Analyzing the impact of transfer learning on explanation accuracy in deep learning-based ECG recognition systems

May Sadiq Khorsheed; AbdulAmir Abdullah Karim

doi:10.1515/eng-2024-0066

Artikel Open Access

Analyzing the impact of transfer learning on explanation accuracy in deep learning-based ECG recognition systems

May Sadiq Khorsheed und AbdulAmir Abdullah Karim

Veröffentlicht/Copyright: 19. September 2024

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen

Aus der Zeitschrift Open Engineering Band 14 Heft 1

Abstract

Electrocardiogram (ECG) recognition systems now play a leading role in the early detection of cardiovascular diseases. However, the explanation of judgments made by deep learning models in these systems is prominent for clinical acceptance. This article reveals the effect of transfer learning in ECG recognition systems on decision precision. This article investigated the role of transfer learning in ECG image classification using a customized convolutional neural network (CNN) with and without a VGG16 architecture. The customized CNN model with the VGG16 achieved a good test accuracy of 98.40%. Gradient-weighted class activation mapping (Grad-CAM), for this model, gave the wrong information because it focused on parts of the ECG that were not important for making decisions instead of features necessary for clinical diagnosis, like the P wave, QRS complex, and T wave. A proposed model that only used customized CNN layers and did not use transfer learning performed 99.08% on tests gave correct Grad-CAM explanations and correctly identified the influencing areas of decision-making in the ECG image. Because of these results, it seems that transfer learning might provide good performance metrics, but it might also make things harder to understand, which could make it harder for deep learning models that use ECG recognition to be reliable for diagnosis. This article concludes with a call for careful consideration when using transfer learning in the medical field, as model explanations resulting from such learning may not be appropriate when it comes to domain-specific interpretations.

Keywords: convolutional neural networks; gradient-weighted class activation mapping; VGG16; electrocardiogram; transfer learning

1 Introduction

Cardiovascular deaths are still the leading cause of death worldwide, so the development of new diagnostic techniques is an important requirement [1,2,3]. Electrocardiogram (ECG) analysis holds a central place among examining tools because it is a completely non-invasive method that provides information about the state of cardiac activity [4,5]. The application of deep learning, especially convolutional neural networks (CNNs), is the next breakthrough in the ECG interpretation field; it is more precise and more productive than traditional approaches [6]. The suggested approach was chosen because it could shed light on the importance of using explanation methods for deep learning models as well as how transfer learning affects the precision of explanations in deep learning-based ECG recognition systems. As artificial intelligence (AI) in healthcare has progressed significantly in recent years, there is a greater reliance on AI models in crucial clinical areas, including ECG analysis. However, the black-box nature of these models hinders their implementation, making it challenging to comprehend the connection between the model and the problem at hand. Also, while transfer learning is considered an effective approach that has been applied in numerous studies and yields promising results in different image recognition tasks, including ECG analysis tasks, the impact of this technique on the interpretability of models has not been sufficiently studied yet. Understanding how transfer learning influences explanation accuracy in deep learning for ECG recognition is as crucial as ever. Solving these problems is vital to ensuring that the AI systems used in the medical field are credible. For instance, imagine that a deep learning model is applied to identify arrhythmias using ECG signals. Despite its high performance in detecting abnormal rhythms, doctors and clinicians only see the final decision made by the model, not how the decision was made. This can be attributed to the lack of clarity in how the model arrived at certain predictions, meaning that there is doubt about using the model for clinical decision-making. Take another example where a deep learning model trained with transfer learning diagnoses arrhythmias perfectly. However, on closer observation, clinicians can determine that the actual reasons offered by the model are not credible based on medical evidence. This brings about doubt and would cause one to pause when trusting an AI-powered system to make life-altering decisions in the context of medical treatment, thus hinting at the larger issue of how to make AI systems trustworthy in healthcare. Therefore, we conducted the current research to tackle these specific issues and illuminate crucial aspects of the advancement of AI in the healthcare sector. To investigate how transfer learning affects the precision of explanations in deep learning-based ECG recognition systems, we selected the application of a CNN trained from scratch, without transfer learning, to a hybrid CNN with VGG16. VGG16, a state-of-the-art deep CNN architecture, has been employed in many image recognition tasks as a pre-trained feature extractor; therefore, the same approach is also used in this research. The purpose of this study was therefore to try and close this gap in the literature and to help more people get to know how various model architectures and training approaches can impact the ability of deep learning models in the identification of complex ECG signals. It is necessary to enhance future research to utilize deep learning models in medical diagnosis in a more comprehensible and trustworthy manner.

2 Background and related work

The application of CNNs and deep transfer learning techniques used in diagnosing heart diseases through ECG analysis have resulted in tremendous changes. Deep learning models have demonstrated their ability to outperform conventional diagnostic modalities, leading to a transformation in how cardiac abnormalities are detected and analyzed [7]. Salehi et al. have made significant contributions in this area by conducting an exhaustive empirical comparison of transfer learning techniques across various ECG datasets and neural structures. They highlight the significant benefits of transfer learning, which infuses the models with information from unrelated large datasets. However, as the training dataset expands, the marginal benefit of transfer learning could potentially decrease [8]. Refereed research like that of Herman et al. has their AI-based systems for reading ECG interpreted as being better than the traditional computerized means. The presented AI systems, which have processed enormous amounts of ECG data, have improved their accuracy and reliability level to such an extent that in individual cases, these systems have turned out to be more reliable and accurate in the diagnosis of ECG than even experienced cardiologists [9]. Albahri et al. specifically focus their systematic study on the reliability and clarity of artificial intelligence in medical technological applications. Their research brings to light the critical feature of building AI with precision, interpretation, and justification in clinical settings to eliminate the chances of bias and error [10]. Qiu et al. and his team focus their study on the explainability challenge, exploring the relevance of gradient-weighted class activation mapping (Grad-CAM) across various deep-learning architectures. In this study, they reported that the visualization results from Grad-CAM were highly dependent on the architecture and depth of the underlying neural network model. As such, they highlighted the importance of careful consideration when choosing a network for diagnostic tasks [11].

While ECG interpretation with AI shows much promise, the limitations of transparency and operationalization of these models become apparent in real clinical settings. The question of how the architectural decisions made influence the effectiveness of the models and their clinical relevance is pertinent and therefore must remain an area of continued research. This research will further explore these aspects, with emphasis on learning strategies and advanced explanation techniques in creating more accurate, understandable, and safer ECG recognition models.

3 Research gaps

The following are the research gaps that were filled in our study.

3.1 The lack of explanation techniques is the major challenge when using transfer learning in various applications

Many of the research papers employing transfer learning for ECG and other medical image classification have not used any explanation approaches. This exclusion creates a significant knowledge gap about how these models make decisions, which is vital for the clinical endorsement that enables trust between the general public and AI-based diagnostic systems.

3.2 Marginal use of explanatory methods

Although some research works have used explanation methods, including Grad-CAM, they have not investigated the reliability and depth of these explanations. Determining the level of explanation is critical to avoid misrepresentation of the visualized decision-making process or features of the medical images being investigated.

4 Methodology

4.1 Model architectures

We compare two distinct architectures to assess the transfer learning effect on the explanations generated by CNNs:

Customized CNN without VGG16: This baseline model’s general goal is to classify ECG images. It consists of the following layers:

Input layer: Takes inputs in 224 × 224 pixel format.
Convolutional layers: To simplify the non-linear process [12], ReLU activation follows each of the four convolutional layers with a number of filters (128, 256, 512, and 512).
Global average pooling: Placed after the customized convolutional layers to lower the feature dimensions and help with classification.
Dense layers: There are two dense layers: the first with 512 units and the second with four units, each corresponding to a class of cardiac condition. The first layer uses ReLU activation, while the second uses the Softmax function to output probabilities for each class.

For a detailed visualization of this model’s configuration, refer to Figure 1.

Figure 1

The proposed CNN architecture.

This figure illustrates the architecture of the customized CNN model without using VGG16 as one of its layers. The architecture is sequential in terms of shape, from the input layers to the dense layers. Variations in the architecture are as follows: the first convolutional layer has various filter sizes, which are essential in the feature extraction of ECG signals for classification.

Hybrid CNN with VGG16: This model leverages the robust feature extraction capabilities of the VGG16 architecture, pre-trained with the ImageNet dataset [13]. It consists of the following layers:

Input layer: Takes inputs in 224 × 224 pixel format.
Pre-trained VGG16 base: During the initial stage of training, all layers will remain frozen to preserve the learned features, starting from the model’s base to the “block5_conv3” layer.
Customized top layers: We position four new convolutional layers (128, 256, 512, and 512 filters) at the output of the last retained VGG16 layer to better define the specific features of ECG images. Each of these layers undergoes subsequent ReLU activations.
Global average pooling: Placed after the customized convolutional layers to lower the feature dimensions and help with classification.
Dense layers: There are two dense layers: the first with 512 units and the second with four units, each corresponding to a class of cardiac condition. The first layer uses ReLU activation, while the second uses the Softmax function to output probabilities for each class.
Trainable layers: At first, only the layers after VGG16 are trainable. Later stages unfreeze the layers from VGG16’s “block5” to enable fine-tuning on the ECG dataset.

To understand VGG16 integration, refer to Figure 2.

Figure 2

The hybrid CNN with VGG16 architecture.

This figure represents the structure of the hybrid CNN model, where the VGG16 base model is incorporated with extra customized layers. The shape is a combination of the VGG16 base model and customized layers, and variations include the base VGG16 (freeze layers) and adding more layers with 128/256/512 filters, respectively. This design makes use of the pre-trained features while at the same time learning other features specific to ECG.

4.2 Implementation of Grad-CAM for visual explanations

This study chose Grad-CAM to generate visual explanations of a deep learning model’s decisions [14]. It supports the identification of the regions in the ECG images that models rely on to make their predictions, thus giving useful information about model understandability.

The Grad-CAM implementation includes the following processes:

Layer selection: We used Grad-CAM on the output of each model’s final convolutional layer to identify the most spatial features that determine classification.
Gradient calculation: We used backpropagation to compute the gradients of the target class.
Weighted activation map: By multiplying the feature maps by the established set of weights, we were able to create a localization map with the most features that discriminate between classes.

We then scaled the generated class activation maps to the original ECG images’ size and transformed them into heatmaps. We superimposed the heatmaps over the initial ECG images, effectively adding a visual layer that highlighted the regions the model most distinguished in its predictions. This, in turn, led to a clearer understanding of what the models were looking at, indicating how they came to their conclusions.

4.3 The relation of proposed method parameters to system parameters

The parameters used in the proposed method in this study are carefully selected to correspond with the system parameters of the ECG recognition task, such that there are no discrepancies between the general architecture of the model and the nature of ECG data. The customized CNN model comprises the following architecture: input layer: ECG images are fed into the model after resizing the images to a standard size of 224 × 224 pixels. The convolutional layers have a ReLU activation function, which serves the purpose of feature extraction of important details of ECG signals that include edges and other patterns associated with cardiac conditions such as P waves, QRS complexes, and T waves. The global average pooling layers are applied to avoid overfitting and to summarize the spatial features of the input since the networks will generally be very deep; the dimensions of the feature maps are reduced to provide only the most important features. The dense layers with Softmax activation subsequently evaluate the extracted ECG image features and categorize them into four apparent types of cardiac diseases, namely history of myocardial infarctions (HMI), myocardial infarctions (MI), abnormal heartbeat (AHB), and normal.

In the proposed hybrid CNN model, the pre-trained base that is incorporated in the hybrid CNN model or the CNN architecture is VGG16, which has a strong feature extraction ability from the ImageNet dataset. However, due to the differences in features between ImageNet and ECG, some extra layers are incorporated to optimize the feature extraction of the model for ECG images. The training stages include: first, freezing the VGG16 layers to maintain pre-trained weights and, then, partially, unfreezing some layers to fine-tune the ECG dataset, which enhances the model’s ability to capture specific features of ECG.

The Grad-CAM implementation further improves the interpretability of the model. When Grad-CAM is applied using the last convolutional layers, the method shows which areas of ECG images are crucial for the model’s predictions, thus explaining the decisions made. Backpropagation is applied for calculating gradients, and a weighted activation map helps make the results visible so that one can see which part of the picture has the biggest influence on the model.

4.4 Model evaluation and validation

4.4.1 Quantitative evaluation

Standard performance metrics, including precision, recall, F1 score, precision, and specificity, evaluated both models: the customized CNN model without VGG16 and the hybrid CNN model with VGG16. We calculated these measures for each of the four target categories: HMI, MI, AHB, and normal.

4.4.2 Qualitative validation

A practicing cardiologist evaluated the Grad-CAM output of the customized CNN model without VGG16 to see if it paid attention to diagnostically relevant parameters such as P waves, QRS complexes, and T waves. The cardiologist commented that the model’s visualizations effectively pointed out regions in agreement with clinical knowledge, housing the decision-making logic. This means that the model can help to highlight the most important elements in the ECG data.

5 Experiment design

5.1 Dataset

The present study employed a standard ECG dataset that was made available in the public domain and consists of 1,937 ECG samples annotated for analysis and categorized into four classes: MI, AHB, HMI, and Normal [15]. A total of 77 images describe situations in which patients have MI, which is a sign of serious coronary diseases that might lead to heart attacks and even death. Through the collection of 548 images, the AHB disease category demonstrates people struggling with breathlessness or impaired breathing as a consequence of cardiac diseases. The HMI includes 203 patient images from previous MI. Lastly, normal is the largest category, which includes 859 images from healthy individuals. The dataset comes from medical devices in the EDAN series, with a standard 12 leads and 500 Hz sampling rate. The original dataset included the COVID-19 class, which this research excluded. Normal (548 ECG), MI (548 ECG), HMI (548 ECG), and AHB (548 ECG) images are used in the categorization criteria. This value was chosen to adjust the imbalance in the sample and avoid the inefficiency of classifications that might result.

5.2 Data preprocessing and augmentation

For the preparation of ECG images for analysis, we first do a static crop, selecting only 6–96% of the horizontal region and 21–93% of the vertical region. By performing this step, we eliminate unnecessary elements from the images and retain only those areas that contain the expected diagnostic information. We resize all the images right after cropping and standardize them to 224 × 224 pixels, based on the input into our CNN. Moreover, we adjust the brightness by approximately 5% to enhance the model’s resilience against contrast-induced fluctuations in image lighting.

5.3 Model training

Train-validation-test split: We first split the data into training and testing sets using an 80/20 split. We also created the validation subset within the training set, setting its amount to 20%.
Optimizer and loss function: We compiled them using the Adam optimizer and the categorical cross-entropy loss function.
Training process: Customized CNN without VGG16: We trained the customized CNN for 200 epochs, using early stopping to prevent overfitting and reducing the learning rates if the validation loss plateaued. Hybrid CNN with VGG16: To easily adapt to the new top layers, the VGG16-based model initially trains with the frozen VGG16 layers. In the following step, we unfroze and refined the VGG16-specific layers to ensure fine adaptability.
Validation method: Both models used a validation set held out and performance metrics recorded at each epoch to monitor the learning process during iterations as well as guide adjustments.

6 Results

6.1 The performance of customized CNN model without VGG16

We evaluated the customized CNN model without VGG16’s ability to correctly classify ECG images into the desired four categories. Table 1 displays the performance results.

Table 1

The performance results of the customized CNN model without VGG16

	Precision (%)	Recall (%)	F1 score (%)	Accuracy (%)	Specificity (%)
AHB	98.32	98.32	98.32	99.09	99.38
HMI	100	100	100	100	100
MI	100	100	100	100	100
Normal	98.20	98.20	98.20	99.09	99.39

The visualization of Grad-CAM for the customized CNN model without VGG16 (referred to Figure 3a) confirmed that the model successfully highlighted features necessary for clinical diagnosis, like P waves, QRS complexes, and T waves. The visualizations showed the model’s ability to localize precisely on the foundational segments of the image that aimed at providing clinical interpretation to guide its decision-making process.

Figure 3

Grad-CAM visualizations. (a) Grad-CAM visualization for the customized CNN model. The shape of the highlighted regions in the ECG images aims to point at diagnostically significant areas like P-waves, QRS complexes, and T-waves, as well as offer correct visual explanations that match clinical experiences. (b) Grad-CAM visualization for the hybrid CNN model with VGG16. The highlighted regions in the ECG image show that the shape of the regions often emphasizes unimportant areas, which shows that transfer learning adversely affects explanation accuracy. (c) Grad-CAM visualization for the hybrid CNN model with VGG19. The highlighted regions in the ECG image show that the shape of the regions often emphasizes unimportant areas, which shows that transfer learning adversely affects explanation accuracy.

6.2 Hybrid CNN with VGG16 model performance

We also evaluated a hybrid CNN model incorporating VGG16, and Table 2 presents the corresponding results. When compared with the customized CNN, the hybrid model’s Grad-CAM representation (Figure 3b) was clearly out of place. The model pointed out the areas that were not pertinent for clinical interpretation, showing that VGG16 pre-trained features may not have the specific fixings required for the ECG classification.

Table 2

The performance results of the hybrid CNN with VGG16

	Precision (%)	Recall (%)	F1 score (%)	Accuracy (%)	Specificity (%)
AHB	95.16	99.16	97.12	98.41	98.12
HMI	100	100	100	100	100
MI	100	99.06	99.53	99.77	100
Normal	99.07	95.50	97.25	98.63	99.70

6.3 Comparison and observations

Customized CNN Model without VGG16: The customized CNN model without VGG16 was successful because of its high ability to spot diagnostically relevant features in the ECG images. This success is reflected in the high-performance metrics achieved (Table 1). The Grad-CAM visualizations showcased the model’s ability to identify regions significantly affected by cardiac conditions, indicating its performance aligns with expectations.

Hybrid CNN with VGG16: The model displayed some regions on Grad-CAM visualizations that do not contain clinically recognized features used in an ECG interpretation. The model performed as well as the customized one, but the Grad-CAM scores did not always reflect a deep comprehension of the data, which means that transfer learning did not necessarily make the model more interpretable in this context.

There is an indication thus that transfer learning can bring certain advantages into play sometimes, but individually designed and trained models for ECG classification, such as customized CNN without VGG16, can be superior in interpretability by focusing on specific Grad-CAM visualizations.

7 Robustness evaluation

Further experiments were conducted as a measure of testing the robustness of the developed ECG recognition system; the experiments included data augmentation and training the model using a different pre-trained model (VGG19).

7.1 Data augmentation

To simulate real-world variations, data augmentation strategies were employed on the original ECG dataset. In particular, the value of 1.3 pixels of blur and changes in exposure from −10 to +10% were applied. These augmentations assist in evaluating how the models perform in context to seemingly different forms of the ECG image.

7.2 Additional pre-trained model (VGG19)

Besides the customized CNN and hybrid CNN with VGG16, the experiment included a hybrid CNN with VGG19. Similar to VGG16, the VGG19 model was also pre-trained on the ImageNet dataset and later adopted for the ECG classification. Table 3 below displays the performance results.

Table 3

The performance results of the customized CNN model without VGG19

	Precision (%)	Recall (%)	F1 score (%)	Accuracy (%)	Specificity (%)
AHB	100	91.60	95.61	97.72	100
HMI	94.87	100	97.37	98.63	98.17
MI	96.26	100	98.10	99.09	98.81
Normal	97.17	97.17	97.17	98.63	99.10

7.3 Performance metrics with augmented data

We used data augmentation approaches to test the robustness of the models. This was done by evaluating the models on the augmented dataset to test for generalization with slightly different ECG images. Table 4 presents the obtained test results.

Table 4

Performance metrics with augmented data

Model	Accuracy (%)	Precision (%)	Recall (%)	F1 score (%)	Specificity (%)
Customized CNN	99.24	99.25	99.22	99.23	99.74
Hybrid CNN with VGG16	96.64	96.70	96.63	96.62	98.88
Hybrid CNN with VGG19	90.77	91.49	90.98	90.84	96.95

It is clear from the performance metrics that although the customized CNN had much higher accuracy, precision, recall, F1 score, and specificity, the hybrid models consisting of VGG16 and VGG19 are also acceptable but not as good as the customized CNN.

7.4 Grad-CAM visualizations with VGG19

We also created Grad-CAM visualizations for the hybrid CNN with VGG19 to identify which regions of the ECG images affected the model’s prediction. Figure 3c displays the Grad-CAM results for the hybrid CNN with VGG19.

Similar to the VGG16 model, the Grad-CAM output of the VGG19 model also displayed falsely highlighted regions.

8 Selectivity of the proposed method for practical systems

The proposed method is based on a comparison between a customized CNN trained without transfer learning and a hybrid CNN with transfer learning. We conducted this comparison to assess the impact of transfer learning on the interpretability of ECG recognition systems. We conclude that the use of transfer learning reduces the interpretability based on the presented Grad-CAM display results. For practical systems and user-specific applications, this method is beneficial for the following reasons:

Improved interpretability without transfer learning: The customized CNN without transfer learning gives better and more explicit visualization of the explanations using Grad-CAM as illustrated in Figure 3(a). This is particularly helpful for more clinical applications where understanding the decision-making process behind a model significantly enhances trust and the overall functionality of the model.
Clinical relevance and trust: Interpretation is a matter of clarity, and accurate interpretability translates to clinical trust. Thus, when the explanations line up with existing clinical cues, clinicians are more likely to use the AI system. The customized CNN’s reliable explanations make it a better candidate for practical application in the real world, hospitals included.
Enhanced model performance: Our investigation shows that the customized CNN is more appropriate for ECG analysis than the hybrid CNN which uses transfer learning. This makes it especially effective when used to address real-world cases in ECG recognition to improve the reliability of the system.

9 Discussion

9.1 Addressing the interpretability issue

The main objective of this article was to investigate how the use of transfer learning affects the interpretability of deep learning-based ECG recognition systems. For models that are implemented in a clinical context, not only the output of the model is important but also the way the model arrives at those decisions. This creates trust and prepares the groundwork for implementing and adopting AI technologies in clinical practice.

9.2 Interpretation of findings

The study’s findings indicate that the customized CNN without transfer learning had improved accuracy and relevant clinical decisions compared to the customized CNN with transfer learning. One of the reasons that the explanations are not accurate for transfer learning with VGG16 and VGG19 could be the variation between the features of the image learned from ImageNet and those of ECG interpretation. VGG16 and VGG19 were pre-trained for images from nature, but ECG signals differ much as regards structure, shapes, and their arrangements. As a result, the features from VGG16 and VGG19 that are captured may not be the best for the ECG image classification, and consequently, the visualization of Grad-CAM might not be very accurate.

The results of the data-augmentation experiments further supported the robustness of the customized CNN model. While we noticed a decrease in the performance of the proposed hybrid models when tested on the new augmented data, the CNN model developed specifically for our purpose was much more robust and explainable. This means that the customized CNN without transfer learning is relatively accurate in identifying the variations that may exist in real ECG images, as opposed to the hybrid models based on transfer learning.

9.3 Implications for clinical applications

The capability for AI models to display outputs that are understandable to end users is of importance for their usage in clinical environments. Correct descriptions that have been illustrated by the customized CNN model can improve clinicians’ trust and help them understand the basis of model predictions. As a result, it may support making informed medical choices. While the hybrid model’s Grad-CAM outputs showed that transfer learning is not generally the optimal choice for interpretability in domains like ECG analysis, it might be the case that it is leading to worse interpretability. This finding emphasizes the necessity of tailoring the model architecture to suit the application’s specific domain of interest to reach a balance between high accuracy and meaningful interpretability.

9.4 Solving the problem

The proposed method has successfully dealt with the issue of interpretability in the following ways:

Enhanced Transparency: The proposed customized CNN model without transfer learning is more accurate and gives clinically useful explanations, which are important for the trust of clinicians.

Reduced Misalignment: Here, the study notes that while pre-trained models such as VGG16 can be employed in medical imaging tasks, they come with several shortcomings that make it important for bespoke medical domain models to be developed.

Improved Trust and Adoption: Hence, it can be said that precise interpretability has a positive influence on the level of clinically applied artificial intelligence trusted by clinicians. With its cleaner and more reliable explanations, the customized CNN model can be deployed in practical health engagements.

9.5 Limitations and confounding factors

This research has some shortcomings that should be taken into account.

Dataset size and diversity: The dataset presented has good coverage, but it might not cover all possible different types of ECG patterns. Therefore, this may influence the model’s generalizability and the relationship between the inputs and the Grad-CAMs produced.

Model architecture: The model complexity can give deeper insights into the data, but at the same time, the architecture of the models could inherently limit their interpretability. The customized model is tailor-made for this case, while VGG16 and VGG19 are not, suggesting that it may have some negative impact on the performance of the hybrid model.

New research needs to be conducted on various architectures of models and data augmentation techniques to improve the interpretability and accuracy of the model in the domain of clinical operation.

10 Conclusion

This research aims to fill the gaps identified in the existing literature by studying the effects of transfer learning on explanation accuracy in deep learning-based ECG recognition systems. The findings could potentially enhance the accuracy of disease identification through artificial intelligence, benefiting researchers in AI and diagnostics, physicians, and developers of diagnostic programs and applications. This research serves as a guide for future studies and developmental work, providing insights into the impact of transfer learning on the interpretability of models. The contributions are long-term, emphasizing AI systems’ explainability to support their implementation and trust in healthcare. Moreover, other subfields of medical image analysis and diagnostic AI can apply the results, offering a broader perspective on explanation techniques across various domains. Since this work focuses on improving the reliability of explanations used in AI models with transfer learning, it provides useful information to researchers, practitioners, and the healthcare sector. Specifically, it fosters the creation of reliable AI systems for healthcare, which will benefit both healthcare providers and patients.

In this study, we compare the performance and interpretability dimensions of a custom-designed CNN and a hybrid CNN with VGG16 to classify ECG images. The customized CNN model without VGG16 showed a more robust performance, along with a more accurate explanatory power indicating diagnostically significant areas in the ECG signals, in distinction from the hybrid model with VGG16, which justified these regions more often with features of less value despite achieving comparable quantitative assessments.

The results demonstrate that the VGG16 network’s learned features from the ImageNet dataset do not effectively extract features from ECG images. The ImageNet dataset primarily includes images of everyday objects, animals, scenes, and so on but excludes medical images like ECGs. Therefore, the weights acquired during transfer learning were less effective compared to those obtained by a custom-made CNN model, mainly in terms of interpretability. This emphasizes the necessity to develop individual models that take into account the specific features of medical data (ECG signals, for example) so that the achieved results have high diagnostic meaning.

Future research will focus on making the models more suitable for ECG analysis and adding domain competence features to the design. Further work on improving the model’s interpretability is needed. Yet, also, the growth of diversity and size of input increases the accuracy of predictions and boosts the reliability of determinations. Through complementing these revelations, future development of ECG recognition systems can have the power to achieve high accuracy as well as robust interpretability, which can then improve their use in clinical practice.

Funding information: Authors state no funding involved.
Author contributions: All authors have accepted responsibilityfor the entire content of this manuscript and consented to its submission to the journal, reviewed all the results and approved the final version of the manuscript. Methodology, MSK and AAK; software, MSK; validation, MSK and AAK; formal analysis, MSK; investigation, MSK; resources, MSK; data curation, MSK; writing – original draft preparation, MSK; writing – review and editing, AAK; visualization, MSK; supervision, AAK; project administration, AAK.
Conflict of interest: The authors state no conflict of interest.
Data availability statement: Most datasets generated and analyzed in this study are comprised in this submitted manuscript. The other datasets are available on reasonable request from the corresponding author with the attached information.

References

[1] Lorenzini G, Conti A. Numerical transient state analysis of partly obstructed haemodynamics using FSI approach. Open Eng. 2013;3(2):285–305. 10.2478/s13531-012-0052-y.Suche in Google Scholar

[2] Abbas AR. Prediction and classification of cardiovascular diseases (CVDs) using averaged one-dependence estimators (AODE) classifier. J AL-Turath Univ Coll. 2017;23(1):216–29.Suche in Google Scholar

[3] Tripathi PM, Kumar A, Komaragiri R, Kumar M. A review on computational methods for denoising and detecting ECG signals to detect cardiovascular diseases. Arch Comput Methods Eng. 2022 May;29(3):1875–914.10.1007/s11831-021-09642-2Suche in Google Scholar

[4] Noor A. Potential of cognitive computing and cognitive systems. Open Eng. 2014;5(1):75–88. 10.1515/eng-2015-0008.Suche in Google Scholar

[5] Sadiq A, Shukr N. Classification of cardiac arrhythmia using ID3 classifier based on wavelet transform. Iraqi J Sci. 2013;54(4Appendix):1167–75.Suche in Google Scholar

[6] Ali AS, Abdulmunem M. Image classification with deep convolutional neural network using tensorflow and transfer of learning. J Coll Educ Women. 2020 Jun;31(2):156–71.10.36231/coedw/vol31no2.9Suche in Google Scholar

[7] Cao M, Zhao T, Li Y, Zhang W, Benharash P, Ramezani R. ECG Heartbeat classification using deep transfer learning with Convolutional Neural Network and STFT technique. In Journal of Physics: Conference Series. Vol. 2547, No. 1. IOP Publishing; 2023 Jul. p. 012031.10.1088/1742-6596/2547/1/012031Suche in Google Scholar

[8] Salehi AW, Khan S, Gupta G, Alabduallah BI, Almjally A, Alsolai H, et al. A study of CNN and transfer learning in medical imaging: Advantages, challenges, future scope. Sustainability. 2023 Mar;15(7):5930.10.3390/su15075930Suche in Google Scholar

[9] Herman R, Demolder A, Vavrik B, Martonak M, Boza V, Kresnakova V, et al. Validation of an automated artificial intelligence system for 12‑lead ECG interpretation. J Electrocardiol. 2024 Jan;82:147–54.10.1016/j.jelectrocard.2023.12.009Suche in Google Scholar PubMed

[10] Albahri AS, Duhaim AM, Fadhel MA, Alnoor A, Baqer NS, Alzubaidi L, et al. A systematic review of trustworthy and explainable artificial intelligence in healthcare: Assessment of quality, bias risk, and data fusion. Inf Fusion. 2023 Aug;96:156–91.10.1016/j.inffus.2023.03.008Suche in Google Scholar

[11] Qiu Z, Rivaz H, Xiao Y. Is visual explanation with Grad-CAM more reliable for deeper neural networks? A case study with automatic pneumothorax diagnosis. In International workshop on machine learning in medical imaging. Cham: Springer Nature Switzerland; 2023 Oct. p. 224–33.10.1007/978-3-031-45676-3_23Suche in Google Scholar

[12] Ahmed WS. The impact of filter size and number of filters on classification accuracy in CNN. In 2020 International Conference on Computer Science and Software Engineering (CSASE). IEEE; 2020 Apr. p. 88–93.10.1109/CSASE48920.2020.9142089Suche in Google Scholar

[13] Abdulhadi MT, Abbas AR. Human action behavior recognition in still images with proposed frames selection using transfer learning. Int J Online Biomed Eng. 2023;19(6):47–65.10.3991/ijoe.v19i06.38463Suche in Google Scholar

[14] Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision; 2017. p. 618–26.10.1109/ICCV.2017.74Suche in Google Scholar

[15] Khan AH, Hussain M, Malik MK. ECG images dataset of cardiac and COVID-19 patients. Data Brief. 2021 Feb;34:106762.10.1016/j.dib.2021.106762Suche in Google Scholar PubMed PubMed Central

Received: 2024-05-15

Revised: 2024-06-24

Accepted: 2024-07-05

Published Online: 2024-09-19

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/eng-2024-0066

Schlagwörter für diesen Artikel

convolutional neural networks; gradient-weighted class activation mapping; VGG16; electrocardiogram; transfer learning

Creative Commons

BY 4.0