Machine Learning Sepsis Alert Notification Using Clinical Data

Overview

Machine learning is a powerful method to create clinical decision support (CDS) tools, when training labels reflect the desired alert behavior. In our Phase I work for this project, we developed HindSight, an encoding software that was designed to examine discharged patients' electronic health records (EHRs), identify clinicians' sepsis treatment decisions and patient outcomes, and pass those labeled outcomes and treatment decisions to an online algorithm for retraining of our machine-learning-based CDS tool for real-time sepsis alert notification, InSight. HindSight improved the performance of InSight sepsis alerts in retrospective work. In this study, we propose to assess the clinical utility of HindSight by conducting a multicenter prospective randomized controlled trial (RCT) for more accurate sepsis alerts.

We will evaluate the performance of HindSight in a randomized controlled trial (RCT). HindSight is a novel encoding software designed to optimize alerts for sepsis alert notification. HindSight identifies clinicians' sepsis-related decisions in the electronic health records of former patients and passes those events to InSight, thus supplying InSight with labeled examples of true positive sepsis cases for retraining. In our retrospective work, we have shown that HindSight enables InSight to adapt to site-specific deviations of real-world clinical deployment by successfully reducing false and irrelevant alarms, without human supervision. The goal of this project is to demonstrate that the retrospective success of HindSight can be successfully translated to live clinical environments. In our Phase I work, HindSight achieved an area under the receiver-operating characteristic (AUROC) of 0.899, 0.831 and 0.877 for clinician sepsis evaluation, treatment, and onset, respectively. By using an online learning algorithm to incorporate HindSight-labeled data into the InSight predictor, we showed that the online-trained InSight can adapt to the HindSight-labeled data and outperform both baseline and periodically re-trained versions of InSight (p \< 0.05). In Aim 1, we will prospectively validate HindSight's performance on real-time patient data streams in three diverse hospitals non-interventionally. In Aim 2, we will evaluate the effect of the tool in a prospective, interventional RCT. HindSight will first be evaluated by live deployment at four academic and community hospitals, during which time it will not provide alerts of future sepsis onset. Following any necessary algorithm optimization arising from live hospital validation, we will perform an RCT to evaluate reductions in false alerts from InSight trained on HindSight sepsis labels (experimental arm), compared to InSight trained on gold standard Sepsis-3 labels (control arm).