a 2-D CNN achieved an accuracy of 99.52% on the MIT-BIH dataset by combining wavelet-based spectral features with CNN's temporal features. This demonstrates the effective combination of advanced signal processing techniques and DL. These advances highlight the strong flexibility and growing accuracy of CNN models in diagnosing heart conditions using different datasets and architectural improvements.

Many studies associated with ECG have recently been conducted in 2-D format, mainly because of extracting morphological features (Wang et al., 2021), non-linear and non-stationary features (Gupta et al., 2021), comparing images with signals (Kim, 2021), for statistical features (Alharbey et al., 2022), good performance across different CNN variants (Dessai and Virani, 2023), performing well in other imaging like RGB (Kumar and Ramachandran, 2023) and with transformers (Qiu et al., 2024), it worked well. Compared to 1-D signal ECG, 2-D ECG images provide more insight into abnormalities and the interpretation of complex cardiac conditions due to their visual representation that combines frequency and temporal characteristics in a single image. Furthermore, when DL architectures are employed, spatial relationships can be exploited to extract important features and improve the classification process:

Recent studies (Yoon and Kang, 2023) have highlighted the importance of scalogram-based approaches for the interpretation of ECG through bimodal CNNs, combined with ensemble and Inception-v3 techniques, achieving an accuracy of 95.08% and 95.74% in classifying ARR, CHF, and NSR, while (Ozaltin and Yeniay, 2023) expanded this work to accurately diagnose COVID-19 with accuracies of 96.53% using CNN and 99.21% with CNN-SVM% respectively.

An extensive investigation has been conducted using ARR, CHF, and NSR datasets to study automatic ECG signal classification, with (Mohamed et al., 2023) reported accuracies of 96%, 92.66%, and 95.33% by using architectures GoogleNet, AlexNet, and ResNet; additionally, Sabeenian and Sree Janani (2023) achieved an accuracy of 98.81% with ResNet18, while combining CNN with Naïve Bayes (Ajjey et al., 2022) and AlexNet (Olanrewaju et al., 2021), reported accuracies of 98.76% and 98.7%, showing the dominance of DL architectures in accurately classifying ECG signals.

Studies using the UCDDB dataset (Mashrur et al., 2021) have applied a scalogram-based CNN to identify obstructive sleep apnea, with an accuracy of 94.30%. Research using the PTB and MIT-BIH arrhythmia datasets (Byeon et al., 2019) has demonstrated the adaptability of DL architectures, including GoogleNet, EECCGNet, and ResNet, attaining high classification accuracies of 92.29% to 99%. Normalizing binary images for extracting the QRS complex (Wang et al., 2020) and with AlexNet, VGG-16, and Inception-V3 (Naz et al., 2021), reflecting ECG properties, such as QRS complexes and T waves, might be advantageous when integrated with other forms of ECG data. Normalization is implemented to achieve consistency in image scaling, which in turn enables the integration of analysis for improved interpretation and detection of cardiac problems.

Recent advancements in ECG analysis have been driven by the exploration of complex DL architectures for enhanced feature extraction and classification. Researchers have leveraged these architectures to improve diagnostic accuracy and enable automated interpretation of ECG signals. Another recent study used 1,932 paper-based ECG images, which were divided into five classes (MI, HMI, NHB, AHB, and COVID-19), to evaluate the performance of different DL architectures in classification tasks. The DenseNet201 (Rahman et al., 2022) model was employed for binary classification, attaining an accuracy of 99.1%. For the classification of three classes, DenseNet201 (Rahman et al., 2022) achieved an accuracy of 97.36%, while InceptionV3 (Rahman et al., 2022) achieved an accuracy of 97.83% for the classification of five classes. A different research study utilized the identical dataset and implemented the In-Res106, InceptionV3, ResNet50, DenseNet201, VGG19, and MobileNetV2 architectures (Fatema et al., 2022) to develop an automated system for predicting heart disease. Impressive levels of accuracy were attained, with In-Res106 emerging as the top performer with a score of 98.34%. In addition, a distinct study that specifically examined the categorization of cardiac disorders (ARR, CHR, NSR) found that the SIFT-CNN (Prusty et al., 2024) attained a remarkable accuracy rate of 99.78%. Nevertheless, alternative techniques such as SVM, K-Nearest Neighbors (KNN), Long Short-Term Memory (LSTM), and AlexNet-SVM (Cnar and Tuncer, 2021) produced diverse outcomes, with accuracy rates ranging from 65.63% to 96.77%, when applied to classification and prediction tasks. A slantlet based feature extraction followed by an SVM classifier has resulted in an Area Under Curve (AUC) of 99.25% (Saeed and Yousif, 2021).

In a different scenario, instead of opting for complex DL architectures, a particular study investigated the use of convolution-based heterogeneous activation facility (CHAF) (Narayanan, 2023). This approach involves employing multiple activation functions (AFs) in the convolution layer blocks, with each block having its own AF. The aim is to extract features more effectively and enhance accuracy. The study achieved an accuracy of 99.55% with an execution time of 0.008 seconds using the CHAF-KNN method on the PTB dataset. Similarly, with the MIT-BIH dataset, the CHAF-KNN method achieved an accuracy of 99.08% and executed in 0.07 seconds.

In similar to our proposed work, there is research work that combines phase and magnitude (Scarpiniti, 2024) of CWT and got 98.5% accuracy; in another work, a time-frequency-based DL framework (Karimulla and Patra, 2025) achieved 94.60% accuracy; and lastly, in another research (Ahmad et al., 2021) uses multimodal image fusion and multimodal feature fusion techniques were used to achieve 99.2%.

Table 1 presents a summary of DL architectures for ECG classification. During our investigation into different DL architectures and ML algorithms for a range of datasets, with the goal of improving feature extraction and model performance in healthcare applications, we have discovered a potential area for innovation: combining the features of two images created from ECG signals.

This section presents a detailed description of the proposed BlendNet architecture, an analysis of BlendNet's computational complexity and an overview of its advantages and limitations.

This section contains dataset description, implementation details, definitions of metrics used for evaluation and the experimental results with related discussions.

This work proposed “BlendNet,” a novel DL architecture that effectively extracted the features of an ECG signal using a blending approach termed “alpha blending.” The blended feature map is subjected to a dense neural network layer (as in traditional CNN) that classifies the image. The utilization of alpha blending facilitated the generation of a composite feature set that incorporated different characteristics of a 2D-represented ECG signal from its scalogram and binary versions. Experimental results on the PhysioNet dataset showed that the BlendNet has its best performance for α = 0.7. The result of the ablation study showed that, in the case of α = 0.7, BlendNet's performance was better than the performance of its traditional counterparts (i.e., CNN on only scalogram images and CNN only on binarized images). Proposed BlendNet is shown to be flexible in terms of dense layer settings. For applications demanding complicated neural network architectures, BlendNet can be deployed with dense layer settings as in InceptionV3, ResNet152V2, DenseNet169, or MobileNetV2. For applications demanding faster execution times, the dense layer can be replaced with ML algorithms such as XGBoost for faster convergence.

Limitation of the proposed BlendNet: As the blending proportion α deviates from 0.7, the performance starts to degrade. It is also dataset dependent. An approach needs to be devised for estimating an optimal value of α.

Recommendations for future work: (i) As an extension of this work, the effectiveness of blending can be improved by considering other imaging modalities; (ii) The composite feature set resulting from blending can be used in generalized adversarial networks popularly known as GANs.