AbstractImportanceHeart failure constitutes a high burden on patients and society, but although lifetime risk is high, it is difficult to predict without costly or invasive testing. Knowledge about novel risk factors could enable early diagnosis and possibly preemptive treatment.ObjectiveTo establish new risk factors for heart failure.DesignWe applied supervised machine learning in UK Biobank in an agnostic search of risk factors for heart failure. Novel predictors were then subjected to several in-depth analyses, including multivariable Cox models of incident heart failure, and assessment of discrimination and calibration.SettingPopulation-based cohort study.Participants500,451 individuals who volunteered to participate in the UK Biobank cohort study, excluding those with prevalent heart failure.Exposure3646 variables reflecting different aspects of lifestyle, health and disease-related factors.Main OutcomeIncident heart failure hospitalization.ResultsMachine learning confirmed many known and putative risk factors for heart failure, and identified several novel candidates. Mean reticulocyte volume appeared as one novel factor, and leg bioimpedance another; the latter appearing as the most important new factor. Leg bioimpedance was significantly lower in those who developed heart failure (p=1.1x10-72) during up to 9.8-year follow-up. When adjusting for known heart failure risk factors, leg bioimpedance was inversely related to heart failure (hazard ratio [95%CI], 0.60 [0.48–0.73]) and 0.75 [0.59–0.94], in age- and sex-adjusted and fully adjusted models, respectively, comparing the upper vs. lower quartile). A model including leg bioimpedance, age, sex, and self-reported history of myocardial infarction showed good predictive capacity of future heart failure hospitalization (C-index=0.82) and good calibration.Conclusions and RelevanceLeg bioimpedance is inversely associated with heart failure incidence in the general population. A simple model of exclusively non-invasive measures, combining leg bioimpedance with history of myocardial infarction, age, and sex provides accurate predictive capacity.Key pointsQuestionWhich are the most important risk factors for incident heart failure?FindingsIn this population-based cohort study of ~500,000 individuals, machine learning identified well-established risk factors, but also several novel factors. Among the most important were leg bioimpedance and mean reticulocyte volume. There was a strong inverse relationship between leg bioimpedance and incident heart failure, also in adjusted analyses. A model entailing leg bioimpedance, age, sex, and self-reported history of myocardial infarction showed good predictive capacity of heart failure hospitalization and good calibration.MeaningLeg bioimpedance appears to be an important new factor associated with incident heart failure.
Response to Comment on Segar et al. Machine Learning to Predict the Risk of Incident Heart Failure Hospitalization Among Patients With Diabetes: The WATCH-DM Risk Score. Diabetes Care 2019;42:2298–2306
