Advancing heart failure research using machine learning.

Heart failure is a prevalent1Conrad N Judge A Tran J et al.Temporal trends and patterns in heart failure incidence: a population-based study of 4 million individuals.Lancet. 2018; 391: 572-580Summary Full Text Full Text PDF PubMed Scopus (577) Google Scholar and complex clinical syndrome associated with frequent and resource-intensive hospitalisations,2Barasa A Schaufelberger M Lappas G Swedberg K Dellborg M Rosengren A Heart failure in young adults: 20-year trends in hospitalization, aetiology, and case fatality in Sweden.Eur Heart J. 2014; 35: 25-32Crossref PubMed Scopus (121) Google Scholar reduced quality of life,3Juenger J Schellberg D Kraemer S et al.Health related quality of life in patients with congestive heart failure: comparison with other chronic diseases and relation to functional variables.Heart. 2002; 87: 235-241Crossref PubMed Google Scholar and shortened lifespan.4Alter DA Ko DT Tu JV et al.The average lifespan of patients discharged from hospital with heart failure.J Gen Intern Med. 2012; 27: 1171-1179Crossref PubMed Scopus (38) Google Scholar The heterogeneity of the syndrome is reflected by its many causes and classification systems.5McDonagh TA Metra M Adamo M et al.2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure.Eur Heart J. 2021; 42: 3599-3726Crossref PubMed Scopus (3010) Google Scholar Machine learning has demonstrated significant potential in various medical research fields and has the potential to uncover intricate associations and the ability to identify subtypes of heart failure beyond those that are currently recognised, improve risk prediction, and ultimately pave the way for personalised medicine. In The Lancet Digital Health, Amitava Banerjee and colleagues6Banerjee A Dashtban A Chen S et al.Identifying subtypes of heart failure from three electronic health record sources with machine learning: an external, prognostic, and genetic validation study.Lancet Digit Health. 2023; 5: e370-e379Google Scholar conducted an ambitious study using machine learning methods to subtype and predict the outcomes of heart failure through the analysis of large electronic health record data. The study population included a total of 313 062 patients from The Health Improvement Network and Clinical Practice Research Datalink databases, which were cross-referenced with the Hospital Episode Statistics, the UK death registry, and the UK Biobank. The authors used unsupervised and supervised machine learning methods to analyse the data. Unsupervised machine learning methods are applied to unlabelled data and aim to discover patterns and structures without predefined output or human guidance whereas supervised machine learning techniques are applied when the data is labelled, typically for predicting outcomes such as death or hospitalisation. Using four unsupervised machine learning methods the authors identified five clusters of patients with heart failure: early onset, late onset, atrial fibrillation-related, metabolic, and cardiometabolic with high external validity across datasets. The authors sought to investigate underlying biological mechanisms involved in the different heart failure subtypes by examining polygenic risk scores for 11 relevant traits and 12 single-nucleotide polymorphisms (SNPs). This analysis revealed an SNP associated with the atrial fibrillation-related subtype while polygenic risk scores for hypertension, myocardial infarction, and obesity could be linked to late-onset and cardiometabolic subtypes. Finally, using supervised machine learning the authors developed a prediction model with an online risk calculator available for both patients and clinicians. The authors’ comprehensive approach in this study deserves recognition. Unsupervised machine learning methods to subtype heart failure have been used before. Shah and colleagues7Shah SJ Katz DH Selvaraj S et al.Phenomapping for novel classification of heart failure with preserved ejection fraction.Circulation. 2015; 131: 269-279Crossref PubMed Scopus (634) Google Scholar characterised 397 patients with heart failure and preserved ejection fraction identifying three distinct clusters of patients. Consistent with Banerjee and colleagues’ study, the clusters included younger patients with a lower risk of death, patients with cardiometabolic risk factors, and older patients with the highest risk of death. Additionally, Gevaert and colleagues8Gevaert AB Tibebu S Mamas MA et al.Clinical phenogroups are more effective than left ventricular ejection fraction categories in stratifying heart failure outcomes.ESC Heart Fail. 2021; 8: 2741-2754Crossref PubMed Scopus (16) Google Scholar clustered 1693 patients into six groups, including clusters related to atrial fibrillation and clusters with higher and lower rates of cardiovascular risk factors. Supervised machine learning methods have also been used in the study of heart failure to predict outcomes such as death and heart failure hospitalisation.9Mortazavi BJ Downing NS Bucholz EM et al.Analysis of machine learning techniques for heart failure readmissions.Circ Cardiovasc Qual Outcomes. 2016; 9: 629-640Crossref PubMed Scopus (196) Google Scholar, 10Kwon JM Kim KH Jeon KH et al.Artificial intelligence algorithm for predicting mortality of patients with acute heart failure.PLoS One. 2019; 14e0219302 Crossref Scopus (57) Google Scholar The present study distinguishes itself from previous research by using a substantially larger study population and reinforcing their findings through external validation. The genetic validation sheds light on how big data associations can be linked to underlying biological processes. The study has its limitations, most of them discussed in the Article, including lack of data on ejection fraction and imaging study results in both the unsupervised and supervised models, which could have strengthened the results. The prediction model showed only moderate discriminatory ability concerning 1-year mortality and could have benefitted from focusing on individual subtypes of heart failure to reduce heterogeneity of data and, use more advanced artificial intelligence (AI) algorithms such as convolutional neural networks and eXtreme gradient boosting to improve risk prediction. The prediction tool does not have a benchmark comparison such as to ejection fraction or N-terminal pro-B-type natriuretic peptide, which could have provided further insights into the model's predictive capability. Subtyping and predicting outcomes of heart failure patients using machine learning methods trained on big data is lucrative. Identifying new ways of subtyping heart failure might facilitate the development of new treatment strategies that might benefit patients. Heart failure is evaluated and diagnosed using various imaging modalities. To advance the field, future research should also investigate the potential of AI-powered pattern recognition techniques in echocardiography and cardiac MRI to subtype heart failure, predict outcomes, and predict treatment response. I declare no competing interests. Identifying subtypes of heart failure from three electronic health record sources with machine learning: an external, prognostic, and genetic validation studyAcross four methods and three datasets, including genetic data, in the largest study of incident heart failure to date, we identified five machine learning-informed subtypes, which might inform aetiological research, clinical risk prediction, and the design of heart failure trials. Full-Text PDF Open Access