Background: Profound therapies for heart failure with preserved ejection fraction (HFpEF) are still lacking. Recent data hinted towards possible beneficial treatment effects of spironolactone in HFpEF patients, but randomized trials failed to reach significance for hard endpoints. The heterogeneity of HFpEF has been proclaimed as the culprit of futile study results, but traditional statistical approaches are limited in their ability to truly model patients’ heterogeneity. Therefore, we aimed to identify heterogenous patient characteristics that allow to predict a positive response to spironolactone therapy among patients with HFpEF by using modern artificial intelligence algorithms in the ALDO-DHF (derivation, Aldosterone Receptor Blockade in Diastolic Heart Failure) and TOPCAT (validation, Treatment of Preserved Cardiac Function Heart Failure With an Aldosterone Antagonist) trial.

Methods: By using a novel, random allocation, -k-fold validation technique, patients from the derivation trial were randomly assigned to clusters in a reiterative pairwise permutation algorithm to identify patient groups which had a greater benefit of spironolactone with respect to improvement in E/e'. Heterogenous features of group assignment (responders and non-responders) were identified by training an extreme gradient boosting (XGBoost) algorithm. The trained model was used to identify responders and non-responders in the validation cohort. Kaplan-Meier curve analysis was performed to investigate different therapy response according to spironolactone therapy in the groups with respect to the combined endpoint of cardiovascular mortality, aborted cardiac arrest, or heart failure hospitalization.

Results: The derivation cohort consisted of 422 patients (213 spironolactone treatment, 209 placebo). Reiterative pairwise permutation identified 159 patients (38%) that showed beneficial therapeutic response to spironolactone. XGBoost identified main features are presented in Figure 1. Namely E-wave velocity, serum potassium and haematocrit were the three most important variables for therapy response.

Applying the XGBoost model criteria to the validation cohort, 185 (35 %) patients were predicted to show beneficial response to spironolactone and 340 (65 %) to show no response to spironolactone. In fact, patients with a predicted favourable response to spironolactone had a significant improved event-free survival when treated with spironolactone as compared to placebo (p=0.0076), while there was no difference according to treatment in the predicted non-responder group (p=0.52). Applying this algorithm, the number of patients needed to treat was reduced from 56 patients to 5 patients when compared to the initial results of the TOPCAT study.

Conclusion: We successfully derived and validated a machine learning model that is able to predict the therapeutic response of HFpEF patients to spironolactone. Using this algorithm, the number of patients needed to treat can drastically be reduced and allows for individualized patient selection that might show the highest benefit of spironolactone, paving the way for precision medicine.

Figure 1: Derivation and validation of a machine learning based selection algorithm for patients with heart failure with preserved ejection fraction receiving spironolactone treatment