Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · ML for Cardiovascular Risk Prediction

A Systematic Review of Machine Learning Methods for Predicting Cardiovascular Disease

A 2025 systematic review screened 397 records to identify 26 eligible studies of machine learning and deep learning models for cardiovascular disease prediction. Performance estimates varied across the included studies. This article sets out what the review examined and how comparisons between model types should be interpreted, with attention to validation type, data type and metric definitions.

Pubrica AcademyWritten by a clinical data science reviewer Reviewed bya cardiovascular epidemiology reviewer Published10 October 2026 Last checked10 October 2026 Reading time7 minutes Sources10 peer-reviewed

Machine learning models for predicting cardiovascular disease range from relatively simple classifiers, logistic regression, decision trees, random forests, to deep learning architectures built on neural networks, and the choice between them is often assumed to favour complexity: more parameters, more layers, better performance. A 2025 systematic review of this literature examined both classical machine learning and deep learning techniques across the studies it included, using the metrics, sensitivity, specificity, and AUROC, conventionally used to judge predictive performance1.

Learning objectives

  • How a 2025 systematic review narrowed 397 identified records down to 26 included studies
  • Why comparisons between classical machine learning and deep learning models require careful interpretation
  • Why a systematic review's screening and inclusion process matters as much as the performance figures it reports

Common misconceptions

  • “Deep learning models are inherently more accurate than classical machine learning models for predicting cardiovascular disease.” Model complexity alone does not determine predictive accuracy. A 2025 systematic review examined both classical machine learning and deep learning techniques for cardiovascular disease prediction, and performance estimates varied across the included studies. Any comparison between the two depends on the validation approach, the type of data and the metric definitions used in each study.

Screening of 397 records to 26 included studies

Suasnabar Perez and colleagues searched Scopus and IEEE Xplore and identified 397 documents, including articles and systematic reviews, related to machine learning for cardiovascular disease prediction1. After applying inclusion and exclusion criteria, the review retained 26 open-access articles for full analysis1, a process the PRISMA 2020 reporting standard expects any systematic review to document transparently at every stage9, an exclusion rate of roughly 93% that is broadly typical of systematic reviews in a rapidly evolving, heavily published computational field, in which many identified records prove to be conference abstracts, preprints without peer review, studies using proprietary or inaccessible data, or papers that do not report the specific performance metrics required for cross-study comparison. The review grouped the approaches of the retained studies into two broad categories, machine learning (ML) techniques and deep learning (DL) techniques, and compared both with traditional, non-computational diagnostic approaches such as ECG signal interpretation, intensive care unit clinical review and echocardiographic examination1.

The screening process of a systematic review merits scrutiny before its downstream performance comparisons are relied upon, because the inclusion and exclusion criteria determine which portion of a rapidly evolving literature the review represents. The restriction to open-access articles is a reasonable and transparent choice that improves reproducibility for readers wishing to verify the included studies, but it may also have systematically excluded strong studies published in subscription-only venues. Such a non-random exclusion could, in principle, bias the resulting ML-versus-DL comparison in either direction, depending on whether open-access and subscription venues differ systematically in the modelling approach their authors tend to favour. Scopus and IEEE Xplore, the two databases searched, are strong sources for computational and engineering literature, but a search limited to them may under-represent studies published primarily in clinical or biomedical-informatics venues indexed more heavily in PubMed or Embase. This gap is relevant because clinical and engineering-oriented venues can differ in how rigorously they require external validation before publication.

Interpretation of the reported performance figures

Across the 26 included studies, classical ML and DL methods were assessed on sensitivity, specificity, precision, F1 score and AUROC (area under the receiver operating characteristic curve), a metric whose statistical meaning and interpretation derive from the foundational work applying receiver operating characteristic analysis to diagnostic test performance5. Performance estimates varied across the included studies1. The review grouped these estimates by modelling category, classical ML and DL, across the literature it included1. When such grouped comparisons are interpreted, the type of input data deserves particular attention. In general, deep learning architectures tend to show their clearest advantages on unstructured data, such as raw imaging or signal waveforms, from which they can learn feature representations directly, whereas classical ML methods often perform comparably or better on structured, tabular clinical variables (age, blood pressure, cholesterol panels) of the kind on which much cardiovascular risk prediction research is based. A dedicated empirical comparison of deep learning against classical methods on tabular data reports the same pattern4.

Critical appraisal of machine learning (ML) performance claims

Summary performance figures reported across five different metrics simultaneously should be read with caution rather than accepted at face value, because the abstract and body of the review did not report fully consistent performance figures, and performance estimates varied across the included studies1. Inconsistencies of this kind are sufficiently common in systematic reviews of a rapidly evolving computational literature that the figures should be checked directly against the individual studies summarised, rather than reliance being placed on the headline figures of the review alone. This is particularly important because many of the pooled primary studies are evaluated on different datasets, different cardiovascular endpoints and different train-test splits, factors that a review comparing only final performance metrics cannot fully reconcile. Metric definitions also warrant scrutiny, since identically named metrics may be calculated for different endpoints, thresholds or populations and may therefore not be directly comparable between studies. For a researcher using this review to inform the choice of a modelling approach, the most defensible interpretation is that performance estimates varied across the included studies, and that no summary figure should be treated as a precise comparison across 26 independently conducted studies using different datasets and evaluation protocols.

A second reason for caution concerns how performance was most probably measured across the included primary studies, although the review does not detail the validation methodology of each study. A model evaluated through internal cross-validation on a single dataset, in which one pool of patients is split into training and testing subsets drawn from the same source, will typically report higher performance than the same model would achieve on an external, independently collected validation dataset. Internal cross-validation cannot fully capture differences in patient population, data-collection protocol or measurement equipment between the original dataset and any new population to which the model might be applied. Many published machine learning studies in clinical prediction, including cardiovascular disease prediction, report only internal validation results, which tends to inflate apparent performance relative to what a model would achieve on a genuinely external population in real-world deployment. A systematic review and meta-analysis of externally validated AI-based scoring systems in intensive care found this gap to be common and often substantial6. Without the full text of all 26 included studies, the proportion that used internal rather than external validation cannot be determined, but the general pattern in this literature indicates that a high reported AUROC should be interpreted as an upper-bound estimate of likely real-world performance rather than as a guaranteed figure for clinical deployment. This caveat applies equally to classical ML and deep learning results and should temper the confidence with which the absolute or comparative figures for either category are cited.

The practical implication for a researcher designing a new cardiovascular prediction study is that the modelling approach should be selected on the basis of the data rather than on an assumption that greater complexity yields better performance. With structured clinical variables and a dataset of thousands rather than hundreds of thousands of patients, a rigorously tuned classical ML approach is a reasonable starting point, since the typical advantage of deep learning emerges mainly when sufficient data are available to learn representations that domain knowledge would otherwise have to encode manually. Deep learning is more readily justified for inputs in which it has a clearer, previously documented advantage: raw ECG waveforms, echocardiographic video, or retinal imaging linked to cardiovascular risk. In either case, the model should be evaluated with a clearly specified validation strategy, preferably including external validation, and reported with explicit metric definitions so that it can be compared fairly with other studies.

The reported performance of any machine learning model should also be compared explicitly with an established, non-machine-learning clinical benchmark before a new model is concluded to represent a meaningful improvement. The ACC/AHA Pooled Cohort Equations2, a long-established, regression-based risk calculator incorporating age, cholesterol, blood pressure, smoking status and diabetes status, built on decades of cohort-based coronary heart disease risk modelling dating back to the original Framingham risk equations7, have been validated across large population cohorts for estimating 10-year atherosclerotic cardiovascular disease risk and serve as a natural comparison point for any newer model claiming superior predictive performance. A machine learning model that only marginally outperforms this established calculator, while requiring substantially more data, computing infrastructure and specialised expertise to deploy and maintain, may not represent a practically meaningful improvement in most clinical settings, even where its AUROC is numerically higher. The figures of the 2025 review should therefore be read alongside, not instead of, the performance of these models against the benchmarks that clinicians already use, reported to the transparent standard that the TRIPOD statement sets for any multivariable clinical prediction model3. Regulatory bodies are converging on a similar expectation: the FDA action plan for AI- and machine-learning-enabled medical software calls for this kind of ongoing, transparent performance monitoring once a model moves from a published study towards clinical deployment8, a concern reinforced by documented cases in which the apparent performance of a deployed healthcare algorithm masked a significant, uncorrected racial bias in the patients it flagged as high-risk10.

Key points

Key points

A 2025 systematic review screened 397 records and included 26 studies of machine learning and deep learning models for cardiovascular disease prediction. Performance estimates varied across the included studies. Comparisons between model types should therefore be interpreted with attention to validation type, data type and metric definitions, and the review’s summary figures should be checked against the individual studies rather than treated as precisely comparable.

Practical application

Practical application

  1. Model type should be matched to the data rather than chosen by default. Deep learning’s advantages are typically clearest on unstructured inputs such as imaging or waveforms; for structured, tabular cardiovascular risk data, the choice of a more complex model should be justified rather than assumed.
  2. The category-level and headline figures of a systematic review should be checked against each other before they are cited. The abstract and body of this review contained a figure inconsistency, which underlines the need to verify a review’s summary statistics against its underlying studies.
  3. Performance-metric ranges should be treated as directional indicators rather than precise cross-study comparisons. The 26 included studies used different datasets, endpoints and evaluation protocols, and performance estimates varied across the included studies, so a summary range is best interpreted as a directional indicator rather than as a single precise figure.

Frequently asked questions

No. The review examined both classical machine learning and deep learning models across its 26 included studies, and performance estimates varied across those studies. In general, deep learning has recognised advantages on unstructured data such as raw imaging or ECG waveforms, so its role depends on the type of data and the prediction task.

The authors applied inclusion and exclusion criteria after an initial search of Scopus and IEEE Xplore, retaining only open-access articles that met their eligibility requirements. This corresponds to an exclusion rate of roughly 93%, which is broadly typical of a rapidly evolving computational research field.

Deep learning architectures tend to show their clearest advantage on unstructured data, such as imaging or waveforms, from which they can learn feature representations directly. Structured, tabular clinical variables of the kind commonly used in cardiovascular risk prediction can often be handled efficiently by classical ML methods without the much larger datasets that deep learning typically requires. A comparison that combines different data types may therefore not indicate which approach is preferable for a particular study.

Not precisely. Performance estimates varied across the included studies, which used different datasets, cardiovascular endpoints and evaluation protocols, so any summary range is best interpreted as a directional summary rather than as a single, precisely comparable statistic.

References

  1. Suasnabar Perez JM, Sulca Gomez KJ, Lozada Flores RM, Bernardo Herrera K. Predicting Cardiovascular Disease Using Machine Learning: A Systematic Review. LACCEI. 2025;1(12). read 10 October 2026. View on LACCEI Proceedings
  2. Goff DC, Lloyd-Jones DM, Bennett G, et al. 2013 ACC/AHA Guideline on the Assessment of Cardiovascular Risk. read 10 October 2026. View on PMC
  3. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement. read 10 October 2026. View on PMC
  4. Shwartz-Ziv R, Armon A. Tabular Data: Deep Learning is Not All You Need. read 10 October 2026. View on arXiv
  5. External validation of AI-based scoring systems in the ICU: a systematic review and meta-analysis. read 10 October 2026. View on PMC
  6. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143(1):29-36. read 10 October 2026. View paper (PDF)
  7. Wilson PWF, D'Agostino RB, Levy D, et al. Prediction of coronary heart disease using risk factor categories. Circulation. 1998;97(18):1837-1847. read 10 October 2026. View on PubMed
  8. U.S. Food and Drug Administration. FDA Releases Artificial Intelligence/Machine Learning Action Plan. read 10 October 2026. View on FDA.gov
  9. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. read 10 October 2026. View on The BMJ
  10. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. read 10 October 2026. View paper (PDF)

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). A Systematic Review of Machine Learning Methods for Predicting Cardiovascular Disease. Pubrica. https://pubrica.com/academy/systematic-review/machine-learning-in-cardiovascular-disease-risk-assessment/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

ML Model Selection Checklist for Cardiovascular Prediction

This checklist should be completed before a machine learning approach is selected for a cardiovascular risk prediction study.

[ ] Identified whether the input data is structured/tabular or unstructured (imaging, waveform)
[ ] Checked comparable published performance for both classical ML and deep learning on similar data types
[ ] Verified any cited review's headline figures against its individual included studies
[ ] Confirmed the dataset size is adequate for the chosen approach (deep learning typically needs larger datasets)
[ ] Specified the exact performance metrics, dataset, and evaluation protocol to be used, for fair comparison

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s data science and systematic review team checks machine learning performance claims against primary studies and matches the modelling approach to the data type, so that the choice of method in a cardiovascular prediction study is based on evidence rather than assumption.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Support for selecting modelling approaches appropriate to the data

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.