Using clinical treatment data in a machine learning approach for sepsis detection

Abstract SignificanceIn this SBIR projectwe propose to develop novel softwareHindSightthat will improve InSighta machine learning based clinical decision supportCDSsystem for sepsis prediction and detectionHindSight will identify clinicianssepsis related decisions in the records of former patientsit will then use these events to supply InSight with labeled examples of sepsis casesincorporating cliniciansjudgement to demonstrate appropriate and inappropriate alarmsTogether with an online training module that accordingly refines InSight s predictorsthis capability will enable InSight to quickly adapt to the idiosyncrasies of a particular clinical deploymentreduce false or irrelevant alarmsand do both without explicit human supervisionResearch QuestionCan a machine learning basedretrospective labeler learn to autonomously label sepsis and sepsis treatments by integrating the total clinical recordthereby providing many high quality examples and labels for training a sepsis CDSIn concert with an online learning algorithmcan this labeler facilitate onlinesupervised learning without explicit human interventionPrior WorkWe have developed InSight for application in a number of sepsis prediction settingsExisting InSight classifiers attain an area under the receiver operating characteristic curveAUROCoffor sepsis detectionandforhour early sepsis predictionSpecific AimsTo identify patients who were evaluated for sepsistreated for sepsisor who actually had sepsis using the retrospective clinical patient record and label them accordinglyAimto use these labels with an online learning algorithm to implement autonomoussupervised learning of alert behavior which reflects clinician judgementAimMethodsWe will identify evaluatedtreatedand septic patients using a machine learning labeler trained on retrospective data from patientselectronic health recordsEHRat time of dischargeUsing a set oftest casesseptichand annotated by our clinician investigatorswe will assess the labeler s performanceLabeling AUROCwill constitute success in AimWe will develop an online learning algorithm which enables InSight to continuously retrain during deploymentWith Aims labelerwe will simulate a deployment with online learningproducing a learning curve of predictive AUROC on a held out test set versus number of observed patientsAimwill be successful if the online training results in superior area under the learning curve versus the initial model and periodic retrainingAll experiments will be executed using the MIMIC III data setFuture DirectionsFollowing the proposed workthe InSight system with an onlineHindSight based retraining module will be deployed at partner hospitals for prospective studies Narrative Clinical decision supportCDSsystems present critical information to medical professionals by examining patient data and providing alertsMachine learning is a powerful method for creating CDS toolsbut it requires labels which reflect the desired alert behaviorWe will develop software that examines discharged patientselectronic health recordsEHRidentifies clinicianssepsis treatment decisions and patient outcomesand passes these labeled examples to an online algorithm for retraining InSightour machine learning based CDS tool for real time sepsis prediction