Main author: Brynjolfur Gauti Jonsson
Institution or Company: Icelandic Heart Association. Faculty of Medicine, University of Iceland.
Co-Authors, Institution or Company:
Valborg Gudmundsdottir, Icelandic Heart Association, Faculty of Medicine, University of Iceland. Valur Emilsson, Icelandic Heart Association. Faculty of Medicine, University of Iceland, Vilmundur Gudnason, Icelandic Heart Association. Faculty of Medicine, University of Iceland, Thor Aspelund, Icelandic Heart Association. Centre of Public Health Sciences, University of Iceland.
Introduction: Heart failure is characterized by the heart’s inability to maintain a sufficient blood supply in response to demand, and it contributes significantly to global disease burden and mortality. Heart failure is the end-stage of a variety of cardiac insults triggered by a range of different known and unknown factors. In this study, we used deep serum proteomic measurements to model the risk of developing heart failure.
Methods: AGES-Reykjavik is a population-based study of the elderly (N=5457), in which each participant is subjected to extensive phenotyping and serum proteomic measurements. The nonparametric bootstrap and LASSO operator were used to approximate the sampling distribution of proteomic variables’ coefficients in logistic and cox proportional hazards regression models as well as their predictive performance as measured by the area under the curve (AUC) and Harrel’s concordance index (c-index). A protein was classified as important to the prediction problem if it had an estimated non-zero coefficient in at least 80% of the bootstrap iterations.
Results: For prevalent heart failure the expected out-of-sample AUC was 0.85 (95%CI 0.81 – 0.88). For incident heart failure the expected out-of-sample c-index was 0.76 (95%CI 0.73 – 0.79). Twelve proteins were deemed important for predicting incident heart failure only, 14 for predicting prevalent heart failure only and one for predicting both.
Conclusions: Analyzing the approximate sampling distributions obtained by applying the nonparametric bootstrap along with the LASSO operator has given valuable insights as to which proteins are important to the problem of predicting the risk of prevalent and incident heart failure.
