A Deep Convolutional Model for Heart Disease Prediction based on

ECG Data with Explainable AI

SREEJA M. U., SUPRIYA M. H.

Department of Electronics,

Cochin University of Science and Technology,

Kochi-22,

INDIA

researches compared to being considered as signals. The architecture follows a stacked Convolutional Neural

Network for extracting features from ECG images followed by fully connected network for classification. The

evaluation of the proposed model on customized public datasets demonstrates its ability to achieve impressive

outcomes by leveraging the characteristics of convolutional neural networks (CNNs) and supervised learning.

Similarly, Explainability in the form of interpretability has been incorporated into the framework thus ensuring

accountability of the model which is crucial in medical domain. Detailed experiments for identification of ideal

model architecture are conducted. Further, local and vision based Explainability has been explored in detail

using LIME and Grad-CAM. The model could achieve a precision, recall and f1-score of 0.982, 0.982, and

0.981 respectively proving the superiority of the model. Moreover, Explainability visualization based on

popular algorithms for true positive and false positive results have shown promising results on the PhysioNet

A heart disease prediction model can be a valuable

tool in healthcare for several reasons, the prime one

being early detection. Heart diseases, such as

coronary artery disease or heart failure, can often be

asymptomatic or exhibit subtle symptoms in the

early stages. A prediction model can analyze various

risk factors from any kind of data pinpoint

individuals who have higher chances of ending with

heart disease. Early detection allows for timely

intervention and preventive measures to reduce the

preventive advice. This optimization helps in

managing limited healthcare resources efficiently.

Heart disease prediction models can assist

healthcare providers in tailoring treatment plans for

context of the patient's overall health, and make

more accurate diagnoses and treatment decisions. It

also helps identify potential biases or errors in the

model's predictions, promoting fairness and equity

in healthcare delivery. Explainable models provide

an opportunity to educate patients about their risk

factors, lifestyle choices, and the impact on their

heart health. Patients who understand the reasons

behind their risk predictions are more likely to

actively participate in preventive measures, make

that aims at categorizing input ECG images into

healthy and other unhealthy classes. The model

further analyses the significance of Explainability

with two popular algorithms Local Interpretable

Model-agnostic Explanations (LIME) and Grad-

CAM which helps to analyze a deep learning model

and to make its predictions comprehensible. Grad-

CAM have extreme application in image

classification tasks which enables the model to

interpret black box model by visualizing the class

well as the best explainable algorithm for the

proposed problem. The main highlights of the paper

are:

1. An explicit deep convolutional model for

ECG based image prediction trained on customized

datasets on most frequently occurring categories of

heart disease in a supervised manner.

2. A comparative analysis of vision-based

Explainability models on ECG data with Grad-CAM

analysis of results in section 5 and conclusion in

section 6.

layers are then followed by a ‘batch normalization’

layer and a ‘max pooling’ layer. The feature vector

obtained from CNN is flattened and three dense

layers follows with dimension 128, 50 and 4 units

each. The purpose of the model is to learn to

classify incoming ECG signal into 4 classes, hence

the count 4 in the final layer. The architecture has

been depicted in Figure 2. The intermediate

Convolution layers follows ReLU activations as in

(4).

where the number of classes and the ith output value

are denoted as N and respectively.

Loss function: The focus of the model is to classify

electrocardiogram (ECG) images, which can be

approached as a multiclass classification task. To

measure the disparity between predicted probability

loss. The computation of the loss function follows

the formula mentioned in (3).

󰇛 󰇜   

where for N number of samples  and,  is the

predicted value and the true value respectively.

Regularization: In order to avoid overfitting several

techniques have shown best results. The proposed

obtained after this process. A dropout of 0.5

best model.

single CNN model, CNN model with dual stacked

convolution layers and with batch normalization.

Evaluation has been performed based on

quantitative metrics and the best model was chosen

as the optimal one. Results of ablation has been

reported and performance evaluation based on

classification metrics such as precision, recall, f1-

score, accuracy, ROC curve, etc. are outlined in

section 5. The model with the best results has been

chosen as the optimal one from the experiments.

For interpreting the decision of the trained model,

algorithms for Explainable AI have been

experimented. Explainability algorithms can be

divided into local and global ones. The proposed

model has experimented with both and have

individual predictions rather than the model as a

whole. When applying LIME to image

classification, the basic idea is to understand which

image parts have been influential in the model's

prediction process. LIME accomplishes this by

introducing perturbations to the input image and

then monitoring how these perturbations affect the

predictions. The major steps in LIME Explainability

are:

a. Generate perturbed instances for the chosen

images and their corresponding predictions to create

a local neighborhood. The selection can be based on

proximity to the original image or using a sampling

method.

c. Generate interpretable features: Convert the

selected perturbed instances into a format that can

be easily interpreted by human observers. For

decision.

By using Grad-CAM, insights can be gained

regarding how the decisions are made in the CNN

and identify visual cues that influenced the

predictions.

5 Results and Analysis

The model performance is analyzed in this section

The evaluation metrics considered for evaluation are

the popular quantitative metrics used in a

classification problem as shown below:

5.1.1 Accuracy

Accuracy is the ratio of total correct classifications

made by the model and is calculated as in (4).

where N is the sum of True Positives (TP), True

Negatives (TN), False Positives (FP), and False

disadvantages of precision and recall and computed

as in (7).

  

  

5.1.3 Area under Receiver Operating

The accuracy and loss plots obtained during the

training phase of the proposed deep learning model

is shown in Figure 3 where 3(a) denotes the

accuracy curves while 3(b) denotes the loss plots.

Figure shows that the loss of training and testing is

almost overlapping indicating that overfitting was

addressed properly in the proposed architecture and

that the results are consistent throughout the epochs.

Similarly, the plots also demonstrate that the model

has converged optimally.

elements show the true positive results. It can be

observed that there is a slight confusion in the

which is negligible. The results obtained for all

other entries are remarkable.

from Figure 4b, which indicates the high accuracy

of classification achieved by the model for every

classes. The high area under the ROC curve

demonstrate that the model could perfectly

elevated accuracy, precision, recall and f1-score of

0.985, 0.982, 0.982, and 0.981 owing to the dual

convolution and batch normalization incorporated.

The results also indicate that the model performs

remarkably for the input data. Recall is often seen as

a metric for quantifying performance, while

precision is typically considered a measure of the

quality of results.

High recall suggests that an algorithm retrieves a

significant portion of relevant results, even if it

performs in a balanced manner for all the classes.

Figure 5 shows the explanation generated for a true

positive and false positive prediction using LIME

algorithm. First, the model generated neighborhood

data by randomly perturbing features from the

instance. Later, the locally weighted models on this

neighborhood data are learned to explain each of the

classes in an interpretable way. The size of the

neighbourhood to learn the model is set to 1000

samples. The samples considered are from class V

as mentioned above. The illustration depicts a

contrast between the initial image from class V and

Q, the Grad-CAM-generated heatmap, and the

significance of the regions on the original image by

superimposing it with the heatmap. Grad-CAM

identifies the regions in the ECG image that

Fig. 4: Evaluation results (a) Confusion matrix for

images for the second convolution layer with LIME

and (b) False positive (sample from class Q)

B., Liang, D., ... & Chen, J, Classification of

cardiac electrical signals between patients with

myocardial infarction and normal subjects by

disease prediction by using novel optimization

Informatics in Medicine Unlocked, 26, 2021,

100696.

Classification Using Fractal Dimensions and

arrhythmia diagnosis based on the combination

features of ECG signals and metaheuristic