MARKER-HF (Machine Learning Assessment of RisK and EaRly mortality inHeart Failure): Development and Validation of a Novel Model MARKER-HF (Machine Learning Assessment of RisK and EaRly mortality inHeart Failure): Development and Validation of a Novel Model that AccuratelyIdentifies High Risk Heart Failure Patientsthat AccuratelyIdentifies High Risk Heart Failure Patients

Introduction The prevalence, costs and mortality associated with heart failure (HF) continue to rise in the United States. Identifying patients at highest risk for mortality is critical to ensure that appropriate resources can be deployed to those in greatest need. This is particularly true for patients with end-stage HF, as the gold standard therapy, heart transplant, requires significant expense and extremely limited resources. Although numerous risk prediction tools have been developed to predict mortality in HF patients, they have not achieved widespread clinical use. Some are difficult to deploy as they have many inputs which are not readily available. Most previous HF prediction tools were derived using statistical analysis methods and do not capture prognostic information in large data sets that contain multi-dimensional interactions as well as non-parametric analysis methods like machine learning (ML). Conversely, ML has long been used by other fields, including high energy physics, to discriminate between signal and background. We hypothesized that ML could be used to predict mortality in HF patients. Methods We designed MARKER-HF, Machine learning Assessment of RisK and EaRly mortality in Heart Failure, a novel model for HF mortality that discriminates between those with very high risk (signal) and very low risk of death (background). The eight variables needed to calculate our score are routinely collected in HF patients and are readily available for extraction through electronic medical record (EMR) queries. To test and validate the score, we extracted de-identified data from a retrospective cohort of 14,589 patients diagnosed with HF at the UC San Diego Health System since 2006. Results MARKER-HF achieved an AUC of 0.88 (95% Confidence Interval 0.85–0.90) in the separation of low and high-risk patients in the validation cohort. In patients who died after 90 days (intermediate-risk), increasing MARKER-HF scores were associated with decreasing time to mortality (Spearman rank correlation –0.22, p Conclusion MARKER-HF, derived using ML, is an accurate tool that discriminates between HF patients at high and low risk of mortality. Its ease of use and technical soundness greatly increase its potential for widespread adoption in guiding timely and targeted interventions, such as transplant or palliation, in HF patients. These results further support the potential of EMR-based ML analysis in medical practice.