Skip to content

Abstract Search

Primary Submission Category: Heterogeneous Treatment Effects

A Statistical Reinforcement Learning Approach to Personalize Renal Replacement Therapy Initiation in the ICU

Authors: François Grolleau, François Petit, Stéphane Gaudry, Raphaël Porcher,

Presenting Author: François Grolleau*

Trials sequentially randomizing patients each day have never been conducted for renal replacement therapy (RRT) initiation. We used data from electronic health records and clinical trials to learn and validate optimal dynamic strategies for RRT initiation in the ICU.

We included participants from the MIMIC-III database for development and two randomized trials for evaluation. We used a doubly-robust dynamic treatment regimen estimator to learn when to start RRT after the occurrence of acute kidney injury. The decision rule to initiate RRT mimicked that of clinicians i.e., decisions are re-evaluated every day—for three days in a row—given patients’ evolving characteristics. The ‘crude strategy’ aimed to maximize hospital-free days at day 60. The ‘stringent strategy’ recommended initiating RRT only when there was evidence at the 0.05 threshold that a patient would benefit from initiation. For evaluation, we estimated the causal effects of implementing our learned strategies versus following current best practices using the advantage doubly robust estimator with terminal states.

We showed that implementing our strategies could improve the average number of days ICU patients spend alive and outside the hospital. The stringent strategy entailed less frequent usage of RRT and could help save important health resources all the while reducing unnecessary treatment burdens. We developed a practical and interpretable dynamic decision support system for RRT initiation in the ICU.