What can machines learn about heart failure: A systematic literature review

Abstract

With the goal of producing a synthesis of pertinent findings and a critical evaluation of methodologies, applicability, and accuracy to guide future research in this area, this paper offers a systematic literature review pertaining to the application of data science and machine learning (ML) to heart failure (HF) datasets. The purpose of this work is to explore potential solutions to the problem of minimal implementation of ML methods in clinical practice. Scopus (2014–2021), ProQuest (2014–2021), and Ovid MEDLINE (2014–2021) were the databases searched for literature. Heart failure, cardiomyopathy, and data analytics, mining, or data science were among the search phrases. The evaluation included 81 out of 1688 articles. Retrospective cohort studies constituted the bulk of the research. Across all investigations, the median size of the patient cohort was 1944, with a range of 46–93,260. Prediction models for readmissions used the biggest possible patient samples, with a median of 5676 (min. 380, max. 93260). Common heart failure (HF) challenges that machine learning algorithms have attempted to address include: HF identification from existing datasets, HF mortality prediction, hospital readmission prediction after index hospitalisation, HF cohort classification and grouping into subgroups with unique characteristics, and HF therapy result prediction. Logistic regression, decision trees, random forests, and support vector machines were the most popular ML algorithms. Data about the verification of models was limited. There was a median ratio of 3:1 between IT experts and doctors, according to the authors' affiliations. Medical and IT experts worked together to write more than 50% of the research. Information technology experts who did not consult with clinicians before writing 25% of the publications. Classification models that aid in assessing the outcomes of HF patients were developed by the application of ML to datasets, namely clustering algorithms. Nevertheless, the prospective value of ML models in clinical practice is sometimes exaggerated. Designing randomised controlled trials (RCTs) using ML as an intervention arm is the next step for this field of study to prospectively evaluate these algorithms for practical clinical usage.