Big Data and Trustworthy AI for Heart Failure: A Review

Abstract

The rapid evolution of machine learning techniques, combined with the growing availability of large and diverse datasets, is poised to transform heart failure research and clinical care. This review first provides an overview of key machine learning and artificial intelligence concepts used in heart failure research and then examines how diverse data modalities—including electronic health records, patient registries, biobanks, imaging, telemonitoring, and synthetic data—are leveraged to develop machine learning applications for heart failure diagnosis, prognosis, risk stratification, and personalized treatment strategies. While the potential is considerable, we highlight key barriers to clinical translation, such as data heterogeneity, algorithmic bias, lack of interoperability, and privacy concerns. The review also examines the need for explainable and equitable artificial intelligence systems and evaluates emerging solutions, including Federated Learning and synthetic data generation to address fairness and data privacy challenges. Beyond technical innovations, we underscore the importance of human-centered design, stakeholder engagement, and regulatory readiness. We conclude by identifying future priorities and calling for interdisciplinary collaboration to ensure the scalable, ethical, and effective integration of AI in heart failure management.

Heart failure (HF) is among the most common, costly, and life-threatening cardiovascular conditions, representing a major global health challenge. It is the leading cause of hospitalization in people over 65 years of age and is projected to increase in prevalence by 46% by 2030 due to population aging and lifestyle factors.¹ Although diagnostics and therapeutics have advanced, outcomes in HF remain suboptimal. High rates of readmission—often driven by comorbidities, polypharmacy, and functional limitations—contribute to significant clinical and economic burden. Moreover, the variability in HF trajectories complicates prevention and management, and studies show little improvement in mortality rates over the past 2 decades. For example, Conrad et al¹ reported that HF burden in the United Kingdom now rivals that of the 4 most common cancers combined, despite a modest decline in incidence.

An emerging solution generating optimism is the potential application of machine learning (ML), including deep learning (DL), in HF combined with the growing availability of large datasets. The application of ML in HF has evolved substantially over the past decade. Early studies in the 2010s primarily leveraged structured electronic health records (EHRs) data to predict incident HF, readmission, and mortality, demonstrating the potential of ML to outperform traditional risk models.² A major conceptual milestone followed with the landmark heart failure with preserved ejection fraction (HFpEF) phenomapping study by Shah et al,³ which used unsupervised learning to reveal previously unrecognised HF subgroups with distinct phenotypes and outcomes. Subsequent work, including the influential clustering analyses by Ahmad et al,⁴ further advanced data-driven phenotyping across both HFpEF and heart failure with reduced ejection fraction (HFrEF) populations, while current guidelines still rely on the coarse categorization of HF. In parallel, DL expanded the scope of HF research by enabling large-scale analysis of cardiac imaging, exemplified by Ouyang et al’s⁵ EchoNet-Dynamic model, which automated EF estimation from echocardiography videos with expert-level performance. More recent multimodal models that combine longitudinal chest X-rays with structured EHR data have improved prognostication in hospitalized patients with HF, outperforming single-modality approaches in predicting in-hospital mortality.⁶ Importantly, ML systems have begun to move into prospective clinical evaluation, with trials such as the AI-ECG randomized controled trial (RCT) by Yao et al⁷ demonstrating improved detection of asymptomatic left ventricular dysfunction in routine care pathways. Collectively, these developments mark a transition from exploratory modeling to clinically actionable AI tools capable of supporting more precise, scalable, and timely HF care.

In this review, we begin by outlining the key ML and DL approaches most commonly used in HF research, providing clear definitions to support readers in interpreting the methods discussed throughout the text. We then examine how the convergence of big data and ML is reshaping HF management, drawing on diverse sources such as EHR, patient registries, biobanks, imaging datasets, telemonitoring, and synthetic data. For each, we discuss their contributions to predictive modeling, risk stratification, and personalized care, as well as the specific challenges they present. We then consider the broader landscape of clinical adoption, highlighting promising applications of AI in real-world settings alongside the barriers that continue to limit their integration into routine care. Recognizing the importance of trust and transparency, we explore emerging ethical and technical guidelines that aim to ensure fairness, robustness, and usability of AI systems in health care. Special attention is given to the role of interdisciplinary collaboration and stakeholder engagement in aligning AI development with clinical needs and societal values. Finally, we reflect on new computational paradigms, such as privacy-preserving ML, and outline the steps needed to foster trustworthy, scalable, and impactful AI solutions for HF care.

Overview of ML Methods Used in HF Research: A Clinician’s Guide

Artificial intelligence (AI) refers to computational methods that can perform tasks normally requiring human intelligence. ML is a subset of AI in which algorithms learn statistical patterns and relationships directly from data, rather than relying on explicitly programmed rules. These models improve their performance through exposure to data and are particularly suited to tasks such as prediction, classification, and pattern recognition. DL is a further subset based on layered computational models, named neural networks, that automatically learn increasingly abstract features through successive nonlinear transformations, without manual feature definition.⁸

Applications in clinical AI rely on different learning paradigms and model types, each with unique strengths, limitations, and interpretability characteristics, which we will review in the following sections to provide clinicians with a practical understanding of ML tools in HF research.⁹

Learning Paradigms

Supervised Learning

Supervised learning refers to models trained using input data paired with known target labels, meaning that the output label for the training samples, for example, mortality¹⁰ or HF hospitalization,¹¹ is also provided to the model. The model learns statistical patterns that link input features to these ground-truth labels, enabling it to make predictions on new patients.

Unsupervised Learning

Unsupervised methods detect hidden structure in data without predefined labels. These algorithms reduce complex data into simplified patterns or cluster groups of patients based on observed similarities, providing data-driven insights that may complement guideline classifications. In HF, this has been particularly influential in phenotyping, for example, discovering HFpEF subgroups with distinct clinical and prognostic profiles.³

Self-Supervised Learning

Self-supervised learning leverages the natural structure of large unlabeled datasets, such as EHR medical reports, imaging frames, or ECG waveforms, to create internal pseudo-labels. These pseudo-labels are generated automatically from the data itself, for example, by predicting the next clinical event in an EHR timeline.¹² By learning these intermediate tasks, the model acquires clinically meaningful representations without requiring manual annotation. In clinical ML, self-supervised models often serve as pretrained foundations for downstream diagnostic or prognostic tasks.¹³

Common Model Families

Decision Trees, Random Forests, and Gradient-Boosted Trees

Decision trees learn sequential if–then rules that split patients into meaningful subgroups based on clinical features, offering high interpretability but often overfitting when used alone,¹⁴ meaning they may memorize noise or institution-specific artifacts rather than generalizable patterns. To address this limitation, ensemble learning, that is, methods that combine multiple models to produce stronger predictions, has been proposed. For example, these include random forests, which aggregate the predictions of many decision trees generated from resampled versions of the original dataset,¹⁴ and gradient-boosted trees (eg, XGBoost), which build trees sequentially so that each new tree helps correct errors made by previous ones.¹⁵ By combining the outputs of many trees or refining predictions step by step, these methods often achieve stronger performance in clinical tasks.¹⁶

Support Vector Machines

SVMs construct decision boundaries in high-dimensional space that separate patient groups (eg, those at higher versus lower risk) as clearly as possible.¹⁷ Because they rely on a subset of key training examples, SVMs tend to perform well when datasets are moderate in size and when features are carefully curated and standardized, which characterized much of the earlier ML research in HF prediction.¹⁸

Neural Networks and DL

Neural networks are ML models composed of layers of interconnected artificial neurons that apply weighted transformations and nonlinear functions to learn representations from data. When organized into multiple hidden layers, they form DL architectures capable of capturing complex, nonlinear relationships in high-dimensional medical data.⁸ DL automatically learns hierarchical features, and has enabled advances in HF diagnosis,¹⁹ phenotyping,²⁰ and prognostic modeling.²¹ However, DL requires large datasets due to many trainable parameters and is prone to overfitting. Its black-box nature makes interpretation challenging, highlighting the need for transparency, explainability tools, and rigorous external validation across diverse patient populations.²²

Transformers

Part of the DL landscape of models and originally developed for language tasks, transformer architectures model long-range dependencies using attention mechanisms, which allow the model to dynamically weigh which parts of the input (eg, prior clinical events, specific ECG segments, or imaging tokens) are most relevant for a given prediction. In medicine, they are increasingly applied to longitudinal EHR data, multimodal clinical inputs, and imaging sequences due to their ability to integrate diverse information streams.¹³

HF DataSets and ML

ML’s success in numerous domains owes much to the accumulation of vast datasets facilitating the training of large ML and DL models.²³ This trend is also emerging in medical domains, especially concerning CVDs, driven by the proliferation of various registries (eg, HF registries), EHRs, biobanks, and more.

Throughout this section, we will describe the present landscape of large HF data repositories, which are characterized by big volumes of diverse datasets—spanning millions of patient records, long-term follow-up, and multimodal data types (clinical, genomic, imaging, and continuous physiological signals), listing some of the tasks being leveraged by ML applications, and mentioning the challenges inherent in their deployment. Table 1 provides a structured comparison of the HF data sources, summarizing their clinical availability, strengths, limitations, commonly observed biases, regulatory considerations, and representative AI-in-HF applications. The table illustrates how each source contributes distinct advantages, such as scale, depth, or temporal resolution, while also presenting unique methodological and ethical challenges for ML. ^{4,10,11,16,18–20,24–50}

Table 1.

Comparison of Major Data Sources Used in AI for HF Research and Clinical Applications

To complement this, a comprehensive summary of referenced and critical studies employing ML to address clinical tasks is provided in Table 2, encapsulating both the opportunities they offer and the caveats they entail in terms of data and algorithm fairness (ensuring equitable outcomes across diverse patient populations), robustness (maintaining model performance across different clinical settings and data variations), and explainability (providing clear insights into how models make decisions to foster trust among clinicians and patients). ^{4,10,11,16,18–20,24–48,51–53}

Table 2.

Machine Learning Studies Leveraging Big Data for HF and Cardiovascular Disease Research

Electronic Health Record

EHRs, now widely adopted across health systems, provide the longitudinal, multimodal data foundation that enables both advanced risk prediction and well-adjusted observational studies in HF research.^12,24,25,28 While many statistical models generally have fared modestly in risk prediction of HF, contemporary modeling solutions like ML and DL models have found more success utilizing minimally processed, raw longitudinal records from administrative EHR datasets like the UK Clinical Practice Research Datalink.²⁵ Especially, in the age of large-scale sequential data models such as transformers, the greater the availability of sequential data—both in volume and data type—the greater the opportunity for robust modeling.^12,25–27 For instance, a transformer model trained on longitudinal UK EHR data achieved strong performance in predicting 6-month incident HF (area under the receiver operating curve AUROC=0.93) and incorporated explainability methods to identify both established and potentially novel risk factors such as particular HF medications.¹² Other notable real-world evidence demonstrated how DL analyses of large EHR cohorts can clarify cardiovascular risk patterns in multimorbid populations, showing, for example, a consistent monotonic association between systolic blood pressure and major cardiovascular events in patients with chronic obstructive pulmonary disease (COPD) and in those with diabetes, without evidence of a J-shaped relationship.²⁸ In addition, rich multimodal EHR datasets have allowed researchers to redefine HF phenotypes beyond traditional ejection fraction classifications, as demonstrated by Banerjee et al,²⁴ who proposed and validated 5 HF subtypes using multimodal longitudinal data combined with genetics. If implemented clinically, such models could enable earlier risk detection, improved stratification of multimorbid patients, and more personalized HF management. However, translation will require prospective and external validation, demonstration of clinical benefit in real-world decision-making, and robust, interpretable integration into clinical workflows.

Despite these advantages, analyzing EHR data with ML presents significant challenges. Data quality and heterogeneity remain primary concerns, as EHRs aggregate information from diverse sources, leading to inconsistencies in formats, terminologies, and completeness.^54,55 Furthermore, ML models often inherit biases present in EHR datasets, particularly if underrepresented populations are involved, potentially resulting in inequitable predictions.⁵⁶ This is the case in Obermeyer et al,⁵⁷ where the authors found that at a particular algorithm-based risk score, Black patients were considerably sicker than White patients. In the cardiology domain, a study by Li et al⁵⁸ found that ML models trained on data from 109 490 patients exhibited substantial gender bias. While models had modest fairness across racial groups, for gender the equal opportunity difference was 0.131 to 0.136 and the disparate impact was 1.535 to 1.587, indicating a markedly lower true-positive rate for women compared with men. Privacy and security concerns add another layer of complexity, as EHRs contain sensitive patient information requiring strict data governance and compliance with regulations. Lastly, the lack of standardized data formats and interoperability across EHR systems hampers the scalability and generalizability of ML models, limiting their broader clinical application.⁵⁹

Registries

HF registries, prospectively collected datasets designed around predefined research objectives, provide structured, high-quality information on well-defined patient cohorts, enabling detailed analyses of disease burden, treatment patterns, and real-world practice variation across health care systems. ML models trained on registry data have demonstrated significant potential for enhancing HF care. ML techniques have been used to uncover predictors of HF onset and progression, including risk factors³⁴ and molecular biomarkers,²⁹ supporting earlier intervention and improved stratification. In addition, these models aid in determining optimal windows for advanced therapies and forecasting comorbid conditions,³⁰ offering data-driven clinical decision support. Registry-based ML studies have also proven instrumental in identifying novel HF phenotypes via unsupervised learning methods,^31,32 improving risk prediction,^10,36–38 and evaluating the effectiveness of pharmacological treatments across patient subtypes.³³ Particularly, using the Swedish Heart Failure Registry, Ahmad et al⁴ applied random forests and cluster analysis to ≈45 000 patients, achieving strong prognostic performance and identifying 4 clinically meaningful HF phenotypes with distinct outcomes. Their work further demonstrated heterogeneous responses to common therapies across these subgroups, but translation into clinical practice remains limited, as these phenotypes require external and prospective validation, assessment of stability over time, and evaluation of whether phenotype-guided treatment decisions improve patient outcomes beyond standard care.

Registries, while offering notable strengths, are not immune to challenges. Many lack adequate adjustments for sex, ethnicity, or socioeconomic status, and ML studies often fail to address these biases. Some works, such as that by Segar et al,³⁵ demonstrated that ML models incorporating social determinants of health outperformed traditional risk scores in predicting in-hospital mortality for HF, particularly improving prognostic accuracy for Black patients, highlighting the value of integrating socioeconomic context into registry-based analyses. Another key limitation is that high-quality registries are available only in a few high-income countries, leaving many health care systems underrepresented.⁶⁰ This geographic skew can limit model generalizability and hinder the global applicability of AI tools. Explainability is also a pressing issue: while models using tree-based algorithms or other techniques provide some interpretability,⁶¹ DL approaches remain largely opaque. Finally, registries offer a vital means for external validation, boosting ML robustness across populations, but to fully harness their value, broader data sharing and transparency are essential.⁶⁰

Biobanks

Biobanks, large-scale, longitudinal repositories of biological specimens linked with detailed health, lifestyle, and environmental data, provide rich multimodal datasets that underpin genetic, epidemiological, and translational research, including studies in HF.⁶² The breadth and depth of data within biobanks provide a fertile ground for training and validating ML models that require large, diverse datasets to generalize effectively. Within the cardiovascular domain, biobank data have helped uncover novel HF risk factors,⁴¹ distinguish subtypes based on clinical heterogeneity,²⁴ and extract features related to cardiac function or structural abnormalities,⁴² contributing to more precise diagnostics and tailored therapeutic strategies. In particular, Nauffal et al²⁰ leveraged cardiac imaging and genetic data from over 41 000 UK Biobank participants to develop a DL model, based on convolutional neural networks, quantifying myocardial interstitial fibrosis, uncovering novel loci and pathways linked to fibrosis and HF, and demonstrating how biobank-based ML can illuminate disease mechanisms and potential therapeutic targets. However, further validation in clinical HF populations and evidence of incremental clinical utility are needed before routine application.

Several challenges remain in the use of biobanks for ML research. A major limitation is the lack of ethnic and geographic diversity, which constrains the fairness and generalizability of resulting ML models. Most national biobanks underrepresent minority populations, an issue that has received insufficient attention in ML-driven HF research. Efforts to connect datasets across borders face considerable privacy and regulatory hurdles.⁶¹ Federated learning (FL)—a decentralized ML approach where multiple institutions collaboratively train a model without sharing their data—emerges as a promising solution, enabling collaborative model training without data sharing, thus safeguarding patient privacy while enhancing data inclusivity.⁶³ Furthermore, for AI models to be clinically useful, they must be interpretable. Tools like decision trees, systematic feature selection methods, SHAP, and Local Interpretable Model-agnostic Explanations (LIME), both explainability techniques, have shown promise in elucidating ML predictions,⁶¹ fostering trust and adoption in clinical environments. Broader international collaboration, increased diversity in biobank cohorts, and continued development of explainable models are essential to fully realize the potential of biobank-based ML in HF care.

Telemonitoring Data

Wearable-based telemonitoring devices continuously capture physiological and behavioral data in patients with HF, enabling early detection of deterioration, remote management, and a dynamic understanding of patients’ health trajectories. ML models using wearable data have shown considerable potential in advancing HF care.⁶⁴ From statistical tools to DL architectures, these models enable real-time risk prediction, HF diagnosis, and therapy response assessment.^11,18,19,44 For example, seismocardiogram signals have helped distinguish decompensated from compensated states.⁴⁴ Prospective studies like the LINK-HF multicenter trial demonstrated that wearable multisensor chest patches can forecast hospitalizations with accuracy comparable to implantable devices.¹¹ More sophisticated AI applications have leveraged smartwatch ECGs and generative models to reconstruct full-lead ECG signals from limited channels.⁴⁵ In the self-management domain, Perramon-Llussà et al⁴⁹ implemented an AI-enhanced telemonitoring system for patients with HF, showing the potential of alert-based solutions to improve outcomes.

Important limitations persist in the use of wearable data for HF. Accessibility and integration remain limited, and device validation varies across contexts and populations, with biases linked to skin tone, physical activity levels, and socioeconomic status.⁶⁵ Moreover, the lack of standardized regulations and unified telemonitoring platforms creates a fragmented ecosystem dominated by proprietary technologies, with most providers restricting open access to data,⁶⁶ which is discussed thoroughly later in this review. This scarcity of labeled, high-quality data hinders the development and validation of robust ML models. Furthermore, current ML tools often fall short in trustworthiness, lacking fairness, usability, explainability, and traceability assessments. Addressing these challenges will require industry-wide data-sharing initiatives, standardization of wearable metrics, and the development of regulatory frameworks to support the ethical and effective deployment of telemonitoring-based AI in HF care.

Imaging Data

Imaging data, particularly from ECG and CMR, play a critical role in the diagnosis and monitoring of HF by providing precise measurements of cardiac morphology and function—such as ventricular volume, mass, stroke volume, and ejection fraction.⁴⁷ These metrics require detailed anatomic segmentation of the heart structures, a task that traditionally demands significant time and expertise. In response, DL solutions have been developed to automate this process. Notably, convolutional neural networks and architectures like nnU-Net, both DL models specialized for image analysis, have dominated recent cardiac segmentation challenges.⁶⁷ For instance, Zhang et al⁴⁶ demonstrated that DL-generated segmentations can match radiologist-level accuracy in just seconds and, at the same time, reduce the variance among different centers CMRs, making them a more robust tool.

Despite these successes in image segmentation, fully end-to-end DL solutions for HF diagnosis or risk stratification are still uncommon. Most workflows extract high-level imaging features—such as cardiac indices or radiomics—which are then fed into conventional ML models like SVMs or Random Forests.^43,68 The emerging field of CMR radiomics has gained particular attention due to its ability to capture complex imaging phenotypes, including texture and shape descriptors, beyond standard clinical indices.⁶⁸ When integrated with ML, radiomic features have improved the identification of HF subtypes and prediction of outcomes. For instance, Pujadas et al⁴³ showed that combining radiomics with conventional CMR indices and vascular risk factors significantly enhanced incident HF prediction using UK Biobank data. Similarly, radiomics has been applied to stratify atrial fibrillation risk in patients with ischemic heart disease, demonstrating superior performance over ECG markers alone.⁴²

Advanced models have also explored multimodal approaches and direct prediction pipelines. For instance, Pandey et al⁴⁷ applied a deep neural network to ECG-derived imaging features to phenotype diastolic dysfunction in HFpEF patients, identifying those most likely to benefit from spironolactone. However, these approaches remain reliant on manually segmented data, limiting scalability and increasing data requirements. A notable exception is the work by Shad et al,⁴⁸ who used a 3D convolutional neural network to directly analyze raw preoperative echocardiography videos and predict postoperative right ventricular failure, bypassing the need for segmentation. Their model achieved superior performance over clinical experts and integrated explainability methods via saliency maps, which visualized the anatomic regions most influential in the model’s decision-making. In addition, supporting the potential of ECG-derived imaging surrogates, the EchoNext model developed by Poterucha et al⁶⁹ used >1 million paired rhythm-imaging records to detect a wide spectrum of structural heart diseases, outperforming cardiologists and generalizing across diverse care environments. Notably, the model was prospectively validated in patients without prior cardiac imaging, uncovering previously undiagnosed disease. Such end-to-end, interpretable models illustrate the future potential of imaging-based DL applications in HF, though data scarcity and demographic biases remain a key bottleneck. In these regards, Kaur and colleagues found that the performance of ECG-based DL models for predicting HF varied significantly by race, sex, and age. For example, model discrimination (eg, C statistic/ROC) was lower in some under-represented subgroups, dropping from ≈0.80 in younger White patients to ≈0.69 in young Black patients and from ≈0.80 in patients ≤40 years to ≈0.66 in those >80 years, highlighting disparities in accuracy according to demographic factors.⁷⁰

Proprietary Data Silos

Recent analyses warn that the increasing concentration of health-relevant data within proprietary ecosystems creates significant challenges for privacy, governance, and equity.⁷¹ These companies collect vast multimodal datasets spanning physiological signals, behavioral and geolocation traces, and even EHR-linked clinical records, enabling powerful AI applications ranging from arrhythmia detection and HF decompensation forecasting to multimodal generative models.⁵⁰ However, because these data are stored in closed, nonstandardized, and commercially controlled environments, they remain largely inaccessible to public researchers, limit reproducibility, and risk shifting control of health-AI development toward a small number of private actors. This raises concerns about reidentification, unclear data ownership, unequal access to advanced analytics, and potential monopolization of the AI infrastructure underpinning future health care systems.⁷¹ To address these challenges, it is imperative to establish robust, transparent data-governance frameworks and shared data-trust infrastructures, ideally built through industry-academic partnerships, that support interoperable, privacy-protecting, and equitable access to health data, ensuring that the benefits of AI-driven innovation are widely distributed (OECD).⁷²

Trustworthy AI in Routine Clinical Practice

Although health care-related AI research has grown rapidly in recent years, including in HF, only a limited fraction of AI-driven solutions have successfully transitioned into clinical practice. A study by Muehlematter et al⁷³ reported that 222 AI-based medical devices were approved by the U.S. Food and Drug Administration (FDA) and 240 received CE marking in Europe between 2015 and 2020, highlighting the regulatory bottlenecks many innovations still face. While US regulatory activity has accelerated—nearly 1016 AI/ML-enabled devices had been authorized by the FDA by December 2024, including 221 approvals in 2023 alone, largely in medical imaging⁷⁴—Europe has seen more gradual progress. A significant shift occurred with the adoption of the EU Artificial Intelligence Act, which entered into force in August 2024 and classifies AI systems in medical devices as high-risk, thus requiring strict conformity assessments before market approval.⁷⁵ Complementing this, the European Society of Cardiology has launched the Digital Cardiology and Artificial Intelligence Committee (DCAI) to guide the responsible integration of AI into cardiology, emphasizing the need for evidence-based, equitable, and patient-centric applications (DCAI 2024–2026). Nonetheless, the practical deployment of AI tools in European clinical environments—especially for complex domains like cardiovascular care and HF—remains in its early stages, highlighting the ongoing gap between technical innovation and widespread clinical implementation. Table 3 exemplifies the range of AI maturity, from interventions with demonstrated clinical impact to predictive tools awaiting prospective validation.^7,69,76,77

Table 3.

Characteristics and Clinical Impact of Recent Pragmatic Trials and Validation Studies of AI-Enabled Cardiovascular Diagnostic and Prognostic Tools

Despite the recently deployed examples, there is a clear slow integration of AI into HF care compared with the rapid development of research, which cannot be attributed solely to clinician mistrust. More fundamentally, most AI systems remain immature for routine practice: they lack robust, multisite prospective evidence demonstrating patient-level benefit, are often trained on heterogeneous and biased datasets, and face substantial implementation, regulatory, and economic hurdles.⁷⁸ Furthermore, a 2021 report for the European Parliament (Panel for the Future of Science and Technology (STOA) 2022) identified 7 major risks of medical AI, including potential patient harm, misuse, inequities, lack of explainability, privacy breaches, accountability gaps, and difficulties with implementation in real-world health care environments.

To address these barriers, the review will next examine some challenges limiting the adoption of AI in HF management. It will explore how concerns such as algorithmic bias, insufficient transparency, and data security impact trust and clinical readiness. In addition, it will highlight real-world examples and innovations that illustrate the transformative potential of trustworthy AI, aiming to support a shift toward more ethical, explainable, and patient-centered HF care. These interrelated challenges—and the opportunities that arise at the intersection of data and algorithms—are summarized in Figure 1.

Figure 1.

Bridging heart failure (HF) data and machine learning (ML). This figure maps the intersection of HF datasets and ML techniques, highlighting the challenges and opportunities of integrating artificial intelligence (AI) into clinical care. HF data sources—including electronic health records (EHRs), registries, biobanks, imaging, and telemonitoring—provide rich but heterogeneous inputs for ML models. Key challenges involve bias, privacy, generalizability, and clinical integration. Nevertheless, combining ML techniques and approaches creates opportunities for improved risk stratification, early detection, personalised treatment, and stakeholder-driven, trustworthy AI in HF management.

Guidelines to Foster Acceptance and Innovation

To build trust in medical AI and support its safe and effective adoption, particularly in complex domains like HF, it is essential to go beyond technical performance and address broader ethical, societal, and usability concerns. While early guidelines focused on standardizing AI study reporting and regulatory evaluation, these efforts often lack domain-specific depth. A major step forward was the Assessment List for Trustworthy Artificial Intelligence (ALTAI), a general-purpose self-assessment tool based on 7 ethical principles.⁷⁹ However, ALTAI does not address the specific challenges of health care AI. The FUTURE-AI framework, developed through the collaboration of 117 interdisciplinary experts from 50 countries representing all continents, including AI scientists, clinical researchers, biomedical ethicists, and social scientists, addresses this gap by operationalizing trustworthy AI principles into 30 actionable recommendations across 6 pillar principles: Fairness, Universality, Traceability, Usability, Robustness, and Explainability.⁸⁰ An overview of these 6 principles and their practical implications in HF research is illustrated in Figure 2.

Figure 2.

The FUTURE-AI framework for trustworthy artificial intelligence (AI) in heart failure (HF) research. This figure illustrates the 6 core principles of the FUTURE-AI framework as applied to the development and deployment of AI systems in HF research and care. For each principle, a concise description, a real-world example of failure, and an actionable recommendation are provided. Together, these guidelines promote the creation of AI tools that are equitable, generalizable, transparent, clinically integrated, reliable, and interpretable, thereby fostering stakeholder trust and supporting responsible innovation in HF management. SHAP indicates Shapley Additive Explanations.

In parallel, adopting FAIR data principles across EHR platforms and other data types offers a practical pathway to reduce data heterogeneity in HF research and enable safer, more trustworthy AI development by promoting findability, standardized accessibility, interoperability through shared ontologies, and reusable de-identified formats.⁸¹ Complementary efforts include the development of standardized interoperability frameworks and common data models, such as Fast Healthcare Interoperability Resources⁸² and Observational Medical Outcomes Partnership,⁸³ which together aim to harmonize heterogeneous EHR data across institutions—Fast Healthcare Interoperability Resources by enabling standardized data exchange, and Observational Medical Outcomes Partnership by providing a unified analytic data representation. These technical standards are now being operationalized within the European Health Data Space, which provides a regulatory and technical framework for secure cross-border secondary use of health data.⁸⁴ In this context, the study by Laleci Erturkmen et al⁸⁵ demonstrated the potential of the HL7 Fast Healthcare Interoperability Resource International Patient Summary Implementation Guide to facilitate access to EHR data for secondary use. To this end, their analysis focused on 2 European Union-funded research projects that aim to reuse EHR data to develop AI models for personalized decision support services for patients with HF.

Increasing Trust in AI Tools Through Multistakeholder Engagement

Building on the challenges outlined in earlier sections, this part focuses on the crucial role of stakeholder engagement in mitigating risks, fostering trust, and ensuring the effectiveness of trustworthy AI tools in health care. While technical solutions are essential, actively involving patients, clinicians, and other key stakeholders from the outset enhances the design, usability, and success of these innovations.⁸⁶ This approach underscores the importance of understanding diverse contexts and involving end-users throughout the development lifecycle to create intuitive and user-friendly AI tools that align with their preferences (Figure 3). For example, in the Trustworthy Artificial Intelligence for Personalised Risk Assessment in Chronic Heart Failure (AI4HF) project, Puig-Bosch et al⁸⁷ and Cabrita et al⁸⁸ used cocreation workshops, interviews, and participatory methods to identify human-centered ethical and clinical requirements for AI in HF across different regions.

Figure 3.

Multistakeholder engagement cycle for trustworthy artificial intelligence (AI) in health care. The central cycle illustrates the iterative process of multistakeholder engagement throughout the AI lifecycle, encompassing 6 key steps: (1) defining objectives; (2) mapping relevant stakeholders; (3) selecting engagement methods; (4) specifying AI requirements; (5) implementing and deploying the model; and (6) gathering feedback for iteration. For example, of stakeholder types and common engagement approaches are provided at each corresponding step. The guiding principles at the bottom serve as the foundation for effective engagement throughout the entire lifecycle. Together, this framework demonstrates how proactive, structured stakeholder involvement can support ethical, trustworthy, and human-centered AI deployment in clinical practice.

Although health care-related AI research has grown rapidly in recent years, including in HF, only a limited fraction of AI-driven solutions have successfully transitioned into clinical practice. This gap contributes to a persistent disconnect between developers and end-users, resulting in AI tools that often fail to meet clinical needs. Addressing this requires structured, interdisciplinary collaboration across technology, health care, ethics, and the social sciences to ensure solutions remain aligned with real-world contexts. An illustrative example is the governance structure implemented by the University of Wisconsin Health, where a multidisciplinary steering committee systematically reviews and endorses predictive AI applications through a formal intake process, subcommittee review, and ongoing lifecycle monitoring to ensure interpretability, accuracy, fairness, and alignment with clinical workflows and ethical standards.⁸⁹ Such models highlight how diverse perspectives—from patients and clinicians to ethicists and regulators—can shape AI systems that are technically sound, ethically robust, and trusted by end-users. Yet, meaningful engagement remains challenging due to differing priorities, potential biases, and unequal digital literacy. To overcome these barriers, clear communication, transparency, education, and sustained mechanisms for dialogue are essential⁸⁶ (Figure 3).

Addressing Data Privacy and Fairness

Federated Learning

Advancements in ML for cardiology and HF research raise significant privacy concerns when handling sensitive health data. Centralized data repositories, although conventional, pose risks of exposing raw data during analysis. However, recent progress in FL provides a solution by enabling the construction of complex ML models without centralized data access. FL enables the use of private health data from devices or health care systems without centralizing the data, thereby advancing research while preserving privacy⁹⁰ (Figure 4). For example, recent work on FL extensions of medical image segmentation frameworks such as FednnU-Net illustrates how decentralized training can be applied effectively to clinical imaging tasks while protecting patient information.⁹¹

Figure 4.

Federated learning architecture across multiple hospitals. (1) The server sends the current global model to the local centers (hospitals). (2) Each center trains the model separately using its own private local patient data. (3) The local centers send only the model updates (eg, weights, gradients, or compressed statistics) back to the server; no raw data is transferred. (4) The server combines the updates from all centers using an aggregation rule (eg, by averaging the parameters) to generate a new, improved global model. Steps 1 to 4 are repeated until the model reaches a desired level of accuracy (convergence) or a stopping criterion is met. Finally, the resulting global model is shared with the hospitals and used for the target clinical task, enabling collaborative learning across institutions without exchanging sensitive patient data.

In HF research, FL promises the integration of diverse datasets, including data from different modalities, and complex models.⁶³ This translates to enhanced models, improved accuracy, and generalizability. This approach accommodates prediction and inference tasks, enabling precise analysis across health conditions. The successful application of FL to HF research showcases its potential to redefine data-sharing practices in the medical domain, and is being tested by European-funded projects DataTools4Heart and AI4HF.⁸⁵

On the other hand, FL also exhibits promising potential to enhance fairness across various demographic groups. Conventional fair learning approaches assume access to centralized training data, a condition often unmet in real-world applications due to privacy, legal, and logistical constraints. FL offers a novel strategy to navigate this challenge, enabling the training of fair classifiers on decentralized data while preserving privacy.⁹⁰ For instance, recent work on fairness-aware FL methods, such as fairness-aware aggregation algorithms⁹² and group fairness frameworks,⁹³ demonstrates that FL can mitigate disparities in model performance across sensitive groups (eg, racial and gender categories) even under heterogeneous client distributions.

Synthetic Data Generation

Synthetic data generation has emerged as a promising solution for supporting HF research and ML development while safeguarding patient privacy. This approach involves training generative models on real datasets to produce synthetic records that maintain the statistical characteristics of the original data but do not correspond to actual individuals.⁹⁴ As such, synthetic datasets can be safely shared and used for algorithm development, benchmarking, or collaborative projects without violating data protection laws. This facilitates faster data management processes and lowers barriers to data access, particularly in multiinstitutional research settings.⁵³

However, privacy risks persist if the generative model memorizes and reproduces data patterns too closely, potentially re-identifying individuals from the original dataset. To mitigate this, privacy-preserving techniques such as differential privacy are essential. Furthermore, since real-world datasets often reflect underlying biases, synthetic data generation can be used to create more equitable training sets by augmenting underrepresented subpopulations without compromising model performance.⁹⁵

Various generative models are now applied in the health care domain. For instance, several studies have demonstrated synthetic data generation using autoGAN for imaging data,⁸⁰ Tabular GAN for structured patient records,⁹⁶ and both data structures using a dual approach.⁵¹ Real-world applications include Guo et al’s⁵² use of MDClone to generate synthetic EHR data for over 26 000 patients with HF, enabling ML model training for mortality prediction without handling protected health information. Similarly, Tucker et al⁵³ applied Bayesian networks to generate synthetic datasets that addressed missing data and bias issues, releasing a synthetic cohort based on nearly 500 000 primary care patients for external validation and model prototyping. These developments underscore the role of synthetic data in fostering secure, scalable, and inclusive ML research for HF.

Conclusions and Future Perspectives

This scoping review underscores the critical role that large datasets play in advancing trustworthy ML applications for improved HF management. The integration of diverse data sources—from clinical records to biobanks and remote monitoring—has established a strong foundation for developing ML models capable of predicting patient outcomes, stratifying risks, and identifying novel HF phenotypes. While the potential for these models to enhance precision and personalization in HF care is significant, their trustworthiness remains paramount.⁸⁰

Importantly, the current level of evidence supporting the application of ML in HF management is still evolving. While initial findings from observational studies and real-world data analyses are promising (Table 3),^7,69,76,77 the absence of robust evidence from large RCTs, meta-analyses, or longitudinal studies limits the clinical readiness and translation of these models into clinical practice. This gap underscores the need for systematic evaluations of ML models similar to the rigorous frameworks used for introducing new pharmaceuticals, such as RCTs or implementation trials (eg, nudging trials), that assess real-world effectiveness and clinical impact.⁷⁵ Establishing such standards is essential to move beyond theoretical potential toward practical, validated applications that genuinely improve patient care.

The review also identifies significant challenges at multiple levels. At the technical level, issues such as data heterogeneity, algorithmic bias, and the need for explainable AI models must be addressed to enhance model reliability and user trust. At the human level, barriers such as clinicians’ understanding of AI, potential resistance to adoption, and the alignment with clinical workflows need to be tackled. In addition, research methodologies must evolve to integrate ML validation into clinical frameworks, ensuring robust study designs that test both accuracy and practical utility. Proposed solutions include the implementation of privacy-preserving techniques, such as FL frameworks and synthetic data generation,⁹⁷ and novel evaluation protocols tailored to AI systems, ensuring compliance with ethical and clinical standards.

Emerging AI paradigms may transform HF care, evolving from isolated workflow support to highly integrated, predictive precision medicine over the coming decades. In the near term, multimodal foundation models, pretrained on vast collections of imaging, physiological signals, and clinical text, are expected to streamline workflows by automating tasks such as radiology report generation, diagnosis, and treatment planning.⁹⁸ As the field progresses toward a medium-term horizon, the maturation of large-scale wearable foundation models could enable a shift from episodic care to continuous, noninvasive home monitoring for early detection of HF decompensation.⁵⁰ Ultimately, in the long-term, the convergence of these data streams may support the realization of digital twins—individualized simulations of cardiac physiology⁹⁹—and autonomous AI agents capable of proposing context-aware clinical actions.¹⁰⁰ Collectively, these approaches point toward a future in which HF management shifts from static risk prediction toward continuously updated, predictive, and individualized precision medicine.

In conclusion, the potential benefits of ML in HF research and management are vast. However, the path forward lies in harmonizing technical innovation with evidence-based validation, ensuring that these models not only deliver accuracy but also demonstrably improve clinical outcomes. By addressing challenges at all levels and fostering trust, ML can fulfill its potential to significantly enhance HF management and improve patient outcomes.

Article Information

Disclosures

S. Rao reports consultancy fees from Lucem Health and grants from Oxford University Hospital Trust. The other authors report no conflicts.

Funding Statement

This work received funding from the European Union’s Horizon Research and Innovation program under Grant Agreement No. 101080430 (AI4HF project). Dr Gkontra was supported by the project TrustAI-ES (PID2023-146751OA-I00), funded by MICIU/AEI/10.13039/501100011033. With the support of the Industrial PhD Plan of the Department of Research and Universities of the Government of Catalonia. This work is also supported by the European Union’s Horizon Europe research and innovation program under Grant Agreement No. 101057849 (DataTools4Heart project).