Squeeze-ICNN architecture for cardiovascular disease detection using ECG image: training of shape, deep, and LGXP feature extractors

Preprocessing the incoming ECG images is a crucial step in improving its quality and efficiency for further analysis or classification work. An ECG signal image is typically exposed to many kinds of noise. Applying filters, like a median or bandpass filter, is helpful in lowering noise and improving the quality of the signal. Let us assume that the input ECG image is I^ECG. An improved median filtering technique is used as part of the preprocessing step to decrease the undesired or irregular value of pixels variance in an image. The IMF’s step-by-step procedure is described below.

The method of extracting features from preprocessed ECG images, IpECG I_p^{ECG} entails choosing specific information or features from images that can be utilized to illustrate and examine the fundamental trends. In this work, LGXP features, Shape, as well as Deep features are extracted from IpECG I_p^{ECG} .

b.ii.

Shape feature F^SHAPE

Extraction of shape features is an essential part of image processing as it collects the geometrical properties of components in an image. In this paper, an improved hierarchy of skeleton features are extracted from the IpECG I_p^{ECG} . Features are extracted from the skeleton representation of the preprocessed image and these features are organized or analyzed hierarchically [30]. To enhance the histogram features of the image, an improved hierarchy of skeleton features is proposed in the shape features.

b.ii.i.

Improved hierarchy of skeleton features

According to the proposed improved hierarchy of skeleton features, we considered 5 × 5 sub blocks of images, which is shown in Figure 2. The following steps are followed in the improved hierarchy of skeleton features.

• Step 1: Initially, iterate through each pixel in the image.

• Step 2: If the current pixel is the target pixel, proceed to the next step to evaluate the next point.

• Step 3: Count the number of target 8-neighbor pixels. Here, the modified neighbor pixel is calculated as per Eqs (7)–(14). P₂₂ is represented as the center pixel, P_ij indicates the Pixel of the i^th row in the j^th column, and np_l indicates the neighboring pixel (l = 1,2,3,4,5,6,7,8): (7) np1=medianP11,P00 {\rm{n}}{{\rm{p}}_1} = {\rm{median}}\left( {{P_{11}},{P_{00}}} \right) (8) np2=medianP12,P02 {\rm{n}}{{\rm{p}}_2} = {\rm{median}}\left( {{P_{12}},{P_{02}}} \right) (9) np3=medianP13,P04 {\rm{n}}{{\rm{p}}_3} = {\rm{median}}\left( {{P_{13}},{P_{04}}} \right) (10) np4=medianP23,P24 {\rm{n}}{{\rm{p}}_4} = {\rm{median}}\left( {{P_{23}},{P_{24}}} \right) (11) np5=medianP33,P44 {\rm{n}}{{\rm{p}}_5} = {\rm{median}}\left( {{P_{33}},{P_{44}}} \right) (12) np6=medianP32,P42 {\rm{n}}{{\rm{p}}_6} = {\rm{median}}\left( {{P_{32}},{P_{42}}} \right) (13) np7=medianP31,P40 {\rm{n}}{{\rm{p}}_7} = {\rm{median}}\left( {{P_{31}},{P_{40}}} \right) (14) np8=medianP20,P21 {\rm{n}}{{\rm{p}}_8} = {\rm{median}}\left( {{P_{20}},{P_{21}}} \right)

• Step 4: Independently compute the length of lines formed by the target pixels and its median of the two neighbors.

Figure 2:

Neighborhood diagram of improved hierarchy of skeleton feature.

Therefore, the histogram shape features from the preprocessed image is represented as F^SHAPE.

b.iii.

Deep features, F^DEEP

Deep feature extraction is the process of automatically recognizing and extracting hierarchical feature representations from images using pretrained DL models. In this proposed model, pretrained NNs such as ResNet [31], VGG 16 [32], and Inception V3 [33] model’s features are extracted from the preprocessed image. To acquire deep features, feed the preprocessed image through the selected layers of the pretrained CNN. An image representation in a high-dimensional feature space is produced by these layers.

b.iii.i.

ResNet

Residual learning is introduced, in which residue functionalities are learned through the application of shortcut linkages. This facilitates the training of extreme DNNs and aids to mitigate the problem of vanishing gradients. Here, the information can move directly via the network due to identity shortcut links. The basic building blocks, known as residual blocks, have connections that skip. ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152 are some of the variations of the network; the number denotes its depth.

b.iii.ii.

VGG 16

VGG16 is simple and uniform while highlighting the network’s depth. It consists of 3 fully connected layers and 13 convolutional layers, making 16 weighted layers. It uses 3 × 3 convolution filters, which are tiny reception fields that are all throughout the architecture. It significantly down-samples using max-pooling layers.

b.iii.iii.

Inception V3

The inception uses parallel filters of various sizes to catch features at various scales. There is a parallel concatenation of 1 × 1, 3 × 3, and 5 × 5 convolutions in inception modules. Prior to fully connected layers, spatial dimensions are reduced using global average pooling. Normalization in batches is done for quicker convergence.

Thus, these features capture a high-level representation of the image, which is valuable for feature extraction. The extraction of deep features is represented as F^DEEP. Lastly, the final feature is represented as F, F = [F^LGXP F^SHAPE F^DEEP], which is extracted from the preprocessed image:

When preprocessing has been carried out, ECG images are passed through ResNet, VGG16, and InceptionV3 pretrained models with the classification layers at the end of the models removed. Pretrained models extract deep features representing various levels of abstraction. VGG16 captures spatial information, ResNet extracts deep residual characteristics, and InceptionV3 provides multiscale representations of the ECG images. The extracted features beyond what comes through a Squeeze-ICNN classifier are then concatenated with handcrafted features to build a hybrid input.

Classifying CVDs using a hybrid model typically involves combining models to improve the overall performance. In this proposed work, the Squeeze-ICNN model is proposed for the classification of CVDs. The extracted LGXP features, Shape features, and Deep features are combined (F) and given as the input to Squeeze-ICNN model, which combines the SqueezeNet and IMP-CNN models. Thus, the average of the classification models gives the output of classifying the CVD as 0 and 1. Here “0” indicates the normal patient and “1” indicates the abnormal patient, who is affected with CVD. The process architecture of the classification models is shown in Figure 3, which is explained in this section.

Figure 3:

Architecture of classification process. CNN, convolutional neural network.

c.i.

SqueezeNet

The SqueezeNet model [34] is used for classifying CVD. Similar to numerous other DL models, it is used to train various data types, encompassing medical images or other pertinent features linked to CVD.

Utilizing 1 × 1 convolutions, often referred to as point-by-point convolutions, is done to lower the network’s parameter number while keeping expressive power is Squeeze Net’s primary feature by using Fire Modules. SqueezeNet provides a computationally efficient solution that can be deployed on low-resource devices because of these tiny filters. The final retrieved features F are fed to the SqueezeNet model. Convolution layer (conv1) is the first in SqueezeNet’s design, while convolution layer (conv10) is the last. It contains eight fire modules in it. Additionally, it carries out max-pooling with a stride of two following conv1, fire4, fire8, and conv10. Following the fire module, a 50% dropout ratio is utilized. Figure 4 shows the SqueezeNet design.

Figure 4:

SqueezeNet model.

c.ii.

ICNN model

CNNs are a class of DNNs that are particularly effective in tasks involving images and spatial data. In the conventional CNN model [35] is included three main layers namely, Convolutional Layers, Pooling Layers, and Fully connected Layers. Figure 5 shows the conventional CNN model. Here, the conv. block of CNNs is the convolutional layer. ReLU is the most common activation function in CNN. The pooling layer is used to down-sample the spatial dimensions of the input volume. Lastly, the fully connected layer follows the convolutional and pooling layers, which is used for making the final classification. As per this proposed work, an IMP-CNN model is proposed to improve its performance and efficiency. An ICNN model consists of five convolution layers that include the input layer with batch normalization (BN) and Leaky ReLU activation function. The architecture of the proposed ICNN model consists of five convolution layers, a flatten layer with 0.5 dropout layer, three dense layers with ReLU activation, and an output layer with the proposed activation layer, which is shown in Figure 6.

Figure 5:

Conventional CNN model. CNN, convolutional neural network.

Figure 6:

Architecture of ICNN model. ICNN, improved convolutional neural network.

c.ii.i.

Convolutional layer

A convolution layer [36] performs a convolution on the given final feature, F. There are five convolutional layer blocks with strides 2 and the kernel is (1, 1) involved in the convolution part. Each convolution block contains the Convolution Layer, Leaky-ReLU Activation function, and BN. The convolution operation in every stage contains element-wise multiplication of the filter with the input feature, F. Let the kernel size be n × m, which is applied on the input feature F and is denoted as f_k. There are n × m input connections involved in the CNN layer and their output is calculated as per Eq. (15), wherein (i,j) indicates the position of the feature map: (15) CFu,v=∑i=−n2n2∑j=m2m2fki,j*Fu−i,v−j C\left( {{F_{u,v}}} \right) = \sum\limits_{i = - {\raise0.7ex\hbox{$n$} \!\mathord{\left/ {\vphantom {n 2}}\right.}\!\lower0.7ex\hbox{$2$}}}^{{\raise0.7ex\hbox{$n$} \!\mathord{\left/ {\vphantom {n 2}}\right.}\!\lower0.7ex\hbox{$2$}}} {\sum\limits_{j = {\raise0.7ex\hbox{$m$} \!\mathord{\left/ {\vphantom {m 2}}\right.}\!\lower0.7ex\hbox{$2$}}}^{{\raise0.7ex\hbox{$m$} \!\mathord{\left/ {\vphantom {m 2}}\right.}\!\lower0.7ex\hbox{$2$}}} {{f_k}\left( {i,j} \right)*{F_{u - i,v - j}}} }

Here, Leaky ReLU is the non-linear activation function, which is used as an extension of the traditional ReLU activation function. It is applied to the resultant features obtained by the convolution layer.

c.ii.ii.

Pooling layer

The convolution layer’s output that passed into the pooling layer generates the feature map. The pooling layer [36] applies the subsampling on the feature map. A pooling layer offers a standard downsampling procedure that lowers the feature maps’ in-plane dimensions to create a translation invariance to slight shifts and distortions and decrease the number of subsequently trainable parameters. The filter size, stride, and padding in pooling operations are the hyperparameters, similar to those in convolution operations, but it is significant that none of the pooling layers are equipped with learnable parameters. Averaging pooling produces the average of the values obtained from overlapped kernel areas and the input image, whereas maximal pooling produces the maximum value. Translational invariance of this max pooling is dependent on the filter size. Let us consider F to be the input and m to be the kernel size of the i^th value. Eq. (16) provides the computation of the maximum pooling output: (16) MFi= maxFi+k,i+1k≤m2,I≤m2k,I∈N M\left( {{F_i}} \right) = \max \left\{ {{F_{i + k,i + 1}}\left| {\left| k \right| \le {\raise0.7ex\hbox{$m$} \!\mathord{\left/ {\vphantom {m 2}}\right.}\!\lower0.7ex\hbox{$2$}}} \right.,\left| / \right| \le {\raise0.7ex\hbox{$m$} \!\mathord{\left/ {\vphantom {m 2}}\right.}\!\lower0.7ex\hbox{$2$}}k,I \in N} \right\}

c.ii.iii.

Fully connected layer

Rich information from each neuron in the preceding layer (the Flatten Layer) is gathered by a fully connected layer. In the proposed work, three Dense Layers with ReLU activation Fully connected layers are included. And the Last layer is the Dense Layer with updated Activation function, which is the output layer.

A Comb SigHyper-Hsine activation function is proposed in the Last layer of ICNN. It is the combination of SigHyper [37] and comb-H-sine [38] (CHS) activation function to enhance the classification process. Eq. (17) shows the comb-H-sine activation function derivatives, where the value of β is taken as 0.5. Eq. (18) shows the proposed comb sigHyper-Hsine activation function.

(17) CHS=sinhβx+sinh−1βx CHS = \sinh \left( {\beta x} \right) + {\sinh ^{ - 1}}\left( {\beta x} \right) (18) comb sigHyper−Hsine=12*ex2+ex−1ex+e−x+e−2x+1+sinhβx+sinh−1βx=12*2ex+e2x−1ex+e−x+e−2x+1+sinhβx+sinh−1βx \matrix{ {comb\;sigHyper - H\sin e} \hfill \cr { = {1 \over 2}*\left( {{{{e^x}\left( {2 + {e^x}} \right) - 1} \over {{e^x} + {e^{ - x}} + {e^{ - 2x}} + 1}} + \sinh \left( {\beta x} \right) + {{\sinh }^{ - 1}}\left( {\beta x} \right)} \right)} \hfill \cr { = {1 \over 2}*\left( {{{2{e^x} + {e^{2x}} - 1} \over {{e^x} + {e^{ - x}} + {e^{ - 2x}} + 1}} + \sinh \left( {\beta x} \right) + {{\sinh }^{ - 1}}\left( {\beta x} \right)} \right)} \hfill \cr }

Finally, both the ICNN and SqueezeNet models are averaged together to get the output of the proposed model. The output of the classification is either in “0” or “1” as normal or abnormal patients.

The Squeeze-ICNN model proposed is structurally and functionally different from AlexNet. It uses the Fire modules of SqueezeNet that provides the same representation with efficiency in the number of parameters and uses a Comb SigHyper-Hsine activation function instead of ReLU to improve the stability of the gradient at a minimum. Adopting a more programmable approach, Squeeze-ICNN’s model is designed to have multiple branches with fusion of deep, shape, and texture features as well as the use of dropout and BN; resulting in a lighter, more robust model suitable for classification of ECG images with limited datasets.

The proposed architecture is designed to tackle the challenges of ECG image noise and limited data. A motion, EMG, and powerline noise-based adaptive median filter was employed to suppress low-to-moderate amplitude noise while preserving possible waveform details. We used a hybrid architecture blending the lightweight Fire modules of SqueezeNet with an IMP-CNN that allowed for efficient feature extraction with robust learning; the IMP-CNN uses a customized activation function that enhanced gradient flow to the classifier block with the small training ECG datasets, all while keeping the model slim and deployable.

The proposed CVD detection using ECG signal images were implemented in PYTHON, especially version “3.7.” The computational tasks were carried out on an “AMD Ryzen 5 3450U processor with Radeon Vega Mobile Gfx, operating at 2.10 GHz,” and with a total installed RAM of “16.0 GB.” In addition, the assessment of the performance of CVD detection using ECG signal images was executed by utilizing the ECG Images dataset of Cardiac Patients [39].

A set of ECG images from individuals with cardiac conditions has been compiled at the “Ch. Pervaiz Elahi Institute of Cardiology in Multan, Pakistan.” The primary aim of this initiative is to support the scientific community in guiding study related to CVDs. The dataset comprises a total of 928 ECG images, which include Normal Person ECG Images (284), ECG Images of MI Patients (239), ECG Images of Patient that have abnormal heartbeat (233), and ECG Images of Patient that have History of MI (172). Furthermore, sample ECG images are depicted in Figure 7.

Figure 7:

Sample images (A) Normal person ECG images (B) ECG images of MI patients (C) ECG images of patient that have abnormal heartbeat (D) ECG images of patient that have history of MI. MI, myocardial infarction.

The proposed model’s performance is assessed on several performance metrics including accuracy, precision, recall (sensitivity), F1-score, specificity, and FNR. While all of these performance metrics are not only standard, they are highly relevant for medical diagnosis application, where not just a correctly detected case (sensitivity), but especially the ability to minimize missed cases (low FNR), is important. For example, the model attained an accuracy of 95.6%, which indicates that the overall performance is reliable. However, the model’s precision of 97.9% is of particular importance because it demonstrates that it avoided false positives – a particularly useful metric for avoiding undesirable follow-ups. The low FN of 5.1%, and sensitivity of 94.9% indicates that the model is good at detecting actual CVD cases, while minimizing a miss rate of potentially affected people, which specifically addresses a significant limitation presented by many other systems. The proposed system also demonstrated improved detection reliability to the extent of both the CNN (FNR = 19.0%) and DNN (FNR = 21.5%) failing to detect numerous likely cases of CVD, particularly among the minority classes. The values that the proposed model calculated contribute in technical detail to the manuscript by demonstrating that the model is not only accurate, but also clinically reliable and applicable in a real-world sense to ECG-based CVD screening.

Figure 8 depicts the identification of CVDs using ECG signal images, utilizing diverse filtering methods, including Improved Median, Gaussian, Mean, Wiener, and Conventional Median Filtering.

Figure 8:

Images for CVD detection using ECG signal (A) Original images, (B) Gaussian filtering, (C) Mean filtering, (D) Wiener filtering, (E) Conventional median filtering, and (F) Improved median filtering. CVD, cardiovascular disease.

The Squeeze-ICNN model represents a noteworthy progression in the domain of medical image investigation, precisely for the detection of CVDs through ECG signal images. In order to assess its effectiveness, a comprehensive analysis comparing various metrics with established models such as LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8] is imperative. One crucial aspect of this evaluation is the consideration of positive metric values, which play a pivotal role in determining the model’s ability to accurately detect instances of CVD. As we delve into the analysis, the focus will be on elucidating how the Squeeze-ICNN model outperforms its counterparts, showcasing higher positive metric values as a testament to its efficacy in accurately identifying and classifying cardiovascular abnormalities in ECG signal images. Figure 9 visually represents the results of our comparison, showing the detailed differences in performance among the various models. Here, for the training data 90%, the SqueezeNet and DenseNet achieved detection accuracies exceeding 90%, yet the Squeeze-ICNN scheme outperformed them with a remarkable 95% detection accuracy. Likewise, the Squeeze-ICNN approach consistently demonstrated superior performance, surpassing 92% detection accuracy at the 70%, 80%, and 90% training data.

Figure 9:

Comparative analysis on positive metrics for Squeeze-ICNN and traditional schemes. CNN, convolutional neural network; DCNN, deep convolutional neural network; DNN, deep neural network; ICNN, improved convolutional neural network.

The accuracy improvement with increased data availability (92%–95.6% at 90% training) is consistent with the expected learning behavior in which the model can generalize better with more labeled data. The increase in accuracy through learning is arguably more important because the model escaped from potential overfitting by applying dropout, BN, and early stopping. Overall, the variance is sufficiently low across the multiple random train-test splits (σ = 0.028), suggesting that the increase in accuracy is due to learning, and not due to bias, and the model is simply learning features from more diverse training samples. These results suggest that the performance improvement are gains in true learning and not overfitting.

Additionally, the Squeeze-ICNN scheme exhibits a sensitivity of 0.923 with a training data of 70%. By contrast, traditional schemes, such as LSTM (0.730), DCNN (0.813), Bi-GRU (0.781), SqueezeNet (0.861), DenseNet (0.827), DNN [1] (0.762), and CNN [8] (0.800), yielded lower sensitivity values, underscoring the superior sensitivity of the Squeeze-ICNN approach. Moreover, the Squeeze-ICNN scheme attains the greatest precision and specificity, with rates of 0.978 and 0.922, respectively. These values markedly exceed those attained through traditional methodologies. The significant outcomes noticed in the positive measure estimation highlight the efficacy of the Squeeze-ICNN method in identifying CVD through the analysis of ECG signal images. This success is credited to the sophisticated detection strategy and enhancements applied in the preprocessing, feature extraction, and classification procedures.

In terms of clinical cardiology, it is essential to make a correct diagnosis in order to provide adequate early management and intervention in the treatment of CVD, such as MI and arrhythmias. High sensitivity (0.923) in the Squeeze-ICNN model allows for unintended zero missed cases of those containing true abnormal cardiac condition. If a real abnormal cardiac condition was missed, there could be severe life-threatening complications when left untreated. In a similar fashion, high precision value (0.978) means the model does not have a high false alarm rate and therefore avoids wasted clinical follow-ups. More importantly, this value aligns with actual practice, since ECG chart interpretation becomes an important differentiating factor when diagnosing patients in acute care with time constraints. The promising outcomes of DL when interpreting ECG images is not new, as described in previous studies such as DeepECG by Changling Li et al. [3], which also revealed better performance detections of arrhythmias using similar DCNN-based architectures.

They simulated its performance against the standard CNN and SqueezeNet models. The results presented in Table 3 demonstrate that ICNN was able to achieve greater accuracy and sensitivity as compared with the baseline CNN. This demonstrates that the proposed additions to the architecture, in combination with the proposed activation function, were able to learn more effectively from the ECG patterns. Moreover, the Squeeze-ICNN model, which was an integration of SqueezeNet and ICNN, achieved the highest F1 score and precision in diagnostic performance across all tested models, even compared with the baseline model, SqueezeNet.

Figure 10 illustrates the assessment of negative metrics for the Squeeze-ICNN method in comparison to other models such as LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8], for the detection of CVD using ECG signal images. Lowering the negative metric value is crucial for improving the effectiveness of disease detection. Upon thorough examination of all the graphs, it is obvious that the Squeeze-ICNN method consistently achieved the least negative metric values across all training data, highlighting its superior performance in CVD detection. Primarily, with a training data of 90%, the Squeeze-ICNN method exhibited an FPR of 0.062. In comparison, LSTM, DCNN, SqueezeNet, DenseNet, DNN [1], and CNN [8] achieved the lowest FPRs of 0.095, 0.190, 0.102, 0.108, 0.142, and 0.190, respectively. Moreover, the FNR of the Squeeze-ICNN scheme spans from 0.113 to 0.039, demonstrating a significantly lower value compared with conventional strategies. Consequently, the Squeeze-ICNN scheme has exceeded the performance of previously utilized schemes, yielding more satisfactory negative metric ratings. This fascinating indication highlights its outstanding ability in accomplishing efficient detection of CVD through the analysis of ECG signal images.

Figure 10:

Comparative analysis on negative metrics for Squeeze-ICNN and traditional schemes. CNN, convolutional neural network; DCNN, deep convolutional neural network; DNN, deep neural network; ICNN, improved convolutional neural network.

The comparative assessment of the Squeeze-ICNN method is contrasted with LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8] concerning other metrics for the detection of CVD using ECG signal images, as depicted in Figure 11. Primarily, the goal is to maximize the values of the other metrics for the effective detection of CVD. In particular, at a training data of 60%, the Squeeze-ICNN scheme achieves an F-measure of 0.919, surpassing the F-measures of other models, namely LSTM (0.826), DCNN (0.807), Bi-GRU (0.844), SqueezeNet (0.871), DenseNet (0.847), DNN [1] (0.821), and CNN [8] (0.822), respectively. Furthermore, training data = 90, the Squeeze-ICNN scheme achieves MCC and NPV rates of 0.874 and 0.869, respectively. By contrast, LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8] obtained the lowest ratings for both MCC and NPV. Certainly, this serves as clear evidence that the Squeeze-ICNN methodology excels in the detection of CVD. This achievement is made possible through the implementation of an enhanced preprocessing technique utilizing median filtering, coupled with a hybrid classification method and an improved hierarchy of skeleton feature-based feature extraction.

Figure 11:

Comparative analysis on other metrics for Squeeze-ICNN and traditional schemes. CNN, convolutional neural network; DCNN, deep convolutional neural network; DNN, deep neural network; ICNN, improved convolutional neural network.

A thorough investigation was systematically conducted through ablation to estimate the effect of the Squeeze-ICNN approach. This detailed process explores a complete sympathetic of the detailed improvements these elements brought to the complete success of the Squeeze-ICNN method. Table 4 displays the ablation assessment performed on the Squeeze-ICNN framework, including the model with traditional preprocessing, the model integrating a conventional skeleton hierarchy, and the model amalgamating SqueezeNet with conventional CNN for CVD detection using ECG signal images. Furthermore, the precision of the Squeeze-ICNN approach is 0.979, whereas the model with conventional preprocessing attains a precision of 0.785, the model with a conventional hierarchy of skeleton attains 0.788, and SqueezeNet + conventional CNN demonstrates a precision of 0.836. Likewise, the FNR for Squeeze-ICNN method is 0.051, while the model with conventional preprocessing, the model with a conventional hierarchy of skeleton, and SqueezeNet + conventional CNN exhibit FNRs of 0.215, 0.212, and 0.164, correspondingly. Additionally, the proposed Squeeze-ICNN models demonstrate superior performance compared with employing models with conventional preprocessing techniques, SqueezeNet + Conventional CNN, and model with conventional hierarchy of skeleton.

To ensure the precision in results, every method undergoes a thorough statistical assessment, involving a scrupulous inspection of key statistical metrics like “Maximum, Minimum, Mean, Standard Deviation, and Median.” Table 5 describes the statistical analysis comparing the Squeeze-ICNN method with LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8] for the detection of CVD using ECG signal images. In terms of the median statistical metric, the Squeeze-ICNN scheme demonstrates an accuracy of 0.934. This outperforms conventional methods, with accuracy scores as follows: LSTM = 0.785, DCNN = 0.834, Bi-GRU = 0.797, SqueezeNet = 0.868, DenseNet = 0.829, DNN [1] = 0.808, and CNN [8] = 0.801, respectively. Moreover, when considering the maximum statistical measure, the Squeeze-ICNN system accomplished a higher accuracy value of 0.956. Meanwhile, LSTM, DCNN, Bi-GRU, SqueezeNet, DenseNet, DNN [1], and CNN [8] yielded lower accuracy ratings.

The present evaluation was conducted on a single dataset, but the model architecture is intended to generalize to new data because of the feature fusion, custom activation function, and regularization methods. The model showed consistent performance across multiple randomized train-test splits, and was low variance—indicating good generalization. However, a completely new and external dataset was not used in this study, and therefore, we are aware of this. Future work will validate the model on publicly available ECG image datasets, to assess its robustness and generalizability across different patient populations and acquisition settings.

While there are benefits to approaching the construction of model and classifier on the EPS framework, several challenges remain. To confirm, the present evaluation has relied on a singular dataset, and the need for external validation with other ECG datasets will facilitate the further understanding of utility and generalizability regarding the use of clinical ECG data. The model, while relatively lightweight when compared with deeper CNN’s, will require real-time deployment and testing on edge devices in future works. The introduction of real-world variabilities and noise types from multiple sources will further improve clinical feasibility (Table 6).