The Role of Artificial Intelligence in the Diagnosis and Prognosis of Heart Diseases: A Systematic Review

Abstract

The integration of artificial intelligence (AI) into the diagnosis and prognosis of heart diseases is transforming cardiovascular and cardiac healthcare, improving predictive accuracy, and personalizing treatment plans. This review presents a novel contribution by providing a comprehensive overview of both diagnosis and prognosis in heart diseases through AI, covering ML and DL models. Following the PRISMA guidelines, a total of 84 recent research articles sourced from significant journals are reported. A bibliometric analysis using the VOSviewer tool was performed to map the impact of AI, enabling a detailed examination of academic connections and contributions. The findings reveal that DL models were employed 63% for diagnosis tasks, while ML models were utilized in 37% of the studies. Key recommendations include the incorporation of essential model evaluation metrics, as clinical validation indicators, integrating explainable artificial intelligence (XAI) to improve the transparency and interpretability of models, and adopting standardized frameworks to enable smooth clinical integration. This review highlights the potential of AI to improve cardiac and cardiovascular diagnosis and prognosis, providing an overview of its strengths, limitations, challenges and the possible application as AI-driven tools in patient monitoring and to support specialists in the decision-making process.

1. Introduction

AI algorithms are increasingly utilized in the diagnosis and prognosis of heart diseases due to their numerous benefits over conventional methods, achieving high performance metrics despite the presence of complex and noisy data as a significant advantage. This is attributable to their capacity to discern subtle or intricate patterns that may evade detection by humans [1]. Additionally, the capacity of AI algorithms to rapidly process vast quantities of data renders them highly efficacious tools for addressing problems that are data-intensive in nature. This is vital for the prediction of heart diseases where multiple types of data, including electrocardiograms (ECGs), echocardiograms, and medical records, often require analysis [2,3,4,5,6].

To address these challenges, specialized AI algorithms are employed, using various techniques such as supervised, unsupervised, and reinforcement learning. For supervised learning, the most common algorithms are regression types [5], decision trees (DT) [6], support vector machines (SVM) [7,8], and k-nearest neighbors (KNN) [9]. These algorithms are widely adopted in the clinical domain due to their ability to strike a balance between performance, interpretability and adaptability to diverse data types, which contribute to their effectiveness.

For the case of unsupervised learning, clustering and dimensionality reduction algorithms are used [10]. On the other hand, artificial neural networks (ANNs) are another type of algorithm used, which has the capacity to emulate the structural and functional characteristics of the human brain [11]. Furthermore, some of the latest AI techniques such as natural language processing (NLP), computer vision, evolutionary computation, and swarm intelligence are also being implemented for early diagnosis of heart diseases and are paramount for effective treatment and prevention.

Within this context, these AI, ML and DL methods are being applied to the diagnosis and prognosis of several prevalent heart diseases, that can include coronary artery disease, cardiac stroke, congenital heart disease, arrhythmia, heart failure, venous conditions, valvular disorders, peripheral artery disease, pulmonary hypertension, myocardial ischemia, aortic stenosis, atrial fibrillation, myocardial infarction, cardiovascular diseases, among others [12,13,14,15,16,17,18,19].

As several of these conditions remain undiagnosed until they progress to severe stages, advancements in AI are becoming increasingly important and promising [20]. However, it is important to note that the majority of research focuses primarily on diagnosis tasks; therefore, it is imperative to also prioritize in the prognosis of heart diseases, as both require improved evaluation metrics and clinical validation by medical professionals.

In the present study, this review offers a novel contribution by providing a comprehensive overview of both diagnosis and prognosis in heart diseases through AI models, including machine learning (ML) and deep learning (DL), presenting a broader scope compared to other surveys that focus only on diagnostic tasks. Moreover, it conducts a comprehensive comparative analysis of the most recent and relevant works, emphasizing a general overview of the strengths, limitations, and challenges associated with AI models, presenting a roadmap of how challenges could be addressed. In this context, trends include advancements in diagnostic precision, telecardiology and AI-powered medical assistants, while challenges remain in terms of data accessibility and quality, and model interpretability.

The findings demonstrate that the majority of the proposed models achieve acceptable performance in the diagnosis and prognosis of heart diseases. However, the incorporation of explainable AI (XAI) to enable both model transparency and interpretability is noteworthy, emphasizing the importance of reporting key evaluation metrics for healthcare-focused models.

2. Materials and Methods

Search Strategy and Criteria Selection

To execute the search strategy and selection criteria, a bibliometric analysis was conducted, followed by the visualization of scientific networks using VOSviewer software and the retrieval of bibliographic references from PubMed. This software is specifically designed for constructing and visualizing bibliometric networks. For the network analysis, the PubMed repository was utilized, and the keywords selected were artificial intelligence and cardiovascular.

The AI term is positioned as the central and largest node in the network. Other terms are grouped in proximity to it, delineated by varying colors that represent the diverse subfields and applications of AI. The network analysis visualization illustrates concepts related to AI, specifically focusing on research topics for heart-related diseases. The words artificial intelligence, humans, cardiovascular diseases, and ML— being the largest ones—are located in the central nodes. These words show the largest connections, suggesting that they are key concepts most frequently used in the field of study. Figure 1 shows the network analysis visualization.

This analysis identified some insights by grouping the colors blue, green, yellow, red, and purple into five categories. A closer examination of the results from these groups reveals the following:

• The blue group represents clinical applications and decision support, including fundamental concepts such as diagnosis, prognosis, stroke, and atrial fibrillation.
• The red group proposes models based on prognosis and treatment outcomes for heart-related diseases.
• The green group shows healthcare systems related to heart disease, precision medicine, and telemedicine.
• The yellow group is related to fundamental techniques for implementing AI systems in heart healthcare, such as medical imaging.
• The purple group is specialized, showing the most sophisticated computational advances used to obtain data for diagnosing heart disease.

The yellow and purple groups are strongly connected to the blue and red groups, which represent clinical applications. This shows that medical images are the most used type of data for AI-based diagnosis and prognosis. The green group, representing healthcare systems, is connected to the other groups, indicating that the objective of developing AI models for detection and outcome prediction is supported by data infrastructure, such as electronic health records.

The blue cluster, where the most repeated keyword is artificial intelligence due to its node size, contains nodes related to cardiovascular disease, as well as those for the identification and disease progression assessment. The present bibliometric analysis focuses on the spatial distribution of these three nodes, conducting a comprehensive review of the most recent and innovative advancements in the field of heart disease diagnosis and prognosis using AI.

The bibliometric analysis substantially facilitated both the visualization of keyword networks and the guidance of the literature selection, thereby identifying influential publications and the diverse technologies employed in healthcare. Furthermore, by mapping the co-occurrence of keywords and the generated clusters, it was feasible to discern the trajectory of trends in AI-driven healthcare, guiding principle for refining the inclusion and exclusion criteria employed in the research search process, ensuring that the most innovative and representative studies pertinent to the review’s subject matter were included. Specifically, for keyword co-occurrence networks from the bibliometric analysis were used to expand and finalize the search terms, thereby ensuring emerging sub-topics were captured. The trend analysis guided the temporal boundaries of the inclusion criteria, guaranteeing the search strategy is focused on the most innovative and representative studies.

The search strategy was conducted using the PRISMA methodology and followed meticulous procedures to ensure the inclusion of pertinent and high-quality studies. To ensure the relevance of the studies, specific criteria were applied, including the year of publication, peer-reviewed journals and conferences, and articles reporting novel AI models and quantitative performance models were applied. The selected period of publication year, from 2019 to 2026, captures the most recent and significant advancements in AI, applied to the diagnosis and prognosis of heart diseases. Furthermore, focusing on this timeframe allows the reported papers to reflect the current state-of-the-art and the rapid evolution of AI methodologies in this area.

The systematic design of this methodology ensures the efficient implementation of essential filters, facilitating access to articles that contribute to the current state of the field and enhancing its efficacy. The databases accessed for the bibliographic search were mainly PubMed, Nature, Springer, Elsevier, and IEEE, among others. To carry out the literature research, combinations of keywords were used, including artificial intelligence, machine learning, deep learning, heart, cardiovascular, cardiac, diagnosis and prognosis.

Table 1. Keywords and search query employed in databases.

Database	Search Strategy
PubMed	(“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)
Nature	(“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)
Springer	TITLE-ABS-KEY (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)
Elsevier	TITLE-ABS-KEY (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)
IEEE	(“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)
MDPI	(“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“diagnosis” OR “prognosis”) AND (“heart” OR “cardiac” OR “cardiovascular”)

illustrates the search strategy employed for each database, utilizing keywords and Boolean operators, while in Figure 2 the PRISMA methodology diagram is shown.

As can be seen in Table 1, some search strategies are the same for certain databases; however, for Springer and Elsevier, the TITLE-ABS-KEY entry is allowed, which enables keywords to be searched specifically within the article’s title, abstract, and keywords. Furthermore, the Boolean operators ‘OR’ and ‘AND’ are used in all cases.

The PRISMA diagram in Figure 2 illustrates the systematic review process, which began with the identification of 300 records, where 120 were from PubMed, 80 from Elsevier, 35 from Nature, 25 from Springer, 25 from IEEE, and 15 MDPI. Following the removal of 40 duplicates through electronic means, 260 titles were screened, resulting in the exclusion of 110. Subsequently, the remaining 150 abstracts were screened, with 64 excluded due to their scope not aligning with the review, articles not written in English, commentaries and position papers. This left 84 full-text articles for eligibility assessment, all of which were included in this review. The diagram illustrates the systematic screening process that narrowed down the initial pool of studies to the final set of 84 novel and relevant studies, ensuring a transparent and reproducible selection process by the accuracy of the results.

To mitigate the influence of publication bias, a comprehensive approach was implemented, encompassing the utilization of multiple databases, the inclusion of both journals and high-quality conference proceedings, and meticulous manual reference verification.

This systematic review was prospectively registered in the Open Science Framework (OSF). The protocol can be accessed at: https://doi.org/10.17605/OSF.IO/P3YQ8 (accessed on 16 April 2026).

3. Results

In the present results section, a total of 84 articles related to the diagnosis and prognosis of heart diseases are reported, where these articles are divided into two categories: diagnosis using ML and DL algorithms, and prognosis with different AI methods. In the context of ML and DL techniques, research studies reported in this paper implement traditional algorithms, including LR, NB, DT, RF, SVM, KNN, as well as some more sophisticated ones, such as ANN, CNN, boosting, and ensemble types. On the other hand, a limited number of works have employed IoT and image processing for the task of prognosis.

The research, spanning from 2019 to 2026, is novel in nature and involves reporting the implementation of various AI models for diagnosing and prognosing heart diseases, including cardiovascular and cardiac diseases.

The diagnosis task is a clinical procedure employed to ascertain the root cause of a patient’s symptoms, which typically involves a process of obtaining a medical history, conducting a physical examination, and requesting diagnostic tests. The diagnostic process comprises several stages: initially, a patient encounters a health issue, identifies their symptoms, and subsequently seeks medical attention. Upon engaging with the healthcare system, a multi-step and iterative process commences to establish a provisional diagnosis. This process involves compiling relevant data about the patient’s health concerns through various methods, such as obtaining a clinical history, conducting an interview, performing a physical examination, ordering essential diagnostic tests, and consulting with other healthcare professionals.

Diagnostic assessment and prognostic evaluation tasks involve an information collection process that employs different methods at various stages, acquiring diagnostic data in varying sequences. To generate accurate results, it is essential to integrate and interpret the collected data; continuous revision of hypotheses and adjustments to prior probabilities as new information becomes available is also essential. Effective and objective communication among healthcare professionals, patients, and their families is vital for the successful gathering, integration, and interpretation of information [21].

For the diagnosis and prognosis of heart diseases through AI, several datasets are utilized—encompassing both public and private sources and including data such as electrocardiograms, X-rays, and medical records. For the purpose of analyzing data as signals, ML and DL algorithms are generally used, while DL is used for image-type data.

The most widely used public datasets for this purpose include the Multi-Ethnic Study of Atherosclerosis [22], Framingham Heart Study [23], Heart Disease UCI Dataset [24], Cardiac Arrhythmia Database [25], PTB Diagnostic ECG Database [26], Medical Information Mart for Intensive Care [27], EchoNet-Dynamic [28], and others.

3.1. Performance Evaluation for AI Models

The most frequently employed metrics for AI diagnosis models reported in research include accuracy (Acc), precision (P), sensitivity (Sen) or recall, specificity (Spe), F1-Score (F1), receiver operating characteristic (ROC), area under the curve (AUC), coefficient of determination (R²), and mean absolute error (MAE).

The accuracy metric quantifies the proportion of correct predictions, including true positives and true negatives, out of the total number of cases examined. Precision measures the proportion of true positive predictions among all positive predictions, indicating the model’s ability to distinguish between positive and negative instances. Sensitivity, or recall, assesses the model’s ability to correctly identify actual positive cases. In contrast, specificity measures the proportion of correctly identified negatives, reflecting the model’s capacity to detect negatives. The F1-score combines precision and recall into a harmonic mean, especially useful in imbalanced class scenarios. For regression, metrics such as R² and MAE are used. R² indicates the variance in the dependent variable explained by the model, while MAE calculates the average absolute error between predictions and actual values, indicating the magnitude of error without regard to error direction. AUC refers to the area under the ROC curve, which evaluates a binary classifier’s ability to distinguish between classes, especially when models produce probability scores.

3.2. AI in Heart Diseases Diagnosis

A total of 72 studies are presented in this section, which have employed DL and ML models for the diagnosis of heart diseases. The data analysis indicates a marked preference for convolutional neural networks (CNN) models, with 35% of the research employing this type of DL. Additionally, there is a tendency to use boosting models, which are ML models used when highly complex learning patterns are present, at a rate of 3% for Cat Boost and 10% for XGBoost.

However, other techniques, such as DT, KNN, linear regression (LR), Naive Bayesians (NB), random forests (RF), SVM, ANN’s, deep neural networks (DNN), Long short-term memory (LSTM), multilayer perceptron (MLP), recurrent neural networks (RNN), variational auto encoder (VAE) and ensemble models are also commonly used. Figure 3 displays how the techniques used to diagnose heart diseases were distributed among the studies reported in this review.

It is remarkable how DL-based models are gaining importance in healthcare, especially in disease diagnosis. This is supported by the fact that 63% of the papers reported in this research use DL models, while the remaining 37% use ML models. This trend is driven by the fact that most research involves diagnosis from medical images, an area where DL excels. This technique has a strong ability to recognize complex patterns, nonlinear relationships, and other characteristics.

Table 2. Summary of ML and DL algorithms implemented for the diagnosis of cardiac and cardiovascular diseases.

Heart Disease Condition	ML Algorithm	DL Algorithm
Heart disease	LR, RF, DT, XGBoost, Cat Boost, SVM	ANN, CNN, LSTM, RNN
Heart failure	NB, RF, SVM	MLP, CNN
Cardiovascular disease	RF, LR, KNN, RF, Cat Boost, LSTM	CNN, ANN
Coronary artery disease	XGBoost, RF	CNN
Pulmonary hypertension	-	CNN
Myocardial ischemia	XGBoost	VAE, ANN
Arrythmia	-	CNN
Aortic stenosis	-	ANN, CNN
Atrial fibrillation	-	CNN, DNN
Myocardial infarction	-	VAE, ANN

presents the summary of the ML and DL algorithms employed in the reviewed literature, concerning to the diagnosis of cardiac and cardiovascular diseases. Additionally, the average accuracy of the models for each heart disease condition is also displayed.

As displayed in Table 2, a general categorization of certain cardiac and cardiovascular diseases is provided, along with the AI algorithms employed for their diagnosis. Based on this data, the implementation of ML and DL algorithms has been explored, where the use of CNNs has been identified in the prediction of most heart diseases, showcasing their effectiveness in diagnosis through images. Furthermore, the average accuracy demonstrates the effectiveness of AI models in the diagnosis of heart disease conditions.

3.2.1. ML Models for Diagnosing Heart Diseases

Currently, the implementation of ML models has become an important tool for the early detection and diagnosis of heart diseases. These models provide speed, accuracy, and reliability compared to traditional methods. There is a wide range of ML algorithms used in clinical analysis, image pattern recognition, and clinical decision support tasks.

Table 3. ML models implemented in the diagnosis of cardiac diseases.

Ref.	Contribution	ML Model	Performance	Validation
[29]	Heart disease prediction	LR	Acc = 92% P = 90% Sen = 96% F1 = 93%	Internal
[30]	Genome transcriptomic data for clinical cardiomyopathy diagnosis	SVM	P = 90%	Unspecified
[31]	Heart failure disease prediction	NB	Acc = 86% P = 73% Sen = 73%	Internal
[32]	Heart disease prediction	XGBoost	Acc = 97%	Unspecified
[33]	Heart disease prediction	XGBoost	Acc = 90% AUC = 94%	Internal
[34]	Heart disease risks prediction	Cat Boost	Acc = 98% Sen = 97% Spe = 96%	Unspecified
[35]	Heart disease early detection	XGBoost	Acc = 97% P = 95% Sen = 96% Spe = 90% F1 = 92% AUC = 98%	Unspecified
[36]	Heart disease prediction	XGBoost	Acc = 97% P = 97% Sen = 98% Spe = 98% F1 = 99% AUC = 98%	Unspecified
[37]	Heart disease prediction	SVM	Acc = 95% P = 96% Sen = 94% Spe = 95%	Unspecified
[38]	Early heart disease prediction	SVM	Acc = 92% P = 86% Sen = 90% Spe = 93%	Internal
[39]	All-cause mortality for 1, 2, 3, 4, and 5-years prediction	RF	AUC = 76%	Unspecified
[40]	Heart failure prediction	RF	Acc = 96% P = 98% Sen = 95% Spe = 98% Spe = 98%	Unspecified
[41]	Heart failure assessment	SVM	Acc = 98% P = 97% Sen = 97% Spe = 96%	Unspecified
[42]	Heart disease prediction	XGBoost	Acc = 96% P = 95% Sen = 98% F1 = 96% AUC = 96%	Internal
[43]	Heart disease prediction	DT	Acc = 80% P = 78% Sen = 65% F1 = 71%	Unspecified

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, Area under the ROC Curve.

offers an overview of the latest research on using ML models for diagnosing cardiac diseases such as myocardial infarction, cardiovascular disease, coronary artery disease, heart failure, and others, while

Table 4. ML models implemented in the diagnosis of cardiovascular diseases.

Ref.	Contribution	ML Model	Performance	Validation
[44]	Acute coronary syndrome outcomes and mortality prediction	RF	Acc= 99% P = 99% Sen= 99% F1 = 99% AUC = 99%	Unspecified
[45]	Cardiovascular disease prediction	RF	Acc = 99% P = 96% Sen= 96% F1 = 96%	Unspecified
[46]	Analyze healthcare data to predict heart disease	SVM	Acc = 91% P = 90% Sen = 94% Spe = 87% F1 = 92%	Unspecified
[47]	Risk of cardiovascular disease prediction	LR	Acc = 87%	Unspecified
[48]	Cardiovascular disease prediction	KNN	Acc = 95%	Unspecified
[49]	Cardiovascular disease detection	RF	AUC = 80%	Unspecified
[50]	Myocardial infarction diagnosis using cardiac troponin concentrations	XGBoost	AUC = 95%	Unspecified
[51]	Coronary artery disease prediction	XGBoost	AUC = 88%	Unspecified
[52]	Cardiovascular diseases prediction	Cat Boost	Acc = 98% P = 97% Sen = 98% F1 = 98% Spe = 97%	Internal
[53]	Coronary heart disease risk prediction	XGBoost	AUC = 82%	External
[54]	Early stage of cardiovascular disease prediction	SVM	Acc = 81% Sen = 93% Spe = 89%	Unspecified

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, area under the ROC curve.

focuses on ML models implemented for the diagnosis of cardiovascular diseases.

The tables contain information such as research references, the contributions of each study, the ML model implemented, the performance reported, and the validation type. For the validation information, the internal information is only applied to test data; external information is applied to real clinical contexts or is unspecified if there is no information related to model validation.

From the table presented above, which showcases ML models applied to the diagnosis of cardiac diseases, the models demonstrate satisfactory performance in their respective tasks. Notably, the research [34,36,41] has yielded significant advancements in heart disease risk prediction, heart disease prediction, and heart failure assessment, respectively. These findings underscore the potential of trained models to effectively diagnose specific cardiac diseases.

From Table 4, which presents ML models applied to the diagnosis of cardiovascular diseases, we can highlight [44], which presents very good metrics and also predicts acute coronary syndrome outcomes and mortality. Additionally, we can highlight study [45], which predicts cardiovascular disease. Both studies obtain very good metrics; however, they require validation in real clinical settings to verify their efficacy.

As shown in Table 3 and Table 4, 25 ML-based models for cardiac and cardiovascular diseases diagnosis achieve acceptable metrics. The models used are of the supervised learning type, where the dataset already contains labels, which helps the model to generalize better. Figure 4 provides a summary of the models used in diagnosing heart diseases that are reported in this research.

The overall performance of all models is generally over 90%, as evidenced by their ability to generalize effectively. This is primarily attributed to the fact that the datasets utilized are labeled with classes, which facilitates the model’s ability to distinguish between them more effectively.

In order to show the trend in the development of ML models for the diagnosis of cardiac and cardiovascular diseases, in Figure 4 a graph is presented that displays the information related to the average accuracy by ML model, the percentage of studies employing each model, and the mean publication year.

As displayed in Figure 4, XGBoost models are the most frequently employed in reported studies, followed by RF and SVM. The boosting models, such as XGBoost and Cat Boost, are commonly implemented when data patterns are complex to learn, indicating that medical data often lacks simple or clear patterns, and these models are highly robust in their algorithms. Moreover, ensemble models are among the most recently explored models, with a mean publication year of 2025.

3.2.2. DL Models for Diagnosing Heart Diseases

DL algorithms have proven highly effective in diagnosing heart disease; they have also been used to develop tools that analyze ECGs, echocardiograms, and other data to automatically detect heart conditions with high accuracy. In addition, DL models have been utilized to develop risk prediction tools for patients with specific risk factors, including diabetes and high blood pressure.

Table 5. DL models implemented for cardiac diseases diagnosis.

Ref.	Contribution	DL Model	Performance	Validation
[55]	Cardiac disease detection	MLP	Acc = 92% P = 95% Sen = 96% F1 = 94%	Unspecified
[56]	Early identification of the disease	CNN	P = 98% Spe = 96% Sen = 100%	Unspecified
[57]	Categorization analysis of electrocardiogram using rhythm or beat features	CNN	Acc = 99% P = 99% Sen = 99% Spe = 99%	Unspecified
[58]	Heart disease diagnosis	ANN	Acc = 82% P = 82% Sen = 94%	Unspecified
[59]	Heart disease prediction	CNN	Acc = 93% Sen = 94% Spe = 91%	Unspecified
[60]	Heart disease prediction	LSTM	Acc = 94%	Unspecified
[61]	Heart disease prediction	LSTM	Acc = 98%	Unspecified
[62]	Heart disease prediction	LSTM	Acc = 96% Spe = 95% Sen = 95%	Unspecified
[63]	Heart disease threat detection	DNN	Acc = 94% P = 98% Sen = 100%	Unspecified
[64]	Heart disease detection	RNN	Acc = 99%	Unspecified
[65]	Possibility of heart diseases prediction	DNN	Acc = 99% Spe = 99% Sen = 99%	Unspecified
[66]	Heart failure diagnosis with preserved ejection fraction	CNN	Sen = 84% Spe = 81% AUC = 95%	Unspecified
[67]	Prediction of acute heart failure	CNN	AUC = 81%	Unspecified
[68]	Classification of heart failure subtypes	CNN	Sen = 100% Spe = 94%	Unspecified
[69]	Heart failure acutely decompensated prediction	CNN	Acc = 94% Sen = 79% F1 = 85%	Unspecified
[70]	Detection of hypertrophic cardiomyopathy by EGC	CNN	AUC = 98% Sen = 92 Spe = 95%	Unspecified
[71]	Detection and classification of left ventricular hypertrophy.	CNN	AUC = 95%	Unspecified
[72]	Heart disease classification	LSTM	Acc = 91%	Unspecified

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, area under the ROC curve.

presents the research findings on diagnosing cardiac disease with DL models, and

Table 6. DL models implemented for cardiovascular diseases diagnosis.

Ref.	Contribution	DL Model	Performance	Validation
[73]	Detection of severe aortic stenosis	CNN	AUC = 97%	Internal
[74]	Aortic Stenosis diagnosis by echocardiography	ANN	-	Internal
[75]	Detect aortic stenosis using ECG data	CNN	Acc = 97% Sen = 98% Spe = 96%	Unspecified
[76]	Screening for aortic valve stenosis using ECG data	CNN	Acc = 74% Sen = 74% Spe = 78% AUC = 85%	Unspecified
[77]	Detection of subclinical Atrial Fibrillation	CNN	Acc = 83% Sen = 82% Spe = 83% AUC = 90%	Unspecified
[78]	Atrial Fibrillation detection model	CNN	AUC = 79%	Unspecified
[79]	Detection of Atrial Fibrillation	DNN	Acc = 99% Sen = 99% Spec = 99%	Unspecified
[80]	Cardiovascular feature data extraction	CNN	Sen = 96% Spe = 92% ROC = 98%	Unspecified
[81]	Prediction of heart disease	ANN	Acc = 73%	Unspecified
[82]	Cardiovascular disease prediction	LSTM	P = 95% Sen = 92% F1 = 95%	Unspecified
[83]	Cardiovascular disease existing prediction	CNN	Acc = 97%	Unspecified
[84]	Polish summary texts of patient hospitalizations	CNN	Acc = 78%	Unspecified
[85]	Cardiovascular disease detection	CNN	Acc = 99%	Unspecified
[86]	Cardiovascular disease detection	ANN	Acc = 95%	Unspecified
[87]	Cardiovascular disease detection	ANN	Acc = 73%	Unspecified
[88]	Detection of Patients Requiring Invasive Coronary Angiography	CNN	Acc = 80% AUC = 87%	Internal
[89]	Myocardial infarction detection using ECG	VAE	AUC = 72%	Unspecified
[90]	Myocardial ischemia detection	ANN	Sen = 88% Spe = 86%	Unspecified
[91]	Myocarditis diagnosis	VAE	Sen = 78% Spe = 92%	Unspecified
[92]	Valvular heart disease by echocardiographic assessment	CNN	AUC = 99%	Unspecified
[93]	Detection of left ventricular hypertrophy	CNN	AUC = 88%	Internal
[94]	Severity Aortic Stenosis prediction	CNN	Acc = 93%	Internal
[95]	Pulmonary hypertension prediction from computed tomography images	CNN	Acc = 85% P = 86% Sen = 86% F1 = 85%	Internal
[96]	Coronary artery disease prediction by tongue image analysis	CNN	Acc = 77% F1 = 60% AUC = 57%	External
[97]	Valvular heart disease prediction	RNN	Sen = 72% Spe = 82% AUC = 83%	External
[98]	Arrhythmia detection from ECG signals	CNN	Acc = 95% F1 = 93% AUC = 96%	Internal
[99]	Pulmonary arterial hypertension prediction from echocardiography	CNN	AUC = 79%	External

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, area under the ROC curve.

presents the studies where DL models are implemented for cardiovascular diseases.

Table 5 presents a comprehensive summary of studies pertaining to the implementation of DL models in the diagnosis of cardiac diseases. Notably, studies [57,65] are highlighted for their contributions to the categorization analysis of electrocardiogram data using rhythm or beat features and the prediction of the possibility of heart diseases, respectively. These studies demonstrate the potential of DL models in predicting the likelihood of heart disease.

It is important to acknowledge that both models presented satisfactory performance metrics. However, the internal or external validation conducted to assess the efficacy of these models remains unexplored.

In Table 5 and Table 6, 47 studies are presented that focus on DL-based models for heart diseases diagnosis. Most of the reported works implementing DL models utilize CNN. This preference stems from CNNs utility, especially in imaging, as they are deep algorithms capable of detecting patterns that other models cannot find.

The performance metrics generally exceed 80%, with some models achieving accuracies and precisions exceeding 95%. However, studies employing smaller or imbalanced datasets often exhibited low performance in classification tasks.

In Figure 5 a graph is displayed that gives the information related to the average accuracy by DL model employed in the diagnosis of heart diseases, the percentage of studies implementing each model, and the mean publication year.

As shown in Figure 5, CNN models are the most widely used for this type of diagnostic tasks, along with other models such as ANN and LSTM. Furthermore, the RNN models have a better performance compared to other algorithms, and these types of models are also the most recently implemented in 2025. This is attributable to the architectures, whereby RNNs demonstrate enhanced robustness in the processing of sequential and temporally corelated data in order to integrate contextual information over time.

DL models are emerging as a powerful tool for discovering new knowledge in healthcare, providing highly favorable metrics and achieving increasingly innovative discoveries.

3.3. AI in Heart Diseases Prognosis

Prognoses in clinical practice involve predicting the likely course of disease progression. This process includes identifying symptoms and signs related to a specific illness or condition, as well as assessing whether these indicators will worsen, improve, or remain stable over time. It also involves recognizing any potential health issues or complications that might arise, along with evaluating the patient’s ability to perform daily activities and their survival prospects. In clinical settings, a wide range of patient data is collected, such as phenotypic traits, genetic information, proteomic profiles, pathology test results, and medical imaging [100]. Prognosis has two stages: first, the outcome, and then treatment. In simple terms, the outcome is the actual result of a disease or medical intervention, which can be measured by survival, symptom relief, quality of life, and other factors. Treatment includes any intervention for preventing or managing a disease, such as medication, surgery, radiation therapy, physical therapy, and other methods.

The field of predictive analytics focuses on using mathematical models and statistical techniques to forecast future events. AI systems can predict disease outcomes, especially in chronic ischemic heart disease, by analyzing various patient parameters [101]. In

Table 7. AI models and other technologies used in cardiac diseases prognosis.

Ref.	Contribution	Technique	Performance	Validation
[102]	Prognosis of major adverse cardiac events	LR	Acc = 87% AUC = 92%	External
[103]	Prognosis of heart failure	FT-Transformer	AUC = 74%	Internal

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, area under the ROC curve.

and

Table 8. AI models and other technologies used in cardiovascular diseases prognosis.

Ref.	Contribution	Technique	Performance	Validation
[104]	Prediction of cardiovascular disease risk	Swin Transformer	Sen = 0.81% Spe = 66% R2 = 0.5 MAE = 1.58	Unspecified
[105]	Cardiovascular outcomes	CNN	AUC = 0.94%	Unspecified
[106]	Cardiovascular prognosis in dystrophinopathy patients	AI-based	AUC = 91%	Internal
[107]	Prognosis of coronary artery disease	ANN	-	Unspecified
[108]	Prognosis of cardiovascular events	KNN	AUC = 79%	Unspecified
[109]	Prognosis of peripheral artery	NLP	AUC = 88%	Unspecified
[110]	Relationship between severity of peripheral arterial disease and symptom severity	RF	AUC = 68%	Unspecified
[111]	Rehabilitation monitoring of stroke	IoT	-	Unspecified
[112]	Post-stroke home rehabilitation	IoT	-	Unspecified

Acc, accuracy; P, precision; Sen, sensitivity; Spe, specificity; F1, F1-Score; AUC, area under the ROC Curve; R², coefficient of determination.

, twelve research studies are reviewed where AI models and other emergent technologies have been implemented for heart diseases prognosis.

It is important to acknowledge the lack of research into the prognosis of heart diseases, as it is a relatively new area of research. Therefore, the presented studies are limited to the most pertinent and relevant research conducted in this field.

The above Table 7 and Table 8 present novel research on heart-related disease prognosis using various emerging tools. Traditional AI methods are evident, as well as other technologies such as IoT, which currently drive a trend in the development of systems that incorporate both emerging technologies. Prognostics can range from risk prediction for a patient to treatment personalization. The latter involves AI models capable of providing personalized treatment based on each patient’s needs, symptoms, and conditions.

Specifically, the IoT is a technology that facilitates the connectivity of devices via the internet for patient monitoring. This functionality is achieved by transmitting data to the cloud, which subsequently serves as input in AI models to generate a prognosis. Furthermore, the IoT contributes to the field of telemedicine by enabling medical specialists to monitor patients remotely.

4. Discussions

The most significant advancements in AI-based models reported in this review focus on early detection of heart-related diseases, which encompasses a wide range of cardiovascular and cardiac problems. Notably, the models proposed by the authors span from simple regression models to more complex models, such as ANNs.

Most studies presented in this review are based on binary classification tasks, which enable them to predict whether a patient can be diagnosed with a heart-related disease. These models can serve as valuable tools for medical specialists in supporting their diagnostic processes.

Based on the reported studies through this review, the papers could be classified into two primary paradigms: signal or tabular data analysis and image analysis. For the case of signal or tabular data analysis, traditional ML methods such as DT, NB, LR, RF, KNN and SVM are implemented. Additionally, more advanced ML techniques, including ensemble or boosting models are employed. DL methods like ANNs, CNNs, LSTM, RNN and XAI, are also used, with ECG data being the dominant approach for physiological signals. On the other hand, image analysis is predominantly handled by 2D or 3D CNNs, fusion models, and ensemble models, reflecting the spatial complexity of echocardiographic and CT imaging data.

The findings based on the analysis of the trend of the years reveals that algorithms that are not implemented frequently tend to show a low performance, which may indicate a gradual transition towards more specialized or sophisticated AI algorithms. In addition, between 2023 and 2025, the models are associated with obtaining better performance metrics, indicating the progress in the development of AI models. This is due to the enhanced robustness of the new architectures employed in relative new AI models, such as ensemble models and RNNs, which are designed to capture better complex patterns and temporal dependencies, helping to mitigate overfitting and obtain an enhanced generalization across diverse data types and distributions.

The studies reported in this review indicate that most of the proposed models achieve acceptable performance, demonstrating their potential. However, it is crucial to emphasize that they require rigorous testing in real clinical settings to ensure their inclusion in healthcare systems. Several studies have reported metrics exceeding 95%, but F1-score, false positive and false negative rates are not provided, which is crucial in clinical settings. Presenting these metrics facilitates the understanding that false diagnostic cases can also occur in the clinical context, where the AI model fails to identify an existing heart disease [113,114].

A limited number of studies provide comprehensive evaluations of all model metrics. This comprehensive analysis is essential to ascertain whether a model is capable of making correct decisions. Furthermore, if all evaluation metrics are not analyzed, there is a risk that the model may exhibit poor generalization, potentially leading to underfitting or overfitting. In this context, reporting the F1-score is essential for a balanced evaluation of an algorithm’s performance, particularly in scenarios where accuracy alone can be misleading. By incorporating both sensitivity and precision, the F1-score provides a robust measure of reliability, making it indispensable for assessing AI applied to medical diagnosis and prognosis.

Regarding model generalization, it is noteworthy that more intricate models, such as ensemble models or ANNs, generally exhibit superior performance. However, these kinds of models offer less interpretability compared to simpler models, including regressions and decision trees, among others. In this context, AI utilized in healthcare must undergo rigorous evaluation by medical specialists and adhere to stringent standards to ensure clinical reliability. To ensure the accuracy of the model, techniques such as Shapley Additive Explanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Gradient-weighted Class Activation Mapping (Grad-CAM) can be employed. XAI has been employed to enhance interpretable predictive models [115], conduct EGC analysis [116], predict future significance [117], and ensure the reliability of heart disease diagnoses.

It is noteworthy that biases can manifest in AI models due to the data utilized for their training. This aspect is frequently overlooked, yet it holds paramount importance, particularly within the medical domain, producing biases pertaining to inclusion, such as the representation of ethnic and gender groups. Consequently, it is imperative to establish a traceability plan and devise a strategy for addressing the bias issues that AI may encounter during the generalization phase. Future research endeavors must ensure the inclusivity of the dataset, meticulous data processing, and the inclusion of an analysis prior to training the models.

5. Challenges and Trends

AI is an emerging technology that has experienced rapid growth in recent years, which reflects a clear trend; it is making important technological advances in specific areas and becoming more widespread. Incorporating AI into disease prognosis and diagnosis presents significant challenges, but it is clear that progress has already been achieved in this field. Due to recent expansion, several new terms, including eHealth, telemedicine, telecardiology, and the Internet of Medical Things (IoMT), have appeared in healthcare [118,119,120,121].

With the increasing prevalence of heart diseases, emotional illnesses can combine to cause serious complications. There is a challenge that when these two illnesses are combined, the chance of recovery becomes much slower [122,123,124,125]; how can AI help address this problem?

However, challenges like data accessibility, model interpretability, and the ability to generalize across diverse populations still hinder progress. Ongoing research aims to address these issues, ensuring that AI can be effectively integrated into clinical practice and ultimately lead to improved patient outcomes in managing heart diseases.

These challenges can be addressed by employing reinforcement learning models and federated learning to enhance predictive accuracy while preserving patient privacy. Additionally, federated learning can facilitate collaboration between models trained on decentralized datasets. Although federated learning offers substantial benefits in heart disease diagnosis, several challenges persist. These include achieving model convergence and developing robust algorithms capable of effectively managing the complexities inherent in decentralized data.

The trends show a significant increase in advancements such as diagnostic precision, early detection, risk stratification, prediction of mortality and readmission, real-time monitoring, integration of multimodal data, wearable technology, and biomarkers. For example, telecardiology is poised to revolutionize cardiology by offering personalized and remote healthcare to patients with heart issues. This method allows doctors to diagnose and treat heart conditions while also providing education and care to their patients [126].

In addition, the trends suggest a promising future for AI in heart diseases treatment, with the potential to enhance diagnosis, prognosis, and personalized care strategies. A notable trend is developing Retrieval-Augmented Generation (RAG) and Large Language Model-based (LLM) applications and chatbots that can serve as medical assistants, utilized to support healthcare professionals [127,128,129]. However, it is essential to note that these tools will not replace medical specialists; instead, they will serve as aids in diagnosing and prognosing diseases.

5.1. Limitations

The development of AI-based models focused on cardiology may have technical, clinical, and ethical limitations, including issues of data quality, clinical validation, bias, privacy, and model explainability. These limitations can directly impact the reliability and reproducibility of clinical cardiology procedures.

In addition, clinical validations are crucial for integration into real-world settings, thereby enabling physicians to access these tools. However, these processes are often time-consuming and even require financial resources.

Ethical and legal issues could limit the credibility of patients and physicians, where it is evident that operational, regulatory, and human resources processes could represent obstacles to the process of validating the AI tool, impacting directly on the scaling, adoption, and sustainability in clinical practice.

5.2. Research Gaps and Future Roadmap

Despite the advancements in AI, there remain numerous opportunities within the research and development of AI-based models for the diagnosis and prognosis of heart-related diseases. While AI applications have experienced exponential growth, data remains a significant obstacle. This is primarily due to the reliance on public databases, which may not contain authentic data obtained from clinical settings. Consequently, the integration of models into real-world contexts faces limitations.

Understanding the decision-making process of an AI can be intricate for humans, as some models may appear complex. To achieve this, advanced techniques such as XAI are essential to elucidate the decision-making mechanisms of models and ensure that predictions are commensurate with reality. On the other hand, ethical, regulatory, and legal concerns continue to be debated. However, it is imperative that these issues be prioritized to facilitate the appropriate development and implementation of AI-driven applications.

To ensure that AI in diagnostics and prognostics truly possesses the transformative potential and can address research gaps, a multifaceted and strategic approach is necessary. In this regard, a roadmap is proposed that prioritizes data ecosystems through real-time acquisition, multimodal data, and data pertaining to federated learning and democratization.

Furthermore, it is suggested to prioritize the development of interpretable AI-based models, applied to health issues under strict supervision and within a multidisciplinary approach, involving not only developers and AI experts, but also medical specialists. This approach can help minimize bias and ensure that all ethical, legal, and regulatory provisions are in place, enabling the models to be introduced into clinical settings and guaranteeing patient safety.

In this case, to address the challenges posed by the application of AI in healthcare and the real-world validation, including methodological quality and the potential for bias, it is recommended to utilize standardized reporting frameworks such as standards for reporting diagnostic accuracy (STARD-AI), transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD-AI), prediction model risk of bias assessment tool (PROBAST), and others. These frameworks can significantly enhance the transparency and reproducibility of AI research in clinical settings [130,131,132].

6. Conclusions

This review presents a novel contribution by providing a comprehensive overview of both diagnosis and prognosis tasks in heart diseases through the utilization of AI models. This approach offers a broader perspective compared to other surveys that only focus on diagnostic tasks, thereby displaying the most implemented ML and DL algorithms. Furthermore, the review conducts a comprehensive comparative analysis of the most recent and pertinent works, emphasizing a general summary of the strengths, limitations, and challenges associated with AI models. Additionally, it underscores the significance of validation practices in validating AI models within a real-world clinical setting. Moreover, it emphasizes the importance of incorporating XAI to facilitate both model transparency and interpretability.

The evaluation metrics employed by the majority of the authors of each study are presented, which are generally the most widely utilized for assessing model performance, including accuracy, precision, recall, F1-score, ROC, and AUC. In addition, some of the potential limitations that may arise when integrating AI into the healthcare domain are discussed.

AI has the potential to improve the diagnosis and prognosis of heart diseases, demonstrating high accuracy in detecting heart diseases using various imaging modalities and data. Moreover, AI can predict the risk of adverse cardiac events, enabling more precise stratification of patients and customization of treatment strategies. Nevertheless, AI possesses the potential to transform heart healthcare and enhance clinical outcomes, leading to a decrease in mortality; this clearly presents itself as a significant problem to solve. The timely diagnosis of a disease, no matter how severe, is crucial in reducing both the number of fatalities and the severity of the illness. AI models hold promise in achieving these objectives and can be integrated into medical devices and a personalized treatment.

The AI models focused on making medical prognoses and diagnoses undoubtedly assist doctors and medical specialists by increasing precision. In this context, it is important to note that while these systems are useful, they cannot replace the expertise of specialists who use their knowledge and experience to determine the presence or absence of a pathology.

The use of AI models in the diagnosis and prognosis of heart diseases presents several challenges and opportunities for future research. Key challenges that need to be addressed include data accessibility and quality, model interpretability through XAI for better understanding the generalization process and predictions, and generalizability across diverse populations. To address these issues, future directions focus on integrating multimodal data from electronic health records, imaging, and wearables to enhance predictive accuracy, as well as advancing DL techniques like transfer learning and self-supervised learning to improve performance with smaller datasets.

To substantiate the clinical impact of AI models in real-world settings, rigorous validation in large prospective studies is imperative. This highlights the importance of interdisciplinary collaboration among clinicians, data scientists, and healthcare systems to drive advancements in this field.

Based on the findings presented throughout this review, a comprehensive roadmap of how challenges could be addressed is proposed, for the effective implementation of AI models in the diagnosis and prognosis of heart diseases. This framework encompasses the utilization of high-quality, real-world data that undergoes meticulous processing and a transparent evaluation of the model, ensuring that all pertinent evaluation metrics are meticulously reported within the clinical trial validation. In addition, it is feasible to integrate a regulatory framework, such as that established by the Food and Drug Administration (FDA), for the use of medical diagnostic devices that use AI, known as Artificial Intelligence-Enabled Medical Devices. Globally, there is a regulatory framework known as software as a medical device (SaMD), which covers all medical devices that use AI under this category, so that AI models can continue to learn and adapt after their integration.