Opportunities and Challenges of Cardiovascular Disease Risk Prediction for Primary Prevention Using Machine Learning and Electronic Health Records: A Systematic Review

Abstract

Background:

Cardiovascular disease (CVD) remains the foremost cause of morbidity and mortality worldwide. Recent advancements in machine learning (ML) have demonstrated substantial potential in augmenting risk stratification for primary prevention, surpassing conventional statistical models in predictive performance. Thus, integrating ML with Electronic Health Records (EHRs) enables refined risk estimation by leveraging the granularity and breadth of longitudinal individual patient data. However, fundamental barriers persist, including limited generalizability, challenges in interpretability, and the absence of rigorous external validation, all of which impede widespread clinical deployment.

Methods:

This review adheres to the methodological rigor of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) and Scale for the Assessment of Narrative Review Articles (SANRA) guidelines. A systematic literature search was performed in March 2024, encompassing the Medline and Embase databases, to identify studies published since 2010. Supplementary references were retrieved from the Institute for Scientific Information (ISI) Web of Science, and manual searches were curated. The selection process, conducted via Rayyan, focused on systematic and narrative reviews evaluating ML-driven models for long-term CVD risk prediction within primary prevention contexts utilizing EHR data. Studies investigating short-term prognostication, highly specific comorbid cohorts, or conventional models devoid of ML components were excluded.

Results:

Following an exhaustive screening of 1757 records, 22 studies met the inclusion criteria. Of these, 10 were systematic reviews (four incorporating meta-analyses), while 12 constituted narrative reviews, with the majority published post-2020. The synthesis underscores the superiority of ML in modeling intricate EHR-derived risk factors, facilitating precision-driven cardiovascular risk assessment. Nonetheless, salient challenges endure heterogeneity in CVD outcome definitions, undermine comparability, data incompleteness and inconsistency compromise model robustness, and a dearth of external validation constrains clinical translatability. Moreover, ethical and regulatory considerations, including algorithmic opacity, equity in predictive performance, and the absence of standardized evaluation frameworks, pose formidable obstacles to seamless integration into clinical workflows.

Conclusions:

Despite the transformative potential of ML-based CVD risk prediction, it remains encumbered by methodological, technical, and regulatory impediments that hinder its full-scale adoption into real-world healthcare settings. This review underscores the imperative circumstances for standardized validation protocols, stringent regulatory oversight, and interdisciplinary collaboration to bridge the translational divide. Our findings established an integrative framework for developing, validating, and applying ML-based CVD risk prediction algorithms, addressing both clinical and technical dimensions. To further advance this field, we propose a standardized, transparent, and regulated EHR platform that facilitates fair model evaluation, reproducibility, and clinical translation by providing a high-quality, representative dataset with structured governance and benchmarking mechanisms. Meanwhile, future endeavors must prioritize enhancing model transparency, mitigating biases, and ensuring adaptability to heterogeneous clinical populations, fostering equitable and evidence-based implementation of ML-driven predictive analytics in cardiovascular medicine.

Graphical abstract

Keywords

cardiovascular disease / machine learning / electronic health records / risk prediction / primary prevention

1. Introduction

Cardiovascular disease (CVD) remains the most significant threat to global population health and has seen an emergent increase in negative impact [1]. One strategy might involve early prediction of CVD risk and prevention before the symptoms of CVD manifest, through prescribing statins and lifestyle intervention [2, 3, 4].

In most Western countries, clinical guidelines have discussed and suggested that primary care utilize CVD risk prediction scores, usually focusing on the individual’s 10-year future risk of CVD based on basic indices such as blood pressure and their previous medical history. For instance, QRISK3 [5] in the UK by NICE (The National Institute for Health and Care Excellence) guidelines [2], the Pooled Cohort Equations (PCE) in the US by ACC/AHA (American College of Cardiology/American Heart Association) guidelines [3], and SCORE [6] in Europe by ESC (European Society of Cardiology) guidelines [4]. These risk scores are all based on conventional statistical models, such as Cox proportional hazards models. These scores have been independently externally validated by various research [7, 8] and have been amended by adding new predictors by developers [9]. Additionally, these scores have been used in clinical settings for years.

However, several studies have shown that the performance of these scores is not satisfactory, including the underestimation or overestimation of risk for certain population groups [10, 11]. Recent findings suggest that machine learning (ML) might be a good method to replace conventional statistical algorithms due to its ability to handle more complex data types [12]. And electronic health records (EHRs) might be a great source for this new technique to achieve this task [13].

1.1 Rationale

The motivation and rationale for this review are grounded in the imperative to enhance the performance of CVD risk prediction. In this context, “performance” encompasses not only the statistical metrics of discrimination and calibration but also the practical applicability and effectiveness of algorithms in real-world clinical settings. The increasing prevalence of CVD underscores the urgency to explore innovative approaches to predicting and managing CVD. The adoption of EHRs and ML technologies has shown promise in refining CVD risk prediction. Despite their potential advantages, the use of these technologies is not without challenges and limitations.

Thus, this review seeks to offer a comprehensive examination of both the potential benefits and the limitations associated with employing EHRs and ML for CVD risk prediction, while also identifying opportunities and challenges for future research endeavours. By enlightening healthcare professionals and researchers about the capabilities of these technologies, we aim to enhance their utilization and, ultimately, improve patient outcomes through improved risk prediction and management strategies.

1.2 Objectives

The objectives of the review are to:

•	1. Examine the current evidence on CVD prediction models and assess the potential of EHRs and ML models for enhancing CVD risk prediction.
•	2. Identify the limitations of using EHRs and ML for CVD risk prediction, covering both clinical and technical aspects.
•	3. Identify elements of an integrative framework for development, validation, and application of ML based CVD risk prediction algorithm.
•	4. Highlight areas for future research directions to optimize the use of EHRs and ML for CVD risk prediction.

2. Methods

This review is conducted based upon the elements which described in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [14] and Scale for the Assessment of Narrative Review Articles (SANRA) [15].

2.1 Study Design and Search Strategy

We conducted our research on publications in March 2024. Firstly, we searched electronic databases, including Medline and Embase via Ovid, for the period from January 2010 to the present. We utilized a combination of Medical Subject Headings (MeSH), and free text related to ‘CVD’, ‘ML’, ‘EHR’, and ‘risk assessment/factors’ to identify relevant studies published since 2010. The search was limited to the selected years, as more contemporary studies are likely to utilize ML algorithms and EHR datasets.

We further refined our research by focusing on publications that involved human subjects, were written in English, and had full-text availability. We extracted the necessary abstract information from these publications. Details of the search log and strategies are available in the Appendix A. Additionally, we conducted a comprehensive search of the reference lists from the selected papers using Institute for Scientific Information (ISI) Web of Science. Based on our knowledge of the research topic, we manually identified and included potential publications of interest.

2.2 Criteria for Study Selection

We extracted all the search results and materials collected from multiple sources and removed any duplicate papers using Rayyan [16], an online application for the initial screening of systematic reviews. Firstly, we screened the titles and abstracts, eliminating any irrelevant publications. Subsequently, we reviewed the full text to extract potentially inclusive studies, enabling us to select the best source materials for inclusion.

The primary interest focuses on ML-based prediction models utilizing dataset derived from EHR, specifically targeting algorithms for the primary prevention of CVD by predicting the long-term incidence of major cardiovascular events. Thus, eligible citations should include qualitative reviews discussing this area or systematic reviews, with or without quantitative meta-analysis, reporting on such models. Studies that solely develop or validate individual models, as well as methodology papers, will not be considered. The target population should be adults with no prior CVD history or any cardiovascular (CV) symptoms (e.g., acute coronary syndrome (ACS)) or those on statin prescriptions. Models developed exclusively for patients with specific comorbidities (e.g., diabetes mellitus (DM), chronic kidney disease (CKD)) or for certain sub-population groups (e.g., minor ethnic groups, gender-specific) will not be included. The models must be suitable for use in predicting long-term outcomes (e.g., 5, 10, 15 years risk, lifetime risk) in outpatient/general practice (GP) settings for the purpose of early prevention, rather than for patients in inpatient hospitalization or emergency departments (ED), where the aim is to predict short-term adverse health outcomes. The reviews or systematic reviews discussed must involve ML/deep learning (DL), with at least a portion of the models discussed being ML-based. Studies reporting only conventional statistical methods, such as survival analysis and cox model, will not be included. Likewise, studies focusing solely on artificial intelligence (AI) in the context of embedding, natural language processing (NLP) subtype definition clustering, will be excluded. The CVD outcomes of interest should include either composite or individual cases of coronary heart disease (CHD), stroke/transient ischemic attack (TIA) and heart failure, excluding management missions such as predicting length of stay, admission or readmission, and mortality or survival after medical operations in hospitalization and ED settings. The required data type should be based on EHR. While this is a novel technique and most data sources will be structured patient-level health data, the review or systematic review must at least mention and discuss EHR data. Studies that do not mention EHR will be excluded. Data sources focusing only on images, sound, and genetic data will not be considered unless combined with EHR data. Note, these inclusion criteria are not overly strict; studies that discuss or mention ML, CVD risk, and EHR in some capacity will be considered. Detailed inclusion and exclusion criteria are presented in Table 1.

2.3 Data Extraction

We utilized the PRISMA guidelines [14] and Microsoft Excel to extract data from the included publications. From each paper, we recorded pertinent information such as the first author’s name, year of publication, and the journal in which the study was published. We also noted whether each publication contained key details related to the domains we were particularly interested in. Additionally, we considered the rationale and objectives outlined in the introductions of the papers. Furthermore, we extracted the main findings of each study and provided a summary.

3. Results

3.1 Study Selection

The study selection process is depicted in the Fig. 1. A total of 1757 studies were identified from all considered sources: 358 from PubMed/Medline, 1311 from Embase, and an additional 88 citations through backward searching in the Web of Science. After the removal of 258 duplicate studies, 1499 underwent an in-depth evaluation based on title and abstract screening. During this phase, 1314 publications were excluded. Subsequently, 185 records were sought for retrieval; however, 23 additional studies were excluded because they were either preprints without peer review or were conference abstracts and posters. As a result, 162 studies were deemed eligible for further full-text assessment of their eligibility. Ultimately, after applying all exclusion criteria, 140 publications were excluded. The Table 2 (Ref. [11, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37]) provides a summary of the characteristics of the remaining 22 publications [11, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37].

3.2 Descriptive Results

Out of the 22 selected publications, 10 (45%) are systematic reviews. Among these, 4 also include quantitative synthesis via meta-analysis and 5 of them register with PROSPERRO [38]. The remaining 12 (55%) are categorized as review articles. The majority of these publications are recent, mostly post-2020.

Among the 10 systematic reviews (SRs), most searched for studies within databases such as Medline/PubMed, Embase, and Web of Science (WoS). Only one SR did not adhere to reporting guidelines [22], while the others followed PRISMA [14] or TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) [39]. For the risk of bias assessment, three used QUADAS (Quality Assessment of Diagnostic Accuracy Studies) [40], two employed PROBAST (Prediction Model Risk of Bias Assessment Tool) [41], and two utilized self-designed methods derived from well-recognized techniques [20, 24]. Regarding data synthesis, six conducted only descriptive analysis. Among the four methods that performed meta-analysis are: random-effects network meta-analysis [28], Bayesian meta-analysis [30], hierarchical summary receiver operating characteristic curve (ROC) [24], and linear mixed-effects meta-analysis [32]. The performance metrics extracted from their selected models were mostly area under the receiver operating characteristic curve (AUC) or accuracy.

Regarding the coverage and focus on CVD of these articles, only two of the selected publications discuss the exact task of interest [11, 32]. 16 investigate risk prediction associated with a broad definition of CVD outcomes and tasks. 4 concentrate on risk prediction for single CVD subtypes as part of primary prevention: 3 on heart failure (HF) [29, 30, 37] and 1 on stroke [25].

All the publications selected discuss the use of EHRs to some extent. 2 of the publications dedicate a section to thoroughly discussing the use of EHRs in CVD risk prediction [23, 33]. 15 publications discuss the integration of EHRs in ML-based CVD prediction within the text, usually highlighting the opportunity of integrating EHR with ML. 5 others mention EHRs and use EHR-based models as a reference but do not fully explore this topic.

Since the selected publications were primarily systematic reviews and reviews, the ML/DL methods considered typically encompassed a range of models. The specific models discussed were determined by the literature selected for review. 6 of these publications include discussions comparing conventional models with ML methods [17, 18, 29, 30, 32, 33].

3.3 Opportunities and Challenges Identified by the Selected Studies

The existing literature has been limited in its discussion of ML-based CVD risk prediction leveraging EHRs in the outpatient setting, particularly regarding the 10–15-year risk of CVD in undiagnosed patients—an area that holds potential for primary prevention strategies. This review addresses this gap by focusing exclusively on this specific domain and exploring the potential to replace existing conventional risk assessment tools. The opportunities and challenges, summarized from selected publications in the Table 3, relate to the nature of CVD, the characteristics of EHRs, and the complexities inherent in ML/DL models.

3.3.1 Cardiovascular Disease Outcomes

Given the complexity of CVD, a careful evaluation of previous publications and their specific CVD focus areas is crucial. Early intervention plays a critical role in the effective prevention and management of CVD, as patients are often asymptomatic during the early stages of the disease. In fact, early intervention and treatment have been shown to prevent up to 80% of heart disease and stroke events, underscoring the importance of identifying and managing risk factors as early as possible to prevent the development and progression of CVD [42]. All two studies focusing on HF emphasize that efforts should be shifted from stages C/D to stages A/B [29, 37].

Personalized cardiovascular medicine, enabled by ML and EHRs, represents a promising approach to tailoring therapy and treatment strategies. By identifying individualized risk factors and profiles, this approach ultimately leads to better patient outcomes [26, 34]. Moreover, it facilitates the prioritization of screening and intervention efforts at the population level, as well as the efficient allocation of public resources, making it a cost-effective option in clinical practice [27].

By leveraging large EHR datasets, ML algorithms can identify novel CVD risk factors and elucidate relationships between known risk factors and CVD outcomes. This leads to a more comprehensive understanding of underlying mechanisms and enables more targeted investigations. Furthermore, several reviews have discussed the possibility of predicting cardiometabolic diseases (including CKD and diabetes) collectively [43, 44], and some reviews have already reported studies on different cardiometabolic diseases, including CVD [45].

However, CVD is a complex and multifactorial disease characterized by many interrelated risk factors, which can be challenging to quantify. The considerable heterogeneity among individual patients with CVD further complicates the integration of data from multiple sources, especially when it involves sound, image, and genetic data. Some reviews have raised concerns that such complex models may become too overloaded [17, 20, 35].

Nonetheless, the application of ML algorithms in CVD research holds substantial potential for addressing these complexities and improving patient outcomes. It is crucial, however, to ensure that predictive models are adaptable to the various manifestations of CVD and that they incorporate new risk factors as they are identified. This approach will enable effective risk stratification and intervention. Continuous learning and improvement have been identified as challenges by some reviews, emphasizing the need for risk prediction models to be updated continually with developments in CVD research [33].

3.3.2 Electronic Health Records Data Sources

The use of comprehensive and diverse individualized datasets is critical to developing accurate and effective ML based CVD risk prediction model. Most of the selected publications agree that EHR is recognized as a potential data source, which will be increasingly utilized for CVD prediction tasks in the future, potentially replacing traditional data collection methods [20, 23, 24, 25, 26, 27, 28, 31, 33, 35, 37]. EHRs should encompass diverse data sources, including background characteristics, medical history, laboratory results, prescribed therapies, and diagnoses. Additionally, they should feature large in-sample sizes and encompass patient populations that are generalizable across countries and settings [18, 23, 29, 37]. By utilizing such datasets, researchers can provide more complete records and reduce the proportion of missingness related to CVD events, risk factors, and co-morbidities [46, 47].

Real-time longitudinal data is also important as data linkage techniques of EHR can generate longitudinal health data that can be used to timely identify patients at high risk of CVD events [20, 48]. Furthermore, the use of EHR phenotyping provides a standardized platform for extracting, disseminating, and reusing clinical information. This can help improve the accuracy and consistency of data collection [49], which is essential for developing reliable ML models.

One major challenge in utilizing EHR data for CVD research is the issue of data quality and completeness [24]. Data from different EHR sources can vary in quality and completeness, and the unstructured and heterogeneous nature of some EHR sources can make it difficult to integrate data across different platforms [50]. While some studies have explored ML models leveraging EHR for CVD risk prediction [48, 50, 51], systematic evaluations remain limited, and existing research often lacks external validation or standardized methodologies. Moreover, the imbalance in data for some CVD outcomes has also raised concerns among researchers during the training of ML-based CVD models [22, 24].

Interoperability and standardization are also important challenges that must be addressed [52]. Different data formats, codes, and controlled terminologies need to be mapped and standardized to ensure that the data can be effectively integrated and utilized in ML models.

In addition, data privacy and security are significant concerns in using EHR data for CVD research [17, 26, 34, 37]. This is particularly true when the developed models require access to data for validation, a process that is unlikely to occur due to the current protection laws governing most EHR datasets [46]. EHR data contain sensitive patient information, and obtaining consent can cause selection bias [53]. Moreover, even anonymous data can potentially be traced back to individuals, creating potential risks to patient privacy [54]. To overcome these challenges, researchers must develop and implement appropriate data security and privacy protocols, as well as obtain necessary ethical and regulatory approvals [37].

3.3.3 Machine/Deep Learning Technique

One major advantage highlighted in most of the selected publications is the potential for improved accuracy. These publications generally view ML-based models positively, noting their superior predictive performance in terms of both discrimination and calibration when compared to conventional approaches [11, 19, 20, 21, 25, 31, 32, 36, 37]. Specifically, ML algorithms are reported to offer enhanced predictive capabilities for CVD events in subpopulations, surpassing traditional statistical models. Moreover, numerous individual studies have demonstrated the superior performance of DL or neural network models over other ML models [11, 17, 19, 20, 22, 25, 28, 35]. Additionally, other ML techniques, such as SVM and ensemble methods—particularly boosting algorithms—have been frequently mentioned in systematic reviews as offering optimal performance [11, 19, 24, 25, 28].

ML demonstrates remarkable flexibility through its ability to manage non-linear relationships between risk factors, or covariates, and CVD outcomes. It effectively models complex and previously hidden interactions among various clinical and environmental variables, thereby accurately predicting desired CVD outcomes [19, 25, 29, 34, 37].

Furthermore, ML can be personalized for individuals by incorporating both individualized and community-level features [55], significantly enhancing the CVD prediction and therapy recommendation process [22].

ML models can efficiently extract data from large and complex datasets, providing timely responses at a relatively low computational cost compared to conventional approaches [22]. They are also capable of adapting to new data sources for continuous improvement and learning [33].

The use of ML in CVD research brings several advantages over conventional statistical models, but it also presents challenges that must be addressed. One notable challenge is the performance variability of ML models, which can be influenced by factors such as the demographic characteristics of the data source, the selection of features, the optimization of hyperparameters, and the choice of performance metrics [11, 21, 25]. This variability often leads to statistical heterogeneity in reported results, a problem that is extensively discussed in the literature.

Typically, studies assess the performance of different ML models by analysing each CVD subtype individually, subsequently choosing the best-performing model for that subtype. This method, while practical, introduces the risk of selection bias, as the chosen model might not be universally superior across all contexts. Furthermore, selected SRs have highlighted difficulties in identifying the optimal ML models from their evaluated development papers [20, 27]. These challenges arise from inconsistent performance comparisons, attributable to the fact that their included studies define different types of CVD [21, 22, 35]. Moreover, concerns arise regarding whether the performance differences between ML models and conventional statistical approaches are statistically significant [21].

Another challenge arises when an ML model is trained on a dataset without properly addressing the imbalance in CVD data. Its flexible nature may lead to overfitting the training dataset [20, 22], which, in turn, results in poor generalization to new data sources, especially for minority ethnic groups.

In addition, the interpretability of flexible ML models presents a concern, often referred to as the ‘black box’ paradox, which is undoubtedly a major challenge acknowledged by researchers. However, issues with the ‘black box’ extend beyond its inherent nature. A lack of interpretability can also stem from developers failing to disclose technical features. Selected studies have identified flaws in model development, including the absence of clear definitions and measurements for predictors [35], a lack of data-driven feature selection [19, 20], and insufficient details on hyperparameter tuning [24, 35]. This omission of technical details, coupled with a lack of replication instructions by the developers33, breaches the principle of code transparency [17, 25, 27, 35, 36, 37]. Consequently, it makes reproducing the models challenging and hinders their interpretability [17, 23, 31, 35]. Besides these technical reasons, there are also challenge before clinical application. Studies have report that several challenge, firstly, the different definition of CVD outcome may hard to be interpret in clinical settings [18, 35]. Also, selected studies also mention that the lack reporting of clinical features for model developer [22, 24, 35, 36].

Most ML-based CVD prediction models report their performance using the AUC, also known as the c-statistic, to demonstrate the models’ ability to discriminate [17, 20, 24, 25, 30, 31, 32, 35, 37]. But the clinical value of AUC value is hard to interpret. One review criticized that the healthcare impact of those ML models is overestimated and unrelated to patients’ clinical benefits [21]. While AUC quantifies a model’s discriminatory capacity, it provides an incomplete assessment of clinical utility [21, 24]. A robust predictive model should demonstrate strong calibration, reliability, and balanced trade-offs across multiple performance metrics [30, 32, 35]. Calibration, assessed through the Brier score, calibration curve, and calibration slope, reflects how well predicted probabilities correspond to actual outcomes. However, calibration metrics are often underreported, limiting comprehensive model evaluation [25, 36, 45]. Similarly, sensitivity, specificity, precision, recall, and F1-score capture different aspects of performance, yet selective reporting skews comparative assessments [25, 31, 45]. Given these limitations, integrating multiple metrics provides a more rigorous and clinically relevant assessment of ML-based CVD risk prediction models.

Ethical considerations are less frequently discussed in the selected publications, though they are crucial before the final application of ML in CVD prediction in real healthcare settings.

The first consideration is fairness. Health equality might be compromised by incorporating ML-based risk scores. It is noteworthy that most of the selected publications report that the ML models are predominantly developed by researchers in Western countries, focusing on populations of white ethnicity [17, 18, 21, 23, 25, 33, 34, 35, 37]. This situation could result in ‘indicate bias’, particularly disadvantaging patients in rural areas or those who are ethnic minorities. Additionally, these techniques could be inaccessible to patients with limited digital literacy [27].

Another consideration is accountability. Studies have observed a lack of consensus regarding the clinical effectiveness and safety of ML in practice. To date, none of the ML-based studies have conducted any clinical utility tests, and with the lack of clinical impact assessment for these risk prediction models, their usefulness in healthcare settings remains unknown.

Furthermore, the transparency of the ML algorithms themselves is uncertain; few clinicians are involved in the development of ML models, nor do they provide feedback to the developers. This issue has been reported by several selected studies [11, 26, 27, 34, 37].

These ethical issues highlight a significant gap in specific guidelines (e.g., NICE, AHA/ACC) for the development and implementation of ML in healthcare. Moreover, national regulatory oversight (e.g., Medicines and Healthcare Products Regulatory Agency (MHRA), Food and Drug Administration (FDA)) is essential to ensure standardized development, along with the clinical effectiveness and safety of ML. This requirement has been recognized by multiple selected studies [17, 18, 23, 24, 27, 35, 36, 37].

4. Discussion

A significant number of conventional CVD models utilizing EHR, or registry data have been developed, validated, and implemented in practice, with several high quality systematic reviews reporting on such models [56, 57, 58]. Numerous studies and researchers contend that expending extensive effort to develop new risk models is unnecessary at this stage [56, 59]. In contrast, systematic reviews focusing on ML-based models are scarce [32]. Furthermore, to our knowledge, no systematic or narrative review has specifically targeted ML-based CVD prediction models using EHR data for primary prevention.

The debate continues regarding whether ML-based approaches offer superior performance. At present, conventional models still play a significant role in clinical practice due to their simplicity and interpretability. The recently published updates of QRISK4 [9] (derived from UK EHR data) and the Predicting Risk of Cardiovascular Disease Events (PREVENT) equations [60] have demonstrated improved performance over previous conventional models, with the potential to enhance their clinical utility. This underscores the adequacy of conventional models for CVD primary prevention at present. However, given ML’s capacity for continuous refinement and long-term potential, a systematic evaluation of its role in risk prediction is warranted. Rather than viewing ML as an immediate replacement, delineating how it can complement and ultimately augment existing approaches is pivotal for advancing CVD risk prediction.

While some studies suggest that ML/DL outperforms conventional models, the need for appropriate independent external validation of any improvements is still not fully addressed [19, 20, 23, 29, 30, 35, 36, 37, 55]. This issue also applies to comparisons between different ML models. Thus, the development of ML-based models should be approached with focus on clinical utility, ensuring a responsible translation from research to real-world applications. The objective is not to supplant conventional models but to facilitate their gradual integration where ML demonstrably adds value. As ML advances, proactive measures—such as standardization, validation frameworks, and enhanced interpretability—are essential for its future adoption. An effective ML-based model should be both transparent and standardized, particularly when leveraging complex EHR data. Improved explainability will enable replication and iterative refinement, ensuring continuous advancement throughout its lifecycle.

4.1 Prospects of ML Based CVD Risk Prediction Model

The ‘last mile’ problem refers to the situation where the final step of operationalizing a concept into the real world proves to be the most complex and costly.

Therefore, developing an integrative framework for the development, validation, and application of ML-based CVD risk prediction algorithms, as illustrated in Fig. 2, is crucial. Such framework should incorporate both clinical and technical perspectives, guiding researchers in creating practical and effective models that truly benefit clinicians and patients. While not exclusive to EHR-based ML models, use of EHR enhances the implementability of such frameworks.

4.1.1 Clinical Relevance

Developing ML models for CVD requires a focus on clinical relevance and interpretability to ensure they meet the needs of patients and can be effectively used by clinicians for timely interventions. The integration of such models into clinical practice must address concerns around clinical workflow, model validity, and overall value to patient care [22, 24, 35, 36]. Utilizing EHR data for model development calls for transparent reporting of data use, including the phenotyping, mapping, and linkage of data items.

Predictor selection should consider demographics, laboratory data, medical history, and other relevant factors, with detailed reporting on measurement units and criteria. Ensuring transparency at this stage is key to the reproducibility and clinical utility of the models [17, 25, 27, 37].

Validation of these models should yield interpretable and justifiable results, in line with established clinical guidelines from organizations like the ACC/AHA, NICE, and ESC. Although specific reporting guidelines for healthcare ML models are in development, initiatives like TRIPOD-AI [61] are emerging to fill this gap.

Involving healthcare professionals and domain experts is critical for assessing a model’s clinical usability and ensuring its alignment with real-world practice [11, 26, 27, 34, 37]. Prior to clinical application, a model’s performance and impact must be evaluated against existing standards. Its integration into clinical workflows should be feasible, with consideration of healthcare providers’ and patients’ acceptance, as well as its cost-effectiveness analysis.

4.1.2 Technical Robustness

Key considerations for ensuring the technical robustness of CVD risk prediction ML models include selecting an algorithm fit for the task, which should be effective across diverse patient demographics to avoid bias and ensure generalizability [11, 21]. Accurate, reliable, and unbiased performance is also essential, yet details on these factors are often inadequately reported [17, 23, 24, 30].

In development, the choice of ML model should balance performance with computational practicality [34]. Despite a large sample size negating the impact of algorithm choice on predictive accuracy, optimization techniques are vital for preventing overfitting and ensuring generalization.

The complexity of the models should be weighed against their ease of use and understanding. Developers need to document technical details, such as optimization techniques used, to ensure transparency and reproducibility [31, 35, 36, 45].

When implementing models in clinical settings, thorough evaluation using proper metrics, including probabilistic calibration and classification-based confusion matrix metrics, is necessary [30, 37]. Transparency in the model’s decision-making process is also critical, alongside compliance with data privacy and security measures. The lack of specific regulations for ML-based models calls for the development of comprehensive frameworks to maintain ethical standards in healthcare applications.

4.1.3 Summary

Developing, validating and applying ML models for CVD requires that both clinical and technical aspects are considered. Ensuring robustness and reliability enhances the trustworthiness for clinical use. Rigorous development and validation processes establish performance and effectiveness, instilling greater confidence in application. Identification of limitations and continuous evaluation lead to improvements and advancements in CVD prediction and management. Thus, frameworks are needed to support this process and yield more reliable, accurate, and impactful models, driving innovation and improving patient primary prevention of CVD.

4.2 The Definitive EHR Platform for ML Based CVD Risk Prediction Model

To advance ML-based CVD risk prediction, we propose a regulated, open-source EHR platform designed for standardized model development, evaluation, and validation. This initiative addresses key challenges in ML adoption by ensuring high-quality, representative multi-source data, promoting fair comparisons, enforcing regulatory oversight, and facilitating real-world clinical integration.

A fundamental limitation of current ML research is data heterogeneity and lack of representativeness, particularly concerning minority populations and underrepresented clinical cohorts. Existing open-source datasets are often insufficient in size, quality, or diversity, leading to biased models with limited generalizability. This platform would provide a large-scale, curated EHR dataset, encompassing structured and unstructured clinical data (demographics, medical history, lab results, prescriptions, and lifestyle factors), ensuring population representativeness and robust external validity.

A key feature is standardized benchmarking, addressing the lack of reproducibility and selective reporting in ML-based CVD risk prediction. All models would be evaluated using predefined metrics, including discrimination (AUC), calibration (Brier score, calibration curves), fairness audits (across demographic subgroups), and interpretability assessments. Developers must upload models, including all the technical features, the training strategy, and results, ensuring transparency, open peer review, and performance tracking, akin to version control systems. To enable fair comparisons, ML models would be benchmarked against conventional risk scores such as QRISK and Framingham, providing an objective reference standard.

Beyond research, this platform supports real-world clinical validation, bridging the gap between retrospective ML studies and clinical implementation. Top-performing models could undergo prospective validation or pilot implementations in hospital settings and be integrated into clinical decision-support systems, enabling systematic assessment of their impact on patient outcomes. Additionally, various machine learning models can be assessed by the trade-off between interpretability and performance in real-world clinical settings. By eliminating fragmented regulatory pathways, this initiative facilitates scalable, structured ML adoption, allowing policymakers to assess models through real-world performance and practical applicability rather than theoretical benchmarks, which are hard to replicate and interpret.

Ethical and regulatory compliance is a core component. The platform establishes a governance framework, mandating transparent reporting of methodologies, data usage, and compliance with regulatory standards. Unlike traditional models developed in silos, this initiative enforces strict privacy protections while enabling standardized fairness audits and bias mitigation strategies. By consolidating data access, it removes the need for multiple data-use agreements, streamlining research while ensuring accountability and regulatory oversight.

This platform also enables continuous improvement, allowing researchers to refine models based on prior results, incorporate advances in explainable AI, and explore privacy-preserving federated learning techniques. By providing a structured evaluation ecosystem, ML research can shift from ad-hoc comparisons to a standardized, iterative development pipeline, ensuring long-term adaptability and responsible innovation.

By centralizing model development, enforcing transparency, integrating regulatory oversight, and supporting clinical validation, this initiative establishes a structured pathway for ML integration into cardiovascular risk assessment, ensuring that future advancements are scientifically rigorous, clinically actionable, and equitably deployed.

4.3 Limitation

The limitations of this review include potential publication and reporting biases in the selected literature, as no formal quality assessment was conducted. The included studies may be inclined to report favourable results; however, this review qualitatively synthesizes the commonly identified opportunities and challenges across the selected studies rather than assessing biases within individual model reports, which helps contextualize potential reporting discrepancies. Another limitation is that the selected publications span a wide range of tasks and explore various CVD outcomes in diverse populations across different healthcare settings, using various data types. Only a few studies included in their reviews are directly relevant to the specific interests being addressed here. Despite this limitation, the review can still provide valuable insights, as most tasks in this field share similarities and common challenges. Additionally, some publications do not discuss EHRs comprehensively. Nevertheless, we have identified most of their attitudes toward EHRs. Considering that EHRs are used for more than this single task and are rapidly evolving in this field, we find that recent literature pays more attention to this aspect. Finally, the integrative framework we introduce has not been strictly examined and approved by any authority. However, we aim to facilitate its validation in the future to solidify the evidence base. At this point, we can only hope that it serves as inspiration for the future direction of researchers in this area.

5. Conclusions

ML-based CVD risk prediction models derived from EHRs offer immense potential but must overcome significant challenges to be clinically relevant and technically robust. Aligning with medical knowledge and clinical guidelines is crucial, as models need to be interpretable for trust and understanding. The points discussed in this review, including the latest techniques, data sources, model validation, and performance evaluation, should help guide the development of these models. Balancing computational cost, performance, and interpretability leads to the development of risk prediction algorithms that not only benefit patients but also advance scientific understanding.

Input from healthcare professionals and domain experts is invaluable in evaluating the models and identifying areas for improvement. Future research should be directed towards exploring the efficacy, usability, and impact of ML-based CVD prediction leveraging EHRs.

References

[1] Mensah GA, Fuster V, Murray CJL, Roth GA, Global Burden of Cardiovascular Diseases and Risks Collaborators. Global Burden of Cardiovascular Diseases and Risks, 1990-2022. Journal of the American College of Cardiology. 2023; 82: 2350–2473. https://doi.org/10.1016/j.jacc.2023.11.007.

[2] National Institute for Health and Care Excellence. Cardiovascular disease: risk assessment and reduction, including lipid modification. London: National Institute for Health and Care Excellence (NICE). 2023. Available at: www.nice.org.uk/guidance/ng238 (Accessed: 19 April 2024).

[3] Arnett DK, Blumenthal RS, Albert MA, Buroker AB, Goldberger ZD, Hahn EJ, et al. 2019 ACC/AHA Guideline on the Primary Prevention of Cardiovascular Disease: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. Circulation. 2019; 140: e596–e646. https://doi.org/10.1161/CIR.0000000000000678.

[4] Marx N, Federici M, Schütt K, Müller-Wieland D, Ajjan RA, Antunes MJ, et al. 2023 ESC Guidelines for the management of cardiovascular disease in patients with diabetes. European Heart Journal. 2023; 44: 4043–4140. https://doi.org/10.1093/eurheartj/ehad192.

[5] Hippisley-Cox J, Coupland C, Brindle P. Development and validation of QRISK3 risk prediction algorithms to estimate future risk of cardiovascular disease: prospective cohort study. BMJ (Clinical Research Ed.). 2017; 357: j2099. https://doi.org/10.1136/bmj.j2099.

[6] SCORE2 working group and ESC Cardiovascular risk collaboration. SCORE2 risk prediction algorithms: new models to estimate 10-year risk of cardiovascular disease in Europe. European Heart Journal. 2021; 42: 2439–2454. https://doi.org/10.1093/eurheartj/ehab309.

[7] Collins GS, Altman DG. An independent and external validation of QRISK2 cardiovascular disease risk score: a prospective open cohort study. BMJ (Clinical Research Ed.). 2010; 340: c2442. https://doi.org/10.1136/bmj.c2442.

[8] Goh LGH, Welborn TA, Dhaliwal SS. Independent external validation of cardiovascular disease mortality in women utilising Framingham and SCORE risk models: a mortality follow-up study. BMC Women’s Health. 2014; 14: 118. https://doi.org/10.1186/1472-6874-14-118.

[9] Hippisley-Cox J, Coupland CAC, Bafadhel M, Russell REK, Sheikh A, Brindle P, et al. Development and validation of a new algorithm for improved cardiovascular risk prediction. Nature Medicine. 2024; 30: 1440–1447. https://doi.org/10.1038/s41591-024-02905-y.

[10] D’Agostino RB, Sr, Grundy S, Sullivan LM, Wilson P, CHD Risk Prediction Group. Validation of the Framingham coronary heart disease prediction scores: results of a multiple ethnic groups investigation. JAMA. 2001; 286: 180–187. https://doi.org/10.1001/jama.286.2.180.

[11] Allan S, Olaiya R, Burhan R. Reviewing the use and quality of machine learning in developing clinical prediction models for cardiovascular disease. Postgraduate Medical Journal. 2022; 98: 551–558. https://doi.org/10.1136/postgradmedj-2020-139352.

[12] Goldstein BA, Navar AM, Carter RE. Moving beyond regression techniques in cardiovascular risk prediction: applying machine learning to address analytic challenges. European Heart Journal. 2017; 38: 1805–1814. https://doi.org/10.1093/eurheartj/ehw302.

[13] Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine-learning improve cardiovascular risk prediction using routine clinical data? PloS One. 2017; 12: e0174944. https://doi.org/10.1371/journal.pone.0174944.

[14] Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ (Clinical Research Ed.). 2021; 372: n71. https://doi.org/10.1136/bmj.n71.

[15] Baethge C, Goldbeck-Wood S, Mertens S. SANRA-a scale for the quality assessment of narrative review articles. Research Integrity and Peer Review. 2019; 4: 5. https://doi.org/10.1186/s41073-019-0064-8.

[16] Ouzzani M, Hammady H, Fedorowicz Z, Elmagarmid A. Rayyan-a web and mobile app for systematic reviews. Systematic Reviews. 2016; 5: 210. https://doi.org/10.1186/s13643-016-0384-4.

[17] Mohd Faizal AS, Thevarajah TM, Khor SM, Chang SW. A review of risk prediction models in cardiovascular disease: conventional approach vs. artificial intelligent approach. Computer Methods and Programs in Biomedicine. 2021; 207: 106190. https://doi.org/10.1016/j.cmpb.2021.106190.

[18] Patel R, Peesay T, Krishnan V, Wilcox J, Wilsbacher L, Khan SS. Prioritizing the primary prevention of heart failure: Measuring, modifying and monitoring risk. Progress in Cardiovascular Diseases. 2024; 82: 2–14. https://doi.org/10.1016/j.pcad.2024.01.001.

[19] Suri JS, Bhagawati M, Paul S, Protogeron A, Sfikakis PP, Kitas GD, et al. Understanding the bias in machine learning systems for cardiovascular disease risk assessment: The first of its kind review. Computers in Biology and Medicine. 2022; 142: 105204. https://doi.org/10.1016/j.compbiomed.2021.105204.

[20] Zhao Y, Wood EP, Mirin N, Cook SH, Chunara R. Social Determinants in Machine Learning Cardiovascular Disease Prediction Models: A Systematic Review. American Journal of Preventive Medicine. 2021; 61: 596–605. https://doi.org/10.1016/j.amepre.2021.04.016.

[21] Jeong K, Mallard AR, Coombe L, Ward J. Artificial intelligence and prediction of cardiometabolic disease: Systematic review of model performance and potential benefits in indigenous populations. Artificial Intelligence in Medicine. 2023; 139: 102534. https://doi.org/10.1016/j.artmed.2023.102534.

[22] Azmi J, Arif M, Nafis MT, Alam MA, Tanweer S, Wang G. A systematic review on machine learning approaches for cardiovascular disease prediction using medical big data. Medical Engineering & Physics. 2022; 105: 103825. https://doi.org/10.1016/j.medengphy.2022.103825.

[23] Banerjee A, Chen S, Fatemifar G, Zeina M, Lumbers RT, Mielke J, et al. Machine learning for subtype definition and risk prediction in heart failure, acute coronary syndromes and atrial fibrillation: systematic review of validity and clinical utility. BMC Medicine. 2021; 19: 85. https://doi.org/10.1186/s12916-021-01940-7.

[24] Krittanawong C, Virk HUH, Bangalore S, Wang Z, Johnson KW, Pinotti R, et al. Machine learning prediction in cardiovascular diseases: a meta-analysis. Scientific Reports. 2020; 10: 16057. https://doi.org/10.1038/s41598-020-72685-1.

[25] Chahine Y, Magoon MJ, Maidu B, Del Álamo JC, Boyle PM, Akoum N. Machine Learning and the Conundrum of Stroke Risk Prediction. Arrhythmia & Electrophysiology Review. 2023; 12: e07. https://doi.org/10.15420/aer.2022.34.

[26] Bozyel S, Şimşek E, Koçyiğit Burunkaya D, Güler A, Korkmaz Y, Şeker M, et al. Artificial Intelligence-Based Clinical Decision Support Systems in Cardiovascular Diseases. Anatolian Journal of Cardiology. 2024; 28: 74–86. https://doi.org/10.14744/AnatolJCardiol.2023.3685.

[27] Ciccarelli M, Giallauria F, Carrizzo A, Visco V, Silverio A, Cesaro A, et al. Artificial intelligence in cardiovascular prevention: new ways will open new doors. Journal of Cardiovascular Medicine (Hagerstown, Md.). 2023; 24: e106–e115. https://doi.org/10.2459/JCM.0000000000001431.

[28] Baashar Y, Alkawsi G, Alhussian H, Capretz LF, Alwadain A, Alkahtani AA, et al. Effectiveness of Artificial Intelligence Models for Cardiovascular Disease Prediction: Network Meta-Analysis. Computational Intelligence and Neuroscience. 2022; 2022: 5849995. https://doi.org/10.1155/2022/5849995.

[29] Hammond MM, Everitt IK, Khan SS. New strategies and therapies for the prevention of heart failure in high-risk patients. Clinical Cardiology. 2022; 45: S13–S25. https://doi.org/10.1002/clc.23839.

[30] Nadarajah R, Younsi T, Romer E, Raveendra K, Nakao YM, Nakao K, et al. Prediction models for heart failure in the community: A systematic review and meta-analysis. European Journal of Heart Failure. 2023; 25: 1724–1738. https://doi.org/10.1002/ejhf.2970.

[31] Amal S, Safarnejad L, Omiye JA, Ghanzouri I, Cabot JH, Ross EG. Use of Multi-Modal Data and Machine Learning to Improve Cardiovascular Disease Care. Frontiers in Cardiovascular Medicine. 2022; 9: 840262. https://doi.org/10.3389/fcvm.2022.840262.

[32] Liu W, Laranjo L, Klimis H, Chiang J, Yue J, Marschner S, et al. Machine-learning versus traditional approaches for atherosclerotic cardiovascular risk prognostication in primary prevention cohorts: a systematic review and meta-analysis. European Heart Journal. Quality of Care & Clinical Outcomes. 2023; 9: 310–322. https://doi.org/10.1093/ehjqcco/qcad017.

[33] Javed Z, Kundi H, Chang R, Titus A, Arshad H. Polysocial Risk Scores: Implications for Cardiovascular Disease Risk Assessment and Management. Current Atherosclerosis Reports. 2023; 25: 1059–1068. https://doi.org/10.1007/s11883-023-01173-4.

[34] Gautam N, Mueller J, Alqaisi O, Gandhi T, Malkawi A, Tarun T, et al. Machine Learning in Cardiovascular Risk Prediction and Precision Preventive Approaches. Current Atherosclerosis Reports. 2023; 25: 1069–1081. https://doi.org/10.1007/s11883-023-01174-3.

[35] Cai Y, Cai YQ, Tang LY, Wang YH, Gong M, Jing TC, et al. Artificial intelligence in the risk prediction models of cardiovascular disease and development of an independent validation screening tool: a systematic review. BMC Medicine. 2024; 22: 56. https://doi.org/10.1186/s12916-024-03273-7.

[36] Friedrich S, Groß S, König IR, Engelhardt S, Bahls M, Heinz J, et al. Applications of artificial intelligence/machine learning approaches in cardiovascular medicine: a systematic review with recommendations. European Heart Journal. Digital Health. 2021; 2: 424–436. https://doi.org/10.1093/ehjdh/ztab054.

[37] Khan MS, Arshad MS, Greene SJ, Van Spall HGC, Pandey A, Vemulapalli S, et al. Artificial intelligence and heart failure: A state-of-the-art review. European Journal of Heart Failure. 2023; 25: 1507–1525. https://doi.org/10.1002/ejhf.2994.

[38] Page MJ, Shamseer L, Tricco AC. Registration of systematic reviews in PROSPERO: 30,000 records and counting. Systematic Reviews. 2018; 7: 32. https://doi.org/10.1186/s13643-018-0699-4.

[39] Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD Statement. BMC Medicine. 2015; 13: 1. https://doi.org/10.1186/s12916-014-0241-z.

[40] Whiting PF, Rutjes AWS, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Annals of Internal Medicine. 2011; 155: 529–536. https://doi.org/10.7326/0003-4819-155-8-201110180-00009.

[41] Wolff RF, Moons KGM, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Annals of Internal Medicine. 2019; 170: 51–58. https://doi.org/10.7326/M18-1376.

[42] World Health Organization. Cardiovascular diseases: Avoiding heart attacks and strokes. 2015. Available at: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds) (Accessed: 21 April 2024).

[43] Saputro SA, Pattanaprateep O, Pattanateepapon A, Karmacharya S, Thakkinstian A. Prognostic models of diabetic microvascular complications: a systematic review and meta-analysis. Systematic Reviews. 2021; 10: 288. https://doi.org/10.1186/s13643-021-01841-z.

[44] Burlacu A, Iftene A, Popa IV, Crisan-Dabija R, Brinza C, Covic A. Computational Models Used to Predict Cardiovascular Complications in Chronic Kidney Disease Patients: A Systematic Review. Medicina (Kaunas, Lithuania). 2021; 57: 538. https://doi.org/10.3390/medicina57060538.

[45] Khan SS, Coresh J, Pencina MJ, Ndumele CE, Rangaswami J, Chow SL, et al. Novel Prediction Equations for Absolute Risk Assessment of Total Cardiovascular Disease Incorporating Cardiovascular-Kidney-Metabolic Health: A Scientific Statement From the American Heart Association. Circulation. 2023; 148: 1982–2004. https://doi.org/10.1161/CIR.0000000000001191.

[46] Herrett E, Gallagher AM, Bhaskaran K, Forbes H, Mathur R, van Staa T, et al. Data Resource Profile: Clinical Practice Research Datalink (CPRD). International Journal of Epidemiology. 2015; 44: 827–836. https://doi.org/10.1093/ije/dyv098.

[47] Li Y, Sperrin M, Belmonte M, Pate A, Ashcroft DM, van Staa TP. Do population-level risk prediction models that use routinely collected health data reliably predict individual risks? Scientific Reports. 2019; 9: 11222. https://doi.org/10.1038/s41598-019-47712-5.

[48] Amirahmadi A, Ohlsson M, Etminani K. Deep learning prediction models based on EHR trajectories: A systematic review. Journal of Biomedical Informatics. 2023; 144: 104430. https://doi.org/10.1016/j.jbi.2023.104430.

[49] Pendergrass SA, Crawford DC. Using Electronic Health Records To Generate Phenotypes For Research. Current Protocols in Human Genetics. 2019; 100: e80. https://doi.org/10.1002/cphg.80.

[50] Li I, Pan J, Goldwasser J, Verma N, Wong WP, Nuzumlalı MY, et al. Neural Natural Language Processing for unstructured data in electronic health records: A review. Computer Science Review. 2022; 46: 100511. https://doi.org/10.1016/j.cosrev.2022.100511.

[51] Shickel B, Tighe PJ, Bihorac A, Rashidi P. Deep EHR: A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record (EHR) Analysis. IEEE Journal of Biomedical and Health Informatics. 2018; 22: 1589–1604.

[52] de Mello BH, Rigo SJ, da Costa CA, da Rosa Righi R, Donida B, Bez MR, et al. Semantic interoperability in health records standards: a systematic literature review. Health and Technology. 2022; 12: 255–272. https://doi.org/10.1007/s12553-022-00639-w.

[53] Kho ME, Duffett M, Willison DJ, Cook DJ, Brouwers MC. Written informed consent and selection bias in observational studies using medical records: systematic review. BMJ (Clinical Research Ed.). 2009; 338: b866. https://doi.org/10.1136/bmj.b866.

[54] Rocher L, Hendrickx JM, de Montjoye YA. Estimating the success of re-identifications in incomplete datasets using generative models. Nature Communications. 2019; 10: 3069. https://doi.org/10.1038/s41467-019-10933-3.

[55] Zhao Y, Yu BYM, Liu Y, Tong T, Liu Y. Weight reduction and cardiovascular benefits: Protocol for a systematic review and meta-analysis. Medicine. 2018; 97: e13246. https://doi.org/10.1097/MD.0000000000013246.

[56] Damen JAAG, Hooft L, Schuit E, Debray TPA, Collins GS, Tzoulaki I, et al. Prediction models for cardiovascular disease risk in the general population: systematic review. BMJ (Clinical Research Ed.). 2016; 353: i2416. https://doi.org/10.1136/bmj.i2416.

[57] Karmali KN, Persell SD, Perel P, Lloyd-Jones DM, Berendsen MA, Huffman MD. Risk scoring for the primary prevention of cardiovascular disease. The Cochrane Database of Systematic Reviews. 2017; 3: CD006887. https://doi.org/10.1002/14651858.CD006887.pub4.

[58] Willis A, Davies M, Yates T, Khunti K. Primary prevention of cardiovascular disease using validated risk scores: a systematic review. Journal of the Royal Society of Medicine. 2012; 105: 348–356. https://doi.org/10.1258/jrsm.2012.110193.

[59] Talha I, Elkhoudri N, Hilali A. Major Limitations of Cardiovascular Risk Scores. Cardiovascular Therapeutics. 2024; 2024: 4133365. https://doi.org/10.1155/2024/4133365.

[60] Khan SS, Matsushita K, Sang Y, Ballew SH, Grams ME, Surapaneni A, et al. Development and Validation of the American Heart Association’s PREVENT Equations. Circulation. 2024; 149: 430–449. https://doi.org/10.1161/CIRCULATIONAHA.123.067626.

[61] Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ (Clinical Research Ed.). 2024; 385: e078378. https://doi.org/10.1136/bmj-2023-078378.

Funding

Engineering and Physical Sciences Research Council (EPSRC)-funded King’s Health Partners Digital Health Hub(EP/X030628/1)

Metadvice Ltd. and the National Institute for Health Research (NIHR) Biomedical Research Centre based at Guy’s and St Thomas’ NHS Foundation Trust and King’s College London(IS-BRC-1215-20006)