Artificial intelligence in cardiology: a narrative review with focus on patient outcomes

Abstract

Background and Objective

Artificial intelligence (AI) is rapidly transforming cardiology through advancements in diagnostic accuracy, prognostication, and treatment personalization. While evidence for algorithmic performance is robust, its true impact on patient-centered outcomes remains unclear. This review aims to evaluate how AI applications influence patient outcomes in cardiology and identify current limitations and future directions.

Methods

A targeted literature search was conducted in PubMed, Scopus, Embase, and Cochrane databases on May 9 and 23, 2025, using a combination of terms related to AI, cardiology, and patient outcomes. Filters were applied to include human studies, English language, and studies published between January 2015 and May 2025. Two reviewers independently screened articles, and three reviewers reached consensus for final inclusion. A total of 11 studies met inclusion criteria.

Key Content and Findings

AI tools have demonstrated potential benefits across multiple domains, including clinical decision support, cardiac imaging, remote patient monitoring, and patient engagement. Evidence suggests AI can enhance diagnostic accuracy, procedural efficiency, and patient self-management. However, most studies report surrogate or process-related endpoints rather than hard clinical outcomes. Large-scale randomized trials remain scarce, and improvements in mortality, hospitalization, and quality of life (QoL) are inconsistently demonstrated. Ethical considerations, implementation challenges, and cost-effectiveness concerns persist.

Conclusions

AI in cardiology shows promise for improving patient care, but robust evidence linking its adoption to improved clinical outcomes is limited. By synthesizing available findings, this review highlights critical evidence gaps and provides guidance for future research, which should prioritize prospective trials focused on patient-centered endpoints and address barriers to implementation, transparency, and equity.

Introduction

Background

Artificial intelligence (AI) in the field of cardiology is here to stay, with a rapidly growing body of evidence supporting its integration into clinical practice (1-4). Innovations in disease diagnosis, digital biomarkers of disease risk and novel approaches for disease prognostication are some of the main areas that AI has increasing evidence of benefit in the field of cardiovascular care (5).

Rationale and knowledge gap

Despite these advancements, the impact of AI on actual patient outcomes remains an area of ongoing investigation. While numerous studies demonstrate improvements in diagnostics metrics, translating these gains into clinical benefits requires rigorous evaluation and real-world validation (4,5). As highlighted in recent analyses, substantial uncertainty remains regarding the clinical impact of these technologies. Additionally, this research builds upon previous studies by further exploring the cost-effectiveness and productivity paradox (6).

Objective

This narrative review aims to synthesize current evidence on the role of AI in cardiology with a focus on patient outcomes and identify existing gaps and future directions necessary to realize AI’s full promise in cardiovascular medicine. We present this article in accordance with the Narrative Review reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-479/rc).

Methods

On May 9 (researcher one—I.K.) and 23 (researcher two—J.M.O.), 2025, we searched PubMed, Scopus, Embase, and Cochrane databases using a predefined strategy (see Table 1 for complete search details). The search included studies published between January 2015 and May 2025. Filters were applied to include human studies, English language, and the following study types: randomized controlled trials (RCTs), clinical trials, comparative studies, observational studies, meta-analyses, systematic reviews, multicenter studies, and validation studies. The search terms used were the following: ((“artificial intelligence”[MeSH Terms] OR “machine learning”[MeSH Terms] OR “deep learning” OR “neural network*” OR “natural language processing” OR “computer vision”) AND (cardiology OR “cardiovascular disease*” OR “heart failure” OR “myocardial infarct*” OR arrhythm* OR “coronary artery disease*” OR “ECG” OR electrocard* OR echocard* OR “cardiac MRI” OR “cardiac imaging”) AND (“treatment outcome”[MeSH Terms] OR “patient outcome*” OR mortality OR readmi* OR “clinical outcome*” OR “treatment effectiveness”) AND (“real world evidence” OR “real world data” OR implementation OR deployment OR “clinical use” OR “clinical implementation” OR prospective OR pragmatic OR impact)).

Table 1. The search strategy summary.

Items	Specification
Date of search	May 9 (researcher one—I.K.) and 23 (researcher two—J.M.O.), 2025
Databases searched	PubMed, Scopus, Embase, Cochrane
Search terms used	((“Artificial intelligence”[MeSH Terms] OR “machine learning”[MeSH Terms] OR “deep learning” OR “neural network*” OR “natural language processing” OR “computer vision”) AND (cardiology OR “cardiovascular disease*” OR “heart failure” OR “myocardial infarct*” OR arrhythm* OR “coronary artery disease*” OR “ECG” OR electrocard* OR echocard* OR “cardiac MRI” OR “cardiac imaging”) AND (“treatment outcome”[MeSH Terms] OR “patient outcome*” OR mortality OR readmi* OR “clinical outcome*” OR “treatment effectiveness”) AND (“real world evidence” OR “real world data” OR implementation OR deployment OR “clinical use” OR “clinical implementation” OR prospective OR pragmatic OR impact))
Timeframe	January 1st 2015–May 1st 2025
Inclusion criteria	Only included human studies and articles written in English. Only included the following study types: Clinical Trial, Comparative Study, Controlled Clinical Trial, Meta-Analysis, Multicenter Study, Observational Study, Randomized Controlled Trial, Systematic Review, Validation Study
Selection process	Two researchers (I.K., J.M.O.) screened the results and created an initial selection. Three researchers (I.K., J.M.O., M.A.) curated the final selection through group discussion until consensus was reached

Two reviewers (I.K., J.M.O.) independently screened titles and abstracts to identify eligible studies. Full texts of potentially relevant articles were assessed for inclusion. Disagreements were resolved through group discussion, and three reviewers (I.K., J.M.O., M.A.) reached consensus on the final selection (Table 2). Studies focusing on algorithm development without clinical implementation or patient outcome assessment were excluded. Data extraction focused on study design, AI application type, clinical context, and primary patient outcomes.

Table 2. Studies included in the review.

First author, year	Objective	Study design	n	Intervention group	Control group	Outcomes
Luštrek, 2021 (7)	To evaluate a decision support system (HeartMan) for improving QoL and self-care in HF patients	Multicenter RCT	56	HeartMan personal health system + usual care	Usual care	Significant improvement in self-care, anxiety, depression; no difference in mortality risk
Zisis, 2021 (8)	To determine if an avatar-based HF-app improves outcomes by enhancing HF knowledge and QoL	RCT	72	Digital HF coach app + usual care	Usual care	No significant differences in readmissions or mortality; uptake limited by digital literacy
Persell, 2020 (9)	Effect of AI smartphone hypertension coaching app on BP control	Open RCT	333	Smartphone coaching app + BP monitor	BP tracking app + BP monitor	No significant difference in SBP at 6 months; improved self-confidence in intervention group
Aharon, 2022 (10)	To evaluate personalized AI-generated interventions for adherence to cardiac rehab	Prospective cohort vs. historical control	95+500	AI-generated recommendations and motivational messages	Standard CR program	Significant increase in adherence (76% vs. 24%, P=0.001)
Huang, 2021 (11)	Efficacy of AI-enabled portable ECG (BigThumb) for AF recurrence screening post-ablation	Single-center RCT	218	BigThumb ECG + AI algorithm + cardiologist confirmation	Holter monitor screening	Higher AF detection and anticoagulation adherence; improved diagnostic accuracy
Liu, 2018 (12)	Effects of ML-based CDSS on BP management and economic burden	Cluster RCT	430	CDSS-assisted HTN management	Standard HTN care	No significant difference in main analysis; improved BP control and reduced cost in surgical patients
De Backer, 2023 (13)	Impact of computational modeling on efficiency and outcomes of LAA closure	Multicenter RCT	181	AI-based CT simulation planning	Standard planning	Primary endpoint: 41.8% (standard) vs. 28.9% (AI); improved efficiency, fewer device repositioning
Bezerra, 2015 (14)	Robotic-assisted PCI vs. manual PCI: on LGM	Retrospective analysis of trials	1,673	Robotic-assisted PCI	Manual PCI	Lower incidence of LGM (12.2% vs. 43.1%, P<0.001)
Lin, 2024 (15)	Effect of AI-ECG alerts on mortality in hospitalized patients	Multisite RCT	15,965	Real-time AI report and warning to physicians	Delayed AI report	Significant reduction in 90-day mortality (3.6% vs. 4.3%, HR =0.83, 95% CI: 0.70–0.99)
Glessgen, 2025 (16)	To evaluate whether AI-based automated cardiac MRI planning improves scan efficiency and reduces procedural errors	Single-center prospective randomized trial	82	AI-based automated CMR scan planning	Manual CMR planning	Fewer scan errors (71% vs. 45%), reduced scan time by ~6 minutes, increased efficiency
Upton, 2024 (17)	To test if AI-augmented stress echocardiography decision-making is non-inferior to standard clinician-only interpretation in selecting patients for angiography	Multicenter, non-inferiority RCT	2,341	AI-assisted (EchoGo Pro) stress echo interpretation	Standard clinician interpretation	Did not meet non-inferiority margin for appropriate referrals; lower referral rate than expected; possible utility in low-volume centers

AF, atrial fibrillation; AI, artificial intelligence; BP, blood pressure; CDSS, clinical decision support system; CI, confidence interval; CMR, cardiac magnetic resonance; CR, cardiac rehabilitation; CT, computed tomography; ECG, electrocardiogram; HF, heart failure; HR, hazard ratio; HTN, hypertension; LAA, left atrial appendage; LGM, Longitudinal Geographic Miss; ML, machine learning; MRI, magnetic resonance imaging; PCI, percutaneous coronary intervention; QoL, quality of life; RCT, randomized controlled trial; SBP, systolic blood pressure.

Clinical integration of AI and observed outcomes

Treatment optimization and clinical decision support

AI’s capacity to analyze intricate datasets and offer tailored recommendations can greatly enhance patient outcomes by personalizing treatments and refining the precision of medical procedures (18). Furthermore, it holds the potential to augment the efficiency and timeliness of healthcare delivery through the utilization of prognostication and prediction algorithms (19). Clinical decision support system (CDSS) tools, enhanced by AI, have the potential to revolutionize cardiology by assisting physicians in making more informed and timely decisions, thereby improving clinical accuracy, reducing errors, and ultimately, enhancing overall patient care. However, the majority of current AI initiatives have been focused on the development and validation of algorithms designed to phenotype, cluster, or prognosticate patients based on specific interventions (20). Studies examining real-world patient outcomes from the implementation of these tools remain limited.

The RAPIDxAI trial employed a randomized, multisite design across 12 hospitals in South Australia to evaluate the efficacy of an AI-based CDSS for the management of patients with suspected cardiac chest pain in the emergency department. Six hospitals were randomly allocated to use the AI tool, while the remaining six continued with standard care (21). The CDSS tool utilized machine learning (ML) models to provide objective patient-specific diagnostic probabilities, prognostication and evidence-based clinical recommendations based on the most likely diagnosis. While the primary outcome, a composite of cardiovascular death, myocardial infarction, and unplanned cardiovascular readmission at 6 months, showed no difference between the study arms, the AI-implementation group had higher rates of evidence-based therapies such as statins and anti-platelet agents (21).

Another cluster-randomized trial involving 430 hospitalized patients with hypertension, examined whether implementing an ML, graph-based, electronic medical record (EMR)-integrated algorithm could improve blood pressure (BP) control or improve the financial burden of disease (12). While the primary analysis found no difference between the intervention and control groups, a subgroup analysis revealed a statistically significant improvement in BP control and reduced economic burden for patients admitted to the surgical service, with a benefit-cost ratio of 1.44, or 18,186 yuan in absolute terms.

Two RCTs, each involving 56 patients, examined the effectiveness of the HeartMan decision support system (7,22). The investigators combined advanced technologies with a human-centered design, in order to create a personal health system designed to help patients better manage their congestive heart failure (HF) (23). The findings demonstrated that the intervention successfully enhanced self-care behavior, leading to improved disease management, as evidenced by an 11% increase in the Self-Care of HF Index. Additionally, the intervention significantly improved psychological outcomes, with reduced rates of depression, anxiety and perceived sexual problems (22). No effect was observed on illness perception, health-related quality of life (QoL) or exercise capacity. Notably, a subgroup analysis revealed a significant improvement in the left ventricular ejection fraction (LVEF) in the intervention group, along with a reduction in the predicted 1-year mortality risk score based on MAGGIC and 3C-HF.

Cardiac imaging

AI applications in cardiac imaging have expanded rapidly, particularly in enhancing image quality, interpretation speed, and diagnostic accuracy (24-26). This domain is probably the one with the highest potential for AI integration. In a recent survey, most interventional cardiologists agreed that the field in cardiology that will benefit the most from AI is advanced cardiac imaging (27). While these innovations are well-documented, far fewer studies have evaluated whether such improvements translate into measurable benefits in patient outcomes.

De Backer et al. performed an RCT with 200 patients with nonvalvular atrial fibrillation (AF) that underwent left atrial appendage (LAA) closure, either through transcatheter computer tomography (CT) simulation-based planning, or a standard planning without AI (13). The primary endpoint composite was an incomplete LAA closure with residual contrast leakage into the LAA and/or presence of a device related thrombus (DRT). In the standard planning arm, the composite primary outcome occurred in 41.8% of patients, compared with 28.9% in the CT simulation-guided group [relative risk (RR) 0.69, 95% confidence interval (CI): 0.46–1.04; P=0.08]. Rates of complete LAA closure without residual leak or disc retraction were 44.0% in the standard arm vs. 61.1% with CT simulation (RR 1.44, 95% CI: 1.05–1.98; P=0.03). In addition, AI-supported planning was associated with greater procedural efficiency, requiring fewer Amulet devices (103 vs. 118; P<0.001) and fewer device repositionings (104 vs. 195; P<0.001), suggesting improved workflow and a possible benefit in procedural outcomes.

A retrospective evaluation of the angiographic images from the PRECISE trial, which utilized the robotic assisted percutaneous coronary intervention (R-PCI), and the STLLR trial, which used standard manual stent deployment, and compared the incidence of Longitudinal Geographic Miss (LGM) between both (14). Overall, patients treated with R-PCI had 72% lower incidence in LGM than patients with M-PCI (12.2% vs. 43.1%, P<0.001).

A multisite RCT with 15,965 patients and 39 physicians, evaluated the ability of an AI-enabled electrocardiogram (ECG) to identify hospitalized patients with a high risk of mortality (15). They found that in the intervention group, where a real time report and warning messages through AI-ECG alert was sent to the physicians, was associated with a significant reduction in all-cause mortality within 90 days, with a 3.6% mortality in the intervention, compared with 4.3% in the control group [hazard ratio (HR): 0.83; 95% CI: 0.70–0.99].

A study by Glessgen et al. prospectively evaluated the impact of an AI-based automated cardiac magnetic resonance imaging (MRI) planning software on procedure errors and scan times compared with manual planning (16). Eighty-two patients undergoing non-stress cardiac magnetic resonance (CMR) were randomized into manual or automated scan execution. Automated procedures had fewer errors: 71% vs. 45% without automation. Automated scans significantly reduced examination times by approximately 6 minutes (17%) in free-breathing examinations and increased efficiency by reducing the idle portion of examinations in both breathing strategies.

The PROTEUS trial was a multicenter, parallel-group RCT designed to evaluate whether AI interpretation of stress echocardiography could augment clinician selection of participants for coronary angiography (17,28). It involved adults undergoing a clinically indicated stress echocardiogram to assess significant coronary artery disease. The pre-specified primary end point was defined as appropriate referral for coronary angiography with true positives defined as referrals leading to revascularization and false negatives defined as non-referred patients that suffered an acute coronary event in the pursuing 6 months. The study found that the difference in the area under the receiver operating characteristic curve (AUROC) between the intervention and control groups did not meet the pre-specified non-inferiority margin, indicating uncertain benefit of AI assistance that did not significantly improve the accuracy of stress echocardiography in identifying patients who would benefit from coronary angiography.

Remote patient monitoring

Remote and accurate patient monitoring is increasingly recognized as crucial in several realms of cardiovascular medicine as AI powered tools can aid in early detection of anomalies, provide personalized alerts, engage healthcare providers in a timely manner and reduce gaps in care (29). This could be especially valuable for rural communities (30). AI-driven remote patient monitoring has increasingly been incorporated into wearables, for both screening and management of cardiovascular conditions (31). Patients express a positive attitude regarding AI use in healthcare, with recent results of a cross-sectional survey that included 13,806 patients showing that the attitudes of patients towards AI are highly variable depending on demographic characteristics, health status and technological literacy (32).

An RCT involving 297 adult patients with uncontrolled hypertension, evaluated the effect of a conversational AI smartphone coaching app in improving systolic BP levels, compared with a BP tracking app (33). The AI app utilized cognitive behavioral therapy techniques to promote home monitoring and behavioral changes, based on participants’ prior data and responses. The study did not find significant difference in the primary outcome of systolic BP between the groups at 6 months, nor in most of the secondary outcomes, with the exception of improved self-confidence in controlling BP in the intervention group (9).

A single-center RCT evaluated the use of an AI algorithm implemented on a handheld single-lead ECG monitor (BigThumb^®) vs. traditional follow-up for detecting non-valvular AF recurrence after catheter ablation (11). The study, involving 218 patients (109 in each group), found that using the BigThumb AI algorithm led to more frequent detection of AF recurrence, measured as 64.2% free from AF recurrence in the intervention group vs. 78.9% in the control group. Also, adherence to oral anticoagulation at the end of follow-up was higher in the intervention group compared to the control group with 51% and 25.4% respectively. While the AI algorithm was found to improve the accuracy of ECG diagnosis, the study’s relatively small size did not allow for enough power to detect differences in low-rate complications like thromboembolism and bleeding. There was no significant impact on hard clinical outcomes, a gap that future studies should prioritize addressing. Participants’ compliance with the device was observed to diminish over time, with rates in the first 3 months of use being significantly higher than the post-3-month mark.

Patient engagement

Empowering patients to take a more active role in managing their heart health is essential for achieving optimal outcomes. A plethora of AI applications has already been developed, aiming to improve patient education, adherence to treatment plans and communication between patients and their treatment teams, of which some are discussed below (34). The 2023 American Heart Association Science Advisory highlights the promise of these innovations, especially in expanding access and supporting long-term engagement in cardiac care. Digital technologies, both synchronous and asynchronous, can integrate AI and provide personalized education towards a shift from passive care models to tailored interactions (35). Recent reviews have also highlighted the potential of conversational AI systems such as ChatGPT to enhance patient education, improve communication, and assist clinicians in evidence-based decision-making within cardiovascular care (36). Increasingly, research is shedding light on how AI can not only streamline care delivery but also strengthen the patient’s role at the center of it (37).

Aharon et al. investigated whether personalized interventions could increase patient adherence to cardiac rehabilitation (10). By utilizing an AI-based engine that generated recommendations and messages, investigators achieved improved patient adherence to the cardiac rehabilitation program with 76% of participants in the intervention arm remaining active at 3 months, compared to 24% in the control arm. Moreover, the intervention was perceived as beneficial by 97% of the healthcare providers that took part in the project.

The study by Zisis et al. examined an m-Health intervention for patients admitted with acute decompensated HF, focusing on education, self-management, and outcomes (8). While the study was not able to draw any statistically significant clinical conclusions, it shed light on potential barriers and problems that will need to be addressed. The authors found that cognitive impairment, older age, limited English proficiency, lower education levels, and advanced HF with multiple comorbidities may impact enrollment and engagement in m-Health programs. The study concluded that m-Health-based education and monitoring delivered via an HF app is promising but challenging, especially in older patients with recent HF instability.

Cost-efficiency and the productivity paradox

One of the promised benefits of AI in cardiology is cost-efficiency (38). However, these theoretical gains are often counterbalanced by the productivity paradox, which is the phenomenon where significant technological advancements do not immediately translate into improvements in productivity (39). In healthcare, this paradox arises from factors such as high initial implementation costs and the need for healthcare worker training. The initial complexity introduced in the workflow can increase the burden before efficiency benefits are fully realized. To avoid prolonged stagnation in the so-called “valley of death” on the productivity curve, it is critical to redesign workflows early (39). While this concept is primarily discussed as an economic framework, quantitative cost-effectiveness analyses will be an essential next step to determine the real financial impact of AI integration in clinical practice.

To date, few studies have provided real-world economic evaluations of AI implementation in cardiology. Most available analyses are based on modeled projections or limited pilot data, which may not capture the true costs and savings associated with large-scale adoption. Future research should incorporate pragmatic cost-effectiveness studies assessing not only direct healthcare expenditures but also workflow efficiency, training requirements, and long-term return on investment across diverse clinical settings. AI should deliver on its promise of improving both clinical outcomes and economic sustainability (40).

Ethical considerations

Ethical challenges must be addressed when it comes to deepening the integration of AI in the field of cardiology and especially when studying the impact on patient outcomes (41). A major concern is balancing automation with clinician oversight. Excessive reliance on automated systems risks undermining clinical judgment and accountability. In a recent survey on AI in interventional cardiology, the majority (73.5%) of physicians agreed that they know too little about AI to apply it on patients (27). To address the “black box” nature of these AI technologies, developers must prioritize creating models that are interpretable and transparent, while physicians need targeted training to understand, evaluate and appropriately integrate AI tools into clinical practice (42). An increasingly debated question is whether a physician should be held accountable if an AI model provides highly accurate guidance, yet the clinician disregards it and makes an opposing decision based on clinical judgement, particularly if the outcome is unfavorable (43). Clear legal frameworks will be essential to permit implementation of AI in clinical practice.

Current limitations and future directions

Despite rapid advancements, current evidence supporting the use of AI in cardiology is largely retrospective and focuses on surrogate endpoints rather than direct patient outcomes. Prospective trials that evaluate the clinical impact of AI interventions remain scarce (2). Generalizability is also a concern, as many AI models are trained on data biased toward specific populations and healthcare systems, making unsure the performance of these in diverse real-world settings (44). Furthermore, most of the available studies demonstrated only modest or inconsistent improvements in clinical outcomes compared with standard care (12,13,21,28). While these findings suggest that AI has potential to enhance decision-making and procedural efficiency, its translation into tangible and reproducible patient benefits remains limited (Figure 1).

Figure 1.

Artificial intelligence in cardiology: a focus on patient outcomes. Figure created by the authors using Canva. AI, artificial intelligence; RCT, randomized controlled trial.

In addition, most of the available studies involved relatively small patient populations, which limits statistical power and the generalizability of their findings. Another challenge is the uncertainty regarding the long-term sustainability of these AI implementations. Many reports do not clarify whether the tools remained in active use beyond the study period. Moving forward, the primary goal must be to generate patient outcome-focused evidence. Prospective RCTs are urgently needed to determine whether AI applications in cardiology can improve hard endpoints such as mortality, hospitalization rates, and QoL (45,46). Another promising direction lies in hybrid models that combine AI with clinician expertise, that leverage the computational power of AI without sacrificing clinical judgment and extending into emerging domains such as AI-guided electrophysiology and arrhythmia risk stratification, areas where current studies focus on projected capabilities rather than demonstrated real-world outcomes.

Study strengths and limitations

This narrative review synthesizes current studies on the applications of AI in cardiology, with a focus on patient outcomes, a critical and clinically relevant perspective. However, as a narrative review, it is subject to inherent limitations of this study design, including potential selection bias and the lack of a systematic methodology. Future systematic reviews following established reporting frameworks could strengthen the evidence synthesis in this evolving field. Another limitation of this review is the small number of eligible studies identified, which underscores the need for more large-scale investigations. Furthermore, heterogeneity among the included studies and the rapid evolution of AI technologies may limit the generalizability and long-term applicability of the conclusions.

Conclusions

The implementation of AI in the field of cardiology shows great potential. However, robust evidence of its true clinical impact remains limited. More research is needed to determine whether these technologies translate into meaningful improvements in patient outcomes such as mortality, hospital stay and symptom improvement. Additionally, addressing challenges related to ethical use, transparency, and cost-effectiveness will be essential for sustainable integration into clinical practice. By bringing together the currently available prospective and randomized evidence on AI-driven interventions with reported patient outcomes, this review can help clinicians and researchers, prioritize high-value use cases, design outcome-focused trials, and develop implementation strategies that align AI deployment with benefits for patients.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.