Machine learning driven early prediction of cardiac arrest

Abstract

Background

Cardiac Arrest (CA) is a major cause of mortality globally, often occurring suddenly without prior warning, making early detection and timely intervention crucial to saving lives. Traditional methods of predicting CA have proven inadequate due to the lack of clear warning signs. With the integration of Machine Learning (ML) techniques, the potential for more accurate early detection and intervention can improve survival rates.

Objective

This study proposes a machine learning-based approach for the early prediction of Cardiac Vascular Disease (CVD), which is a primary contributor to CA. The model incorporates various patient data, including lab results, vital signs, and Electrocardiogram (ECG) signal readings, to enhance prediction accuracy.

Methods

The study employs a range of advanced machine learning techniques, including Gradient-Boosting Algorithm (GBA), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANN). To process the data, Wavelet Transform (WT) is used to decompose the ECG signals, isolating important features while minimizing noise. Feature selection is performed through an innovative Modified Recursive Feature Elimination (MRFE) technique.

Results

The machine learning models were validated using the MATLAB simulator, with evaluation metrics including accuracy, precision, recall, and F-score. Among the models, ANN demonstrated the highest performance, achieving 96.3% accuracy, 96.1% precision, 95% recall, and 94.65% F-score.

Conclusion

This work demonstrates the effectiveness of machine learning in the early prediction of CA, enabling timely medical intervention and potentially saving lives. The results suggest that the proposed model could become a valuable tool for healthcare professionals in managing and preventing cardiac arrest.

Keywords

Cardiac arrest cardiac vascular disease machine learning wavelet transform modified recursive feature elimination

1. Introduction

Cardiac Arrest (CA) is a sudden and fatal medical emergency that occurs when the heart unexpectedly stops functioning. It is one of the leading causes of death globally, with a notably high mortality rate due to the lack of visible or measurable symptoms preceding its onset. The unpredictable nature of CA makes early detection and intervention extremely challenging. Typically, patients who suffer from CA have underlying cardiovascular conditions, and the only effective way to prevent death is through early prediction, which can guide timely medical intervention.¹

Traditionally, CA prediction has relied on clinical expertise and standardized risk scores. However, these conventional methods often fall short, particularly when it comes to detecting subtle and complex patterns that might indicate an impending cardiac event. The inherent limitations of traditional diagnostic methods make them less effective in identifying patients at risk, especially in the absence of clear clinical symptoms.²

The advent of Machine Learning (ML) offers a promising solution to enhance the prediction of CA. ML techniques can process and analyze vast amounts of patient data—such as lab results, vital signs, and Electrocardiogram (ECG) readings—by identifying hidden patterns and correlations that might otherwise remain undetected. ML's ability to handle high-dimensional, heterogeneous datasets enables a more accurate and nuanced understanding of an individual's health status, which can significantly improve prediction accuracy for CA and other cardiovascular diseases (CVDs).³

Machine Learning techniques like Gradient-Boosting Algorithm (GBA), Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANN) have shown remarkable success in medical applications. These methods are particularly effective in capturing complex, non-linear interactions within the data and yielding highly accurate predictions. While ANN has demonstrated exceptional predictive capabilities, its “black-box” nature can make model interpretability a challenge. This paper addresses this challenge by not only focusing on prediction accuracy but also emphasizing interpretability through innovative feature selection and pre-processing techniques.⁴

In this study, we focus on improving the early detection of Cardiac Vascular Disease (CVD)—a primary precursor to cardiac arrest—by leveraging machine learning algorithms in conjunction with Wavelet Transform (WT) and Modified Recursive Feature Elimination (MRFE). The WT technique helps in noise reduction and feature extraction from ECG signals by decomposing them into various frequency components. MRFE, on the other hand, selects the most relevant features by iteratively eliminating less important ones, thereby optimizing the model's performance.⁵

Through this approach, we aim to significantly enhance the accuracy of CA prediction, leading to better patient outcomes through proactive care. The findings presented in this paper demonstrate that machine learning can serve as an indispensable tool in the fight against cardiovascular diseases, enabling earlier and more accurate predictions, ultimately saving lives. The main contributions of the work are mentioned below,

➢
The dataset is gathered from the publicly available Kaggle dataset and pre-processed using the Wavelet Transform for considering the overfitting issues the feature selection process is stated and for that MRFE technique is used. Which effectively removes the unwanted features and only selects the features needed for the prediction. The interpretability is analyzed with the proper handling of data.
➢
Machine learning techniques such as GBA, RFA, SVM, and ANN are chosen based on the data handling process and the correlation of the nonlinearity of the features. The complex details provided by them are also analyzed. The ANN uses 10-fold cross-validation which enhances the prediction accuracy and interpretability.
➢
Simulation results analyzed various parameters including, accuracy, precision, recall, F-score, error, computational complexity, and cost. This shows that the ANN is better at analysing the CA prediction rather than other techniques.

The rest of the work is arranged as the relevant works are analysed and listed in Section 2. The proposed work is stated in Section 3. The simulation results are presented in Section 4. The work concluded in Section 5.
2. Related works

Bajpai et al.⁶ presented Decision Tree (DT) and SVM models for early CA prediction. There are 21 features present in the health indicator dataset and hybrid ML models like DT and SVM for prediction. An accuracy of 91.56% was accomplished and the experimental results indicated better outputs than other comparative methods. The prediction of CA utilized better findings during the experimental results with the larger scale that enhanced the CA prediction accuracy and the feature extraction techniques had better potential that appeared to result in higher prediction results.

The Light Gradient Boosting Machine (LGBM) algorithm was suggested by Lee et al.⁷ for heart rate variability in ICU in-hospital CA prediction. Due to the measures of Heart Rate Variability (HRV), the ECG with LGBM validates and predicts the CA in a real-time model. From ICU patients, the nonlinear measures with frequency and time domains included in HRV measures in which the ECG signals from 5 min epochs were calculated. Within 0.5–24 h, the in-hospital CA was predicted via the LGBM algorithm. The experimental finding provided 0.881 for ROC and 0.104 precision-recall curve performances. This method failed to remove the noise from HRV data and complexity for clinicians.

Nasiruddin et al.⁸ presented the Hybrid ML (HML) called Neural Networks, Logistic Regression, SVM, and RF to predict the survival of heart failure. The ST depression, maximum heart rate, and age are the key factors included in the predictors. For clinical practice, the potential and quick results offered significant implications. Because of timely interventions, the patient outcomes were potentially upgraded to develop heart disease risk stratification and early detection enhancement. Due to healthcare settings, the flexibility for implementation was provided via various model architectures. Overall experimental findings proved 94.97% ROC-AUC, 90.30% F1-score, 90.85% recall, 89.76% precision, and 88.41% accuracy. It met the challenges of overfitting, data quality availability, and potential discrimination.

For CA monitoring, hybrid computing with an ML-based Smart Wearable System (ML-SWS) was suggested by Hannan et al.⁹ Due to homage patients, CA in the preliminary stage was detected and it reflected the minimal latency with better response time. Assembled the sensors on the patient's body and required the current heart status for monitoring. The classification model called RF, SVM, and AdaBoost were developed thereby evaluating the performance of response time, rate of error, and accuracy. To monitor the CA risk, the research findings delivered on patients diagnosed results in superiority. Without impeding the everyday life of the person, sudden heart failure was prevented and the heart health risk was remotely monitored. This model had limited contextual data and generalization issues.

Hassan et al.¹⁰ presented an Extreme Gradient Boosting (XGBoost) model for coronary artery disease effective prediction. The diverse ML employed and heart failure linked to pivotal attributes identification. Heart failure is associated with the mortality estimated and it predicts the occurrence. The correlation analysis and feature scaling were the pre-processed methodologies for constructing the comprehensive model. The two distinct datasets selected and the feature relevance evaluated with XGBoost. Within these datasets, the occurrence of death predicted and enhanced model performance that fine-tuned the hyperparameters with 86.36% accuracy obtained. For the whole dataset, the Random Forest achieved 85.23% of the highest accuracy. The limitations of this model were Small Sample Size and limited real-time applicability.

For the prediction of congestive heart failure, Singh et al.¹¹ presented an integrated ML called K-Nearest Neighbor (KNN). The KNN integration performed unique pre-processing depending on the dataset of the Cardiovascular Health Study (CHS). An outlier data was removed the significant features were determined and the data imputation missing were employed with KNN. The methodologies of decision tree, SVM, RF, NB, and logistic regression were the comparative methodologies for the determination of experimental findings. The false positive rate, Matthew's correlation coefficient, precision, F1-score, sensitivity, specificity, and accuracy are the seven statistical parameters that evaluated the performances. More than 95% of overall statistical parameter results were accomplished but, limited data availability feature redundancy, and irrelevance were the limitations.

Muneeb et al.¹² analyzed the risks of how the patients having TB might survive COVID-19 and predicted 1.476 times higher than the normal patients. Meanwhile, the recovery rate is lower at 0.677 times, and most of the TB patients immediately. These are the findings of the authors and paved the way for the prediction of other diseases. Hassan et al.¹³ stated a study related to COVID-19 among the people of Pakistan from the Punjab. The survival and morality were plotted using the Kaplan-Meier curves.

2.1. Research gap identification

This section addresses the research gap identification from the early prediction of CA with the usage of ML algorithms. The critically existing literature contextualized and analyzed is more essential. Several previous studies relied on retrospective and static datasets and never reflected patient conditions in real-time and dynamic. From the ICU system, the continuous monitoring data for enabling early warning systems the lack of utilizing real-time. The clinical applicability and generalizability models were reduced with multi-institutional datasets with limited validation. Due to high-risk cases, the previous research met the challenges of poor sensitivity with biased models led and class imbalance effectively addressed. Because of high-stakes settings like the ICU, the ML model adoption and hinders clinician trust interpretability were limited. In the case of real-time use, computationally intensive and developed offline for most models. Limited scalability with the lack of ethical concerns and data privacy for the settings of real-world healthcare. The prognostic power over time was limited and the clinical conditions evolving were not adapted to a static model. The clinical workflow enhancement and optimizing models were the limited research. The proposed method suggested the ML model for improving clinical integration with generalizability, model interpretability, and prediction accuracy enhancements by addressing these research gaps identified.

3. Materials and methods

The proposed CA prediction is a crucial process since there might be lower signs of CA before it occurs. This CA also relies on the age of the patients. This section is to present the materials and methods of the proposed approach. The details are shown in the following subsection.

3.1. Materials

The dataset used for the early prediction of CA is taken from the Kaggle Website (https://www.kaggle.com/datasets/rahulgupta21/cardiac-risk-prediction). The reason for selecting this dataset is to predict the CA earlier and help in prevention, it also provides the details about risk levels by monitoring the target and supports the healthcare professionals to make critical decisions. CA is the most common disease that causes death and can occur due to a number of factors. The factors or features that are considered while predicting the CA are listed in Table 1. Table 1 will provide insights into the risk of CA among patients with a high possibility. The dataset is distributed in a ratio of 80:20.

Table 1.
Features that are considered for the CA prediction from the dataset.

S.No Features/factors Data type

1 Gender Integer

2 Smoker Integer

3 Age Integer

4 Tobacco Consumer Integer

5 High BP Integer

6 Diabetes Integer

7 BMI Integer

8 High Cholesterol Integer

9 Metabolic Syndrome Integer

10 Family History Integer

11 Respiratory illness Integer

12 History of Preeclampsia Integer

13 Use of stimulant drugs Integer

14 Income Integer

15 Education Integer

16 Physical Activity Integer

17 BMI Integer

S.No	Features/factors	Data type
1	Gender	Integer
2	Smoker	Integer
3	Age	Integer
4	Tobacco Consumer	Integer
5	High BP	Integer
6	Diabetes	Integer
7	BMI	Integer
8	High Cholesterol	Integer
9	Metabolic Syndrome	Integer
10	Family History	Integer
11	Respiratory illness	Integer
12	History of Preeclampsia	Integer
13	Use of stimulant drugs	Integer
14	Income	Integer
15	Education	Integer
16	Physical Activity	Integer
17	BMI	Integer

3.2. Methodology

This section presents the methodology for how to predict the CA earlier to avoid the loss of lives. For this, the collected data are pre-processed using the Wavelet transform and the MRFE is sued for eliminating the irrelevant features and selecting the required features. The prediction stage involves four machine learning approaches GBA, RFA, SVM, and ANN. The overall strategy is illustrated in Figure 1.

Figure 1.

The overlay of the proposed work for the early prediction of CA.

3.3. Wavelet transform

One of the influential pre-processing techniques is WT that utilized to analyze the time-series and non-stationary data.¹⁴ To retain time localization, the signal decomposed into constituent frequencies that generate it ideal for various applications such as CA prediction and biomedical signal processing. To preserve critical low-frequency components, the biomedical signals based on high-frequency noise are filtered out using WT. Moreover, the computational efficiency was enhanced and significant coefficients were retained to minimize the data dimensionality.

3.4. Modified recursive feature elimination

Overfitting is the main problem when the prediction process involves small samples with higher dimensionality. It happens when the CA prediction based on the machine learning approaches is used since it is difficult to use a large number of CA cases. Hence feature selection approach is used. About 51% of the article uses feature selection to predict the disease and success and hence this article also uses feature selection based on MMRFE. Based on the predictive model, the most relevant features are identified with the robust feature selection called MRFE. The least important features are recursively removed and achieve the desired number of features iteratively.¹⁵ The faster and simpler model is made to minimize the number of features. The generalization and model accuracy are enhanced with MRFE to eliminate redundant or irrelevant features elimination. The complete set part conducted for feature selection after dividing the training data into two sets. The classification algorithm performance is enhanced by employing the process of feature selection to avoid overfitting. A better interpretability is provided to select the subsets and minimize the computational time as well as maximize the optimal number of feature identifications.

At each iteration, calculate the accuracy metric and in a recursive manner, eliminate the features. The model performance with each feature contributed to gaining the insights to observe the changes in accuracy metrics. The highest overall accuracy with the classifier determines an optimal feature set. Evaluate the set of features that are selected and complete the process of recursive elimination. The feature selection performance is computed and an internal accuracy is computed with the process of MRFE for each iteration. The risk models of CA with the critical predictors are identified. The model performance and interpretability were enhanced with complexity reduced by focusing on relevant features selection using MRFE. Additionally, the generalization and accuracy are enhanced with MRFE to eliminate redundant or irrelevant features. Iteratively selects the desired number of features and the least significant features are removed.

3.5. Prediction stage

The prediction model is based on four machine learning techniques known as GBA, RFA, SVM, and ANN. The main purpose of selecting these techniques is due to their strength in handling structured data, imbalanced datasets, and feature complexities. The GBA is used for handling the structured dataset with non-linear features. More often, based on the high-risk features it provides insights about the CA. meanwhile, the RFA integrates the multiple decision trees to improve stability and mitigate overfitting problems. The interpretable capacity is high and hence used in the medical application especially for the prediction of CA using the features or attributes.¹⁶ It provides feature scores for the features and helps in making the decision. The SVM can be used to take the prediction of medical applications where it requires binary classification tasks like the presence or absence of high risk of CA. moreover, with the weighted penalties it handles the imbalances classes of the CA dataset. In low dimensional spaces the overfitting is low using the SVM. The ANN uses a sufficient training dataset for learning complex patterns and combines various data modalities obtained from physiological and patient demographics. It is more flexible and hence it can be used in learning the non-linear correlation and hierarchical data. After conducting the experimental study, the best architecture is selected based on factors such as interpretability which provides details about the trust and validation, scalability for handling the real-time data, and prediction accuracy.

3.5.1. Gradient boosting algorithm (GBA)

This GBA is mainly used for classification and prediction issues and is an ensemble approach of machine learning technique. The main purpose of gradient boosting is to reduce the loss function with the gradient reverse direction. The pseudocode for the GBA is shown in Algorithm 1. This is used to predict the early CA.¹⁷ The intricate loss functions are reduced using the Gradient descent boosting which is not directly minimized. The reduced loss function is determined as Los in Algorithm 1. The prediction value is initialized as $F_{0} (x)$ for the mean single prediction. The gradient of loss is evaluated for the iteration $t = 1$ and the value that is predicted as $F_{1} (x)$ . The updated prediction value and the magnifier's step magnitude are taken as $F_{1} (x)$ . The process continues till the prediction of CA attains the value of $F_{T} (x)$ .

3.5.2. SVM classifier

SVM is used to predict the CA and the hyperplane is determined in SVM using the hypothesis function $H_{0}$ and is implied as,

\begin{aligned} H_{0} (x_{j}) = {\begin{cases} + 1, i f & W \times x + B \geq 0 \\ - 1, & O t h e r w i s e \end{cases} \end{aligned}

(1)

Based on the equation above the hyperplane for predicting the CA is split into +1 and −1 classes.¹⁸ To solve the optimization problem, the generated hyperplane will be partitioning the features and the values of W and B can be evaluated and maximized as,

\begin{aligned} _{W, B}^{M a x} Y S u b j e c t t o β_{j} \geq Y, j = 1, 2, \dots . g \end{aligned}

(2)

Here, the values for the $β$ and Y are determined as,

\begin{aligned} β & = z \times (\frac{W}{‖ W ‖} x + \frac{B}{‖ W ‖}) \end{aligned}

(3)

\begin{aligned} Y & =_{j = 1, \dots g}^{min} z \times (\frac{W}{‖ W ‖} x + \frac{B}{‖ W ‖}) \end{aligned}

(4)

Moreover, the value of $Y = \frac{f}{‖ W ‖}$ is set and determined the optimization equation as,

\begin{aligned} _{W, B}^{M a x} Y s u b j e c t t o χ_{j} \geq Y, j = 1, \dots . g \end{aligned}

(5)

The increment/ maximization would not affect the outcome of optimization and so the value of f is set as 1 and the above equation is rewritten as,

\begin{aligned} _{W, B}^{M a x} \frac{1}{‖ W ‖} s u b j e c t t o χ_{j} \geq Y, j = 1, \dots . g \end{aligned}

(6)

Meanwhile, the equation for minimization can be written based on the maximization equation as,

\begin{aligned} _{W, B}^{M a x} ‖ W ‖ s u b j e c t t o χ_{j} \geq Y, j = 1, \dots . g \end{aligned}

(7)

The stability of the l1 is lesser than the l2 and so,

\begin{aligned} _{W, B}^{M a x} \frac{1}{2} ‖ W ‖^{2} s u b j e c t t o z_{j} (n \times x + B) - 1, i = 1, \dots, z \end{aligned}

(8)

For predicting the CA, the SVM optimization parameter H is taken and the limitation are set based on the boundary distance. The optimizer used here is Adam optimizer to reduce the overfitting issues and the overall error during the prediction process.

3.5.3. Random forest algorithm (RFA)

RFA is another approach for predicting the CA from the features that are selected and it is a tree structure. The features that are selected are classified to determine the patient having the chance of occurrence of CA in the future.¹⁸ This prediction process involves, designing various decision trees for mapping the random samples from the selected features. It can be described as,

➢
The features from the dataset are predetermined to set the probability of replacement.
➢
The decision tree is built for every feature from the dataset taken. From the existing features, the training samples are determined based on the subspace created. The valuation of probabilities is made for m features. The prediction outputs are taken from the leaf node and continue until it reaches the termination point.

Integration of unpruned trees takes place with the random forest ensemble and makes the prediction decision. The pseudocode for the proposed RFA is shown in Algorithm 2.

3.5.4. ANN

The next technique used for CA prediction is ANN which is based on the functions of the human brain. It can be determined numerically as,

\begin{aligned} x_{j} = F (\sum_{j} W_{i j} Y_{j} + ρ_{j}) \end{aligned}

(9)

The neuron's input is $Y_{j}$ with the weight vector correlated as $W_{i j}$ and the threshold value of $ρ_{j}$ with biased value. The transfer function is defined as $F ()$ and the output is $x_{j}$ . The input of the network is always compared with the target input. The error of the network is calculated with the goal and network outputs. To attain the intended prediction, the weights are altered with the iterations—the usage of weights to store the knowledge by learning the neuron effectively. After the completion of the training process, the needed outputs are created based on the problem statement during the training. For continuous mapping, backpropagation algorithm is used most effectively.¹⁹ The mapping of input and output is done with the backpropagation which reduces the cost function of sum squared error based on the collected training instances.

The overfitting of the ANN is avoided during the prediction process of CA is attained with the K-fold cross-validation. The datasets are split into 10 folds and to begin the first fold is to test the model and the work provided and the usage of next folds are to train the model. When the 10 folds reach the testing set the system stops and the 10 fold structure of the proposed work is shown in Figure 2.

Figure 2.

Proposed 10-fold cross validation for the ANN.

With the 10-fold cross-validation the ANN achieves better results than the others and overcomes overfitting issues with the Adam Optimizer. This Adam optimizer is the best to solve the overfitting issues.

4. Experimental results and discussion

This section discusses the simulation outputs for the ML models for predicting CA. An effective and simpler model is ML and its efficiency is proved using comparative studies. The Python programming language performs simulation and the Google Colaboratory server trained in which the model incorporated with 16 GB RAM specification and 69 K GPU conducts the experiments.

4.1. Comparative investigation for analogous study

Figure 3 represents the evaluation study for the proposed methods. For predicting CA, metrics like F-score, precision, accuracy, and recall focus on various proposed techniques like GBA, RFA, SVM, and ANN. The finding discussion follows the below-structured comparison. The model's prediction shows the overall correctness. The comparison trial for each measure was computed with various target classification methods like GBA, RFA, SVM, and ANN. Every ML model-based individual model performance is evaluated using these graphical plots. Across all metrics, individual models outperform the better abilities of ML methods. In all cases, ANN yields greater results like 96.3% accuracy, 96.1% precision, 95% recall, and 94.65% F-score findings than other methods like GBA, RFA and SVM. The lower performance results of GBA, RFA and SVM are demonstrated than ANN.

Figure 3.

Evaluation study for proposed methods.

Figure 4 displays the evaluation study for the error performance of MAE and RMSE. Utilizes the methodologies of ANN with GBA, RFA and ANN for analysing the error performances. The error of average magnitude among the actual and predicted values is measured using MAE. The average squared square root measure among the actual and predicted values is measured using RMSE. Both larger and average errors are minimized and indicate superior results and achieve lowest the RMSE and MAE. To compute the MAE, the methodologies like RFA is 1.65%, GBA is 3.48%, SVM is 2.76% and ANN is 1.07%. To compute the RMSE, RFA is 2.32%, GBA is 3.21%, SVM is 2.87% and ANN is 1.29%. Contrasted to GBA, RFA and SVM, the ANN provided minimal MAE and RMSE-based error performances.

Figure 4.

Evaluation study for error performance of MAE and RMSE.

The CA prediction performance for various measures is tabulated in Table 2. A more clinically reliable and effective model is ensured to evaluate the multiple performance measures involved in the prediction of CA. The finding discussion follows the below-structured comparison of each measure. The model's prediction shows the overall correctness. The key performance metrics included the evaluation typically in case of measures like F-score, precision, accuracy, and recall when compared to the state-of-art to predict CA. To vary the epochs 20 to 100, the proposed performance of each measure outperformed superior results at the 100^th epoch.

Table 2.

Ca prediction performance for various measures.

Number of epochs	Accuracy	Precision	Recall	F-score
20	87%	83%	88%	89%
40	89%	85%	89%	90%
60	85%	90%	91%	91%
80	88%	91%	92%	93%
100	96%	95%	96%	94%

The comparative investigation for the prediction of CA performance measures is illustrated in Figure 5 and the results of precision, accuracy, recall, and F-score are mentioned in Figures 5(i) to (iv). For predicting CA, the metrics like F-score, precision, accuracy, and recall focus on various proposed techniques like DT-SVM,⁶ LGBM,⁷ HML,⁸ and ML-SWS⁹ and proposed. The following structured comparison of each measure shows the discussion of the result findings and the overall correctness represented by the model's predictions. To compare the state-of-art CA measures like F-score, precision, accuracy, and recall are the key performance metrics involved in this CA prediction. The proposed method shows 95% accuracy, 96% precision, 95.83% recall, and 95.12% F-score results and these are superior to compare with DT-SVM,⁶ LGBM,⁷ HML,⁸ and ML-SWS.⁹

Figure 5.

Comparative study for the prediction using various measures; (i) precision, (ii) accuracy, (iii) recall, and (iv) F-score.

The comparative study for the prediction of CA with respect to error performance is illustrated in Figure 6. Use the methodologies of DT-SVM,⁶ LGBM,⁷ HML,⁸ and ML-SWS⁹ and propose to investigate the error performances. The error of average magnitude among the actual and predicted values is measured using MAE. The average squared square root measure among the actual and predicted values is measured using RMSE. This graphical representation evaluates an overall error rate with respect to the comparative methods and along with the number of epochs from 20 to 100. This investigation shows error rates of 4.43 for DT-SVM, 3.97% for LGBM, 3.02% for HML, 1.97 for ML-SWS, and 1.09% for the proposed method. However, the proposed shows a minimum rate of error values than DT-SVM, LGBM, HML, and ML-SWS.

Figure 6.

Comparative study for the prediction of CA with respect to error performance.

Table 3 shows the state-of-art results for training time, inference time, and interpretability. It is crucial to provide the competing model's comparative analysis and empirical basis when the ANN suggests more accurate results. The domain problem context evaluates the model requirements when the key metric is accuracy. The efficiency of computational for both inference and training and in healthcare, the capacity to clarify predictions during interpretability. The GBA is more interpretable and due to the rule-based structure, the RFA is intuitive. Compared to GBA, the RFA has moderate interpretability and also ANN and SVM have lower interpretability. The inference and training time of ANN is 18 ms and 2.30 h.

Table 3.

The state-of-art results for training time, inference time, and interpretability.

	Parameters
Methods	interpretability	inference Time	training Time
GBA	High	<1 ms	1 min
RFA	Moderate	9 ms	2 min
SVM	Low	47 ms	29 min
ANN	Low	18 ms	2.30 h

Table 4 shows the results for accuracy and clinician adoption likelihood. The multiple factors such as running and training computational costs included. Patient outcomes are affected by decisions with important healthcare applications. It is in critical scenarios with the particular prediction to texture confidence made. Scrutinized the transparent models to errors and biases. The relationship among the input features are the coefficients with the less interpretable are an individual decisions. The non-linear transformations of multiple layers rely on ANN and it is hard to interpret. As well as the spurious correlations or overfit noise are nontransparent models. The relationships of explicitly modeling among the outcomes and features that combines the predictions.

Table 4.

Results for accuracy and clinician adoption likelihood.

	Parameters
Methods	clinician adoption likelihood	Accuracy
GBA	Moderate	Higher
RFA	Lower	Higher
SVM	Higher	Moderate
ANN	Lower without interpretability improvements	Very Higher

Table 5 provides the comparative analysis, focusing on scope, methodology, data usage, performance, and clinical relevance.

Table 5.

Comparative analysis.

Criteria	Proposed	Šećkanović et al.²⁰	Begić et al.²¹	Ademović et al.²²	Begic et al.²³
Main Focus	Early prediction of cardiac arrest using ML	Review of AI applications in cardiology	Risk reduction from heat-not-burn tobacco on cardiovascular health	Triglyceride as a marker of residual CVD risk	Left atrial strain in pediatric cardiology
Methods Used	ANN, SVM, GBA, RFA with Wavelet Transform + MRFE	Review of various AI/ML models (CNN, SVM, RF, ANN)	Systematic review of RCTs	Statistical analysis of HeartScore and triglycerides	Speckle-tracking echocardiography + ML analysis
Dataset	Public Kaggle dataset (cardiac risk prediction)	Not empirical – literature-based	RCT data across trials	Patient cohort with triglyceride and HeartScore data	Pediatric echocardiography data
Feature Selection	Modified Recursive Feature Elimination (MRFE)	Discussed general techniques (not applied)	Not applicable	Analysis focused on triglycerides	Focused on strain measurements, not general features
Prediction Algorithms	ANN (best), SVM, GBA, RFA	Overview of ML/DL in cardiology (no performance metrics)	Not ML-based; review of intervention outcomes	Not ML; statistical correlation with biomarkers	ML applied to echo data (specific to strain analysis)
Performance Metrics	ANN: 96.3% Accuracy, 96.1% Precision, 95% Recall, 94.65% F-score	Not applicable	Not applicable	Not ML-based; outcomes via clinical scores	Performance reported, but specific metrics not directly comparable
Clinical Relevance	High: Real-time prediction, interpretable model potential	High-level review with implications for clinical AI adoption	Focus on harm reduction, policy and health behavior	Biomarker-focused, potential for ML feature inclusion	Specialized application in pediatric diagnostics
Strengths	Rigorous ML pipeline; strong comparative metrics; feature reduction; real-time use	Broad review of AI in cardiology; helpful for context	Clinical RCT synthesis; public health implications	Identifies novel predictive marker	Advanced imaging ML application; pediatric focus
Limitations	Computational cost of ANN; limited real-world validation	No empirical validation or experimentation	No predictive modeling or ML used	Narrow focus on a single biomarker	Specialized scope; not generalizable to broader CA prediction

4.2. Discussion

4.2.1. Computational cost and efficiency of the ML model

In the implementation of ML models for real-time clinical settings, computational cost is a critical factor, especially in resource-constrained environments. Among the evaluated models, Artificial Neural Networks (ANNs) stand out for their high predictive performance but also exhibit significant computational demands due to their iterative optimization processes and a large number of tunable parameters.

4.2.2. Computational cost

ANNs are inherently computationally intensive owing to their deep architecture and the need for repeated weight adjustments during training. The complexity and size of the input dataset, along with the depth and breadth of the network, directly influence training time and resource consumption. While the proposed ANN model offers faster training compared to traditional models like SVM, DT, and Logistic Regression (LR), it still requires specialized hardware, such as GPUs or edge accelerators, to ensure timely processing and prediction—particularly for continuous monitoring in real-time applications such as cardiac arrest (CA) alerts.

To address this, a proof-of-concept (PoC) approach using lighter versions of the model may be beneficial for initial deployment, minimizing the need for high-end infrastructure. Moreover, the ANN's ability to rapidly generate timely alerts can significantly enhance clinical decision-making, provided the infrastructure supports its computational requirements.

4.2.3. Computational efficiency

Despite its complexity, the computational efficiency of the ANN can be significantly improved through model optimization techniques. This includes pruning unnecessary neuron connections or reducing neuron weights without sacrificing performance, thereby shrinking the model size. Such lightweight structures are better suited for deployment on low-power devices commonly used in portable or wearable health monitoring systems.

Furthermore, dimensionality reduction strategies—such as the MRFE employed in this study—help in minimizing the input feature space, thereby reducing the computational burden. Although ANN models may offer only marginal improvements in accuracy over simpler models, their capacity to model complex, non-linear interactions among clinical variables justifies their computational trade-offs in critical applications.

By integrating well-engineered and relevant features, near-equivalent performance can also be achieved with simpler models like Random Forests or Gradient Boosting, especially where interpretability and low latency are prioritized. Nonetheless, for high-stakes scenarios requiring superior predictive power, optimized ANN models remain a compelling choice.

5. Conclusion

One of the critical challenges facing healthcare applications is the early prediction of CA. The timely intervention might mitigate the loss of lives and enhance the patient's health. The potential risk assessments of the CA are stated with the proposed technique known as GBA, RFA, SVM, and ANN. The simulation outcomes indicate that the proposed models can be easily used to capture the intricate details and correlate the non-linear relationships of the patient's data. Among the ANN shows the best accuracy and provides better solutions for the early prediction of CA. The feature selection is based on the MRFE which also acts as the interpretability framework. This ensures the transparency of the prediction accuracy and can be used for clinical assessments. Prior to this, the dataset collected is pre-processed using the Wavelet Transform. For the analysis, different parameters are taken and the ANN shows a prediction accuracy of 96.3%, 96.1% precision, 95% recall, and 94.65% F-score which are maximum than the other techniques. In the future, we are planning to implement the proposed work by collaborating with the healthcare authorities to provide details about interpretability and real-time applications.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

Data will be made available upon reasonable request.

Permission to reproduce material from other sources

Not Applicable.

References

Chae

Gil

H-W

Cho

N-J

, et al. Machine learning-based cardiac arrest prediction for early warning system. Mathematics 2022; 10: 2049.

Kwon

Jeon

Kim

, et al. Deep-learning-based out-of-hospital cardiac arrest prognostic system to predict clinical outcomes. Resuscitation 2019; 139: 84–91.

Lin

, et al. Machine learning for early prediction of in-hospital cardiac arrest in patients with acute coronary syndromes. Clin Cardiol 2021; 44: 349–356.

Chauhan

Kumar

Chauhan

, et al., (eds). Cardiac arrest prediction using machine learning algorithms. In: 2019 2nd international conference on intelligent computing, instrumentation and control technologies (ICICICT), 5–6 July 2019.

Aqel

Syaj

Al-Bzour

, et al. Artificial intelligence and machine learning applications in sudden cardiac arrest prediction and management: a comprehensive review. Curr Cardiol Rep. 2023; 25: 1391–1396.

Bajpai

Sinha

Yadav

, et al., (eds). Early prediction of cardiac arrest using hybrid machine learning models. In: 2023 17th international conference on electronics computer and computation (ICECCO), 1–2 June 2023.

Lee

Yang

H-L

Ryu

, et al. Real-time machine learning model to predict in-hospital cardiac arrest using heart rate variability in ICU. npj Digit Med 2023; 6: 215.

Nasiruddin

Dutta

Sikder

, et al. Predicting heart failure survival with machine learning: assessing my risk. J Comput Sci Technol Stud 2024; 6: 42–55.

Hannan

Cheema

Pires

. Machine learning-based smart wearable system for cardiac arrest monitoring using hybrid computing. Biomed Signal Process Control. 2024; 87: 105519.

10.

Hassan

Zaman

Rahman

, et al. Efficient prediction of coronary artery disease using machine learning algorithms with feature selection techniques. Comput Electr Eng. 2024; 115: 109130.

11.

Singh

Thongam

Choudhary

, et al. An integrated machine learning approach for congestive heart failure prediction. Diagnostics 2024; 14: 736.

12.

Muneeb Hassan

Ameeq

Jamal

, et al. Prevalence of COVID-19 among patients with chronic obstructive pulmonary disease and tuberculosis. Ann Med. 2023; 55: 285–291.

13.

Hassan

Tahir

Ameeq

, et al. Risk factors identification of COVID-19 patients with chronic obstructive pulmonary disease: a retrospective study in Punjab-Pakistan. Immun Inflamm Dis 2023; 11: 981.

14.

Azzouz

Bengherbia

Wira

, et al. An efficient ECG signals denoising technique based on the combination of particle swarm optimisation and wavelet transform. Heliyon 2024; 10: e26171.

15.

Awad

Fraihat

. Recursive feature elimination with cross-validation with decision tree: feature selection method for machine learning-based intrusion detection systems. J Sens Actuator Netw 2023; 12: 67.

16.

Gilpin

Bau

Yuan

, et al., (eds). Explaining explanations: an overview of interpretability of machine learning. In: 2018 IEEE 5th international conference on data science and advanced analytics (DSAA), 1–3 October 2018.

17.

Bahad

Saxena

(eds). Study of AdaBoost and gradient boosting algorithms for predictive analytics. In: International conference on intelligent computing and smart communication 2019, Singapore, 2020: Springer Singapore.

18.

VijiyaKumar

Lavanya

Nirmala

, et al., (eds). Random forest algorithm for the prediction of diabetes. In: 2019 IEEE international conference on system, computation, automation and networking (ICSCAN), 29–30 March 2019.

19.

Carolin Mabel

Fernandez

. Analysis of wind power generation and prediction using ANN: a case study. Renew Energy 2008; 33: 986–992.

20.

Šećkanović

Šehovac

Spahić

, et al., (eds). Review of artificial intelligence application in cardiology. In: 2020 9th Mediterranean conference on embedded computing (MECO), 8–11 June 2020.

21.

Begić

Aziri

Omeragić

, et al. Heat-not-burn tobacco products and cardiovascular risk reduction: a systematic review of randomized controlled trials. Technol Health Care. 2023; 31: 1457–1491.

22.

Ademović

Čakarić

Durak-Nalbantić

, et al. Triglyceride level as a potential marker of residual cardiovascular risk and its relation to HeartScore. Technol Health Care. 2023; 31: 2145–2153.

23.

Begic

Djukic

Begic

, et al. Left atrial strain analysis in the realm of pediatric cardiology: advantages and implications. Technol Health Care. 2024; 32: 1255–1272.

Number of epochs	Accuracy	Precision	Recall	F-score
20	87%	83%	88%	89%
40	89%	85%	89%	90%
60	85%	90%	91%	91%
80	88%	91%	92%	93%
100	96%	95%	96%	94%

Number of epochs	Accuracy	Precision	Recall	F-score
20	87%	83%	88%	89%
40	89%	85%	89%	90%
60	85%	90%	91%	91%
80	88%	91%	92%	93%
100	96%	95%	96%	94%

Number of epochs	Accuracy	Precision	Recall	F-score
20	87%	83%	88%	89%
40	89%	85%	89%	90%
60	85%	90%	91%	91%
80	88%	91%	92%	93%
100	96%	95%	96%	94%