Individualized Treatment Rule Identification in Heart Failure Patients Using Precision-medicine Methods

Background: In the last decade, the fields of statistical learning and causal inference have blended to generate an alternative to average treatment effect (ATE) estimation: individualized treatment rules (ITRs). Where an ATE generates a single treatment rule for the entire population, ITRs generate different rules conditional on an individual's history. We applied some of the best-performing, ITR-identifying, precision-medicine methods to a heart failure (HF) setting. Torsemide and bumetanide have superior pharmacokinetic profiles compared to furosemide, but which sub-populations benefit the most from superior profiles remains unclear.

Objectives: To identify an optimal ITR for loop diuretic choice at HF hospitalization discharge, and to compare outcomes under this rule with outcomes under a one-treatment-for-all rule.

Methods: We identified Medicare beneficiaries who were new users of furosemide, torsemide or bumetanide from 2007-2017, with an inpatient admission for HF. We excluded users with a loop diuretic in the six months prior to index HF hospitalization. Baseline covariates included patient demographics, comorbidities, concurrent medications, and dose equivalencies. Propensity scores were generated using a subset of baseline covariates identified using random forests (RF). Outcomes included all-cause mortality or a composite outcome of all-cause mortality and HF readmission at multiple time points. We compared the optimized reward (number of days until outcome) under a one-treatment-for-all rule and an ITR identified from three precision-medicine methods to identify optimal treatment rules: efficient augmented relaxation learning (EARL), residual weighted learning (RWL), and RF. Using these methods, we estimated the optimized reward for both the ITR and a one-treatment-for-all rule, and calculated the difference in optimized reward between the two approaches.

Results: The 3 precision medicine models performed similarly. Compared to one-treatment-for-all rule, ITRs generated the greatest increase in optimized reward for the composite outcome of HF readmission and all-cause mortality within 1 year: 5.3 (SE: 1.2) days for RWL, 7.8 (SE: 3.5) days for RF, 8.0 (SE: 2.0) days for EARL.

Conclusions: We demonstrate the use of precision-medicine methods to identify ITRs and generate real-world treatment recommendations that are more tailored than the recommendations from an ATE. While several challenges remain in improving inference, interpretation, and decision support development for precision medicine, these results suggest ITR identification might improve outcomes in a HF setting. This approach can be extended to other pharmacoepidemiologic settings, assuming a causal framework is appropriately fit prior to estimating the ITR.