Abstract Body (Do not enter title and authors here): Introduction: Prognostic metrics like NT-proBNP are often low or even normal in patients in Heart Failure with Preserved Ejection Fraction (HFpEF), particularly in HFpEF patients with severe obesity. Here, we use machine learning survival prediction models to investigate predictors of outcomes in HFpEF.

Methods: Patients enrolled in the Johns Hopkins Hospital HFpEF registry from July 2014 to May 2024 were included. HFpEF was defined based on clinical signs and symptoms, echocardiographic and laboratory findings, or by right heart catheterization. Starting with 44 features, a two-stage model, in which variables associated with at least a 15% change in hazard ratio by univariate Cox regression were included for subsequent machine learning. Models tested include a random survival forest with and without gradient boosting. Model accuracy was assessed on the test set with a C-statistic.

Results: A total of 356 HFpEF patients were included, with 284 (80%) assigned to a train set and 72 (20%) to the test set. A gradient boosted random survival forest was found to be the best model with hyperparameter tuning, with a C-statistic of >0.99 and 0.87 in the train and test set, respectively. The most important prognostic features include N-terminal brain natriuretic peptide, systolic blood pressure, age, and the ratio of cystatin C to albumin. Individually, each of these features associate with survival with a C-statistic < 0.70, which is worse compared to collective model performance (Fig 1A). Other relevant features are shown in Fig 1B. For more effective clinical translation, models including the top 15, top 10, and top 5 features were subsequently trained and were found to have similar performance to the model with all features.

Conclusion: Here we report a prognostic model for mortality prediction in a cohort of obese HFpEF patients. Efforts to benchmark this model against other HFpEF prognostic models and test in an external HFpEF patient cohort are underway.