Source-linked AI summary
Machine Learning-Based Heart Disease Diagnosis: A Systematic Literature Review
Md Manjurul Ahsan, Zahed Siddique
TL;DR
Imbalanced ECG and patient data complicate unbiased machine-learning-based heart disease diagnosis. This study synthesizes the literature through a systematic review and reports continuing open problems, especially the need for real-world data and interpretable predictions.
Problem
Imbalanced ECG and patient data challenge traditional machine-learning models' ability to perform unbiased heart disease diagnosis.
Method
The study conducts a PRISMA-aligned systematic literature review, preceded by a meta-analysis of 451 publications and followed by in-depth analysis of 49 papers.
Results
The review identifies open problems in imbalanced-data handling; deep learning dominates, while SMOTE remains a popular oversampling technique.
Takeaways & Limitations
Real-world systems require experiments with real-time patient data, interpretable predictions, and demonstrations of trustworthy medical performance.
Takeaways & Limitations
Many studies rely on Cleveland and MIT-BIH datasets, while relatively few use real-world data, limiting the authenticity of reported performance.
Abstract
from arXiv · showhide
Heart disease is one of the significant challenges in today's world and one of the leading causes of many deaths worldwide. Recent advancement of machine learning (ML) application demonstrates that using electrocardiogram (ECG) and patient data, detecting heart disease during the early stage is feasible. However, both ECG and patient data are often imbalanced, which ultimately raises a challenge for the traditional ML to perform unbiasedly. Over the years, several data level and algorithm level solutions have been exposed by many researchers and practitioners. To provide a broader view of the existing literature, this study takes a systematic literature review (SLR) approach to uncover the challenges associated with imbalanced data in heart diseases predictions. Before that, we conducted a meta-analysis using 451 referenced literature acquired from the reputed journals between 2012 and November 15, 2021. For in-depth analysis, 49 referenced literature has been considered and studied, taking into account the following factors: heart disease type, algorithms, applications, and solutions. Our SLR study revealed that the current approaches encounter various open problems/issues when dealing with imbalanced data, eventually hindering their practical applicability and functionality.
1 Introduction
Heart disease diagnosis using machine learning has expanded because ECG and patient data can support early detection, but imbalanced data complicate unbiased modeling. This review synthesizes the literature to characterize current approaches, imbalance-handling strategies, and remaining research gaps.
- Machine learning-based heart disease diagnosis is presented as a cheap and flexible approach for detecting disease from clinical data.
- Deep learning models, including CNN and DNN systems, have reported heart disease diagnosis accuracy close to 100%.
- The review analyzes 451 papers from Scopus for metadata patterns and 49 papers in depth for algorithms, applications, disease types, and imbalance strategies.
- The in-depth review asks which machine learning approaches are used and how researchers handle datasets with imbalanced class ratios.
- The stated goal is to provide a reference for theorists and practitioners while identifying research gaps for advanced heart disease diagnosis models.
2 Methods
The study uses a systematic literature review following explicit search, selection, and appraisal procedures, reported according to PRISMA recommendations. Searches in Scopus were filtered by date, document type, language, subject area, and keywords before screening and exclusion.
- A systematic literature review formulates questions and uses explicit procedures to find, select, appraise, and synthesize relevant research.
- The review follows PRISMA recommendations to improve clarity and transparency through an evidence-based checklist and four-phase analysis.
- The Scopus search covered publications from 2012 through November 15, 2021 using heart, machine learning, imbalance, diagnostic, and deep learning keywords.
- 5055 initial Scopus articles were reduced to 2710 after the year filter and 468 after document, language, subject-area, and keyword restrictions.
- The screening excluded studies without imbalance-data analysis, studies focused only on model performance, non-peer-reviewed articles, and inaccessible full texts.
3 Observations and findings
The study presents findings from a metadata analysis of 451 selected papers and a content analysis of 49 publications, with the review process illustrated through a PRISMA flow diagram.
- The findings combine metadata analysis of 451 selected papers with content analysis of 49 publications.
- Figure 1 depicts the PRISMA approach used in the research.
3.1 Metadata analysis
The metadata analysis shows strong recent growth in publications on machine-learning-based heart disease diagnosis, alongside diverse journal and institutional participation. Scientific Reports published the largest number of papers among the listed journals, while publication leadership varied across countries and institutions.
- The metadata study organizes 451 papers by year, journals, authors, countries, subject areas, funding, and institution.
- Publication by year: 92 papers were published in 2021 compared with around 48 in 2020, alongside increased attention to imbalanced classification in heart disease diagnosis.
- Scholarly journal articles published between 2012 and 2021: Scientific Reports published 8 of the 451 papers, representing 1.77 percent, and no single journal dominated the field.
3.1.3 Publication by authors
Publication output among individual authors was modest across the ten-year MLBHDD literature, with Mahek Shah contributing the most papers. Several authors ranked closely behind, while many top-ten authors published only three papers.
- Mahek Shah published the most MLBHDD papers, with 6 of 451 articles.
- Javier Ripollés-Melchor and Gopal Krushna Pal ranked second, each publishing five articles.
- Fourteen other authors published three papers each and were included among the top ten authors.
- Individual author publication counts remained relatively low compared with the ten-year analysis period.
3.1.4 Publication by citations
The citation analysis identifies the most influential MLBHDD publications through November 2021 and shows that the cited papers span the 2012–2021 period. Acharya et al.’s 2017 paper received the highest citation count in the review.
- The ten most cited papers considered MLBHDD between 2012 and 2021.
- Citation totals may differ between Scopus and Google Scholar because their indexing procedures differ.
- 448 citations and 89.6 citations per year were recorded for Acharya et al.’s 2017 paper, the maximum in the citation analysis.
- Citation counts for the cited authors ranged from 28 to 448 times.
3.1.6 Most frequently words used in the titles and keywords
Keyword analysis shows that heart disease terminology dominates both article titles and author-keyword sections, while machine learning and deep learning are also prominent in author-selected keywords. Word-cloud size represents term frequency.
- Titles: 115 occurrences made “heart” the most frequent single title keyword, while “heart failure” and “heart rate variability” led double and triple keywords with 38 and 17 occurrences.
- Author keywords: “Heart failure” and “heart rate variability” were the most common author keywords, each appearing 35 times.
- Author keywords: “Machine learning” appeared 27 times and “deep learning” 17 times in the author-keyword sections.
- Titles: In the titles, only “heart” and “disease” ranked among the most frequently used single keywords despite the review’s focus on machine learning and imbalance.
- Word-cloud interpretation: The word cloud uses larger, bolder fonts for more frequently used words and smaller fonts for less frequent phrases.
3.1.7 Publication by institutions
Institutional contributions were distributed across multiple subject areas and countries. Fukushima Medical University led institutional publication counts, while medicine dominated the subject-area distribution.
- Institutions: Fukushima Medical University published 27 MLBHDD articles, approximately 6% of the 451-paper literature.
- Institutions: The University of Oxford ranked second and the University of São Paulo Medical School ranked third by institutional publication count.
- Subject areas: Medicine accounted for approximately 55% of the pooled literature, followed by computer science at 11.10% and biochemistry at 10.30%.
- Subject areas: Engineering, health professionals, and decision sciences also contributed papers, indicating multidisciplinary participation.
3.1.9 Publication by funding
Funding and publication activity in MLBHDD fluctuated from 2012 to 2018 but increased exponentially from 2019 to 2021, reflecting growing attention from researchers, practitioners, and funders.
- 2019–2021 saw exponential growth in funded and published MLBHDD research.Publication counts associated with funding sources fluctuated between 2012 and 2018 before rising sharply.
- The referenced figures organize publication activity by institutions, subject areas, and funding sources over time.
3.2 Insights of MLBHDD
The review analyzes 49 studies on imbalanced heart-disease data across diseases, applications, algorithms, and imbalance solutions. Arrhythmia and deep learning are prominent themes, while studies apply varied data- and algorithm-level strategies.
- 49 studies were analyzed for concepts, techniques, trends, and future scopes related to imbalanced heart-disease diagnosis.
- Disease types: Arrhythmia is a major focus because abnormal heartbeats have fatal consequences and require early diagnosis.At least 13 selected papers considered arrhythmia.
- Disease types: The literature covers cardiac arrest, coronary heart disease, and myocardial infarction in addition to arrhythmia.
- Machine learning algorithms: At least 27 of 49 studies used deep-learning approaches, exceeding the use of traditional machine learning methods.
- Imbalance solutions: Studies addressed imbalanced classes through data-level and algorithm-level approaches, including feature selection and cost-sensitive classification.Gan et al. proposed AdaC-TANBN as a cost-sensitive method for the Cleveland dataset.
4 Discussions
The review identifies arrhythmia as the most widely studied disease and notes that dataset choice affects reported model performance. It argues that real-world, continuously updated data and interpretable predictions are needed for practical clinical systems.
- Arrhythmia is the most widely studied heart disease among the reviewed MLBHDD literature.
- Most studies use the Cleveland and MIT-BIH arrhythmia datasets because they are available and present imbalance challenges.
- Performance varies between open-repository and real-world data, with real-world experiments described as more authentic.
- Clinical decision-support systems require continuous model adjustment using new patient data because instability limits reliance on old data alone.
5 Conclusions
The reviewed studies apply diverse machine-learning models and imbalance-handling methods to heart-disease diagnosis across several diseases and datasets. The paper concludes that practical deployment requires real-time patient data and interpretable predictions.
- The review calls for experiments using real-time patient data and interpretable machine learning in real-world heart-disease diagnosis systems.
- Datasets: The literature includes ECG, clinical, imaging, and hospital datasets, including Cleveland, MIT-BIH, MIMIC-III, NHANES, and Framingham.
- Imbalance solutions: Data-level methods include SMOTE, oversampling, undersampling, augmentation, and synthetic-data generation, while algorithm-level methods include focal and cost-sensitive losses.
- Applications: Arrhythmia, myocardial infarction, coronary heart disease, cardiac arrest, and heart failure appear among the reviewed applications.