A computational approach to early sepsis detection

Abstract SignificanceIn this SBIR projectwe propose to improve the performance of InSighta machine learningbased sepsis screening systemin situations of limited training data from the target clinical siteThe proposed work will make possible prospective clinical deployments to sites which are smaller or lack clinical data repositoriesby significantly reducing the amount of training data necessary down to a few weeks of clinical observationClassicallya machine learning based system like InSight requires complete retraining for each new clinical settingin turn requiring a new and large collection of data from each target deployment siteWe will circumvent this requirement via transfer learning techniqueswhich transfer knowledge acquired previously in a source clinical setting to a newtarget settingResearch QuestionsWhich transfer learning methods and paired classification algorithms are most suitable for use with InSightrequiring minimal target site training data while maintaining strong performanceAre these methods and algorithms robust across the several common sepsis spectrum definitionsPrior WorkWe have developed InSight using the MIMIC III retrospective data seton which it attains an area under the receiver operating characteristic curveAUROCoffor sepsis detectionandforhour early sepsis predictionWe have also conducted pilot transfer learningexperiments in a different clinical taskmortality forecastingin which transfer learning yields afold reduction in the amount of target site training data required to achieve AUROCSpecific AimsAimto implement and assess side by side four diverse transfer learning methods for a retrospective clinical sepsis prediction taskwhere the source data set is MIMIC III and the simulated clinical target is a data set drawn from UCSFAimto determine which among the best methods from Aimalso provide robust performance when applied to two additional sepsis spectrum gold standardsMethodsWe will prepare implementations of transfer learning methods which use instance transferresidual learning and or feature augmentationkernel length scale transferand feature transferWe will test these methods with applicable classifiers on subsets of the UCSF setusing cross validation and quantifying discrimination performance in terms of AUROCThe best method classifier pairs will require no more thanexamples of septic patients from the target set and attain AUROC superiorities ofinandhour pre onset sepsis prediction detectionrelative to the best tested alternative screening systemsAimThe top three pairs will then be tested for robustness to gold standard choiceusing septic shockandhourand SIRS based sepsishourgold standardsin these testsat least one pair must again attainmargin of superiority in AUROC versus the alternative screening systemsAimFuture DirectionsThe results of these experiments will enable InSight to be robustly deployed to diverse clinical sitesyielding high performance without the need for extensive target site data acquisition Narrative Clinical decision supportCDSsystems present critical information to medical professionals by examining patient data and providing relevant informationMachine learning is a powerful method for creating CDS toolsbut accessing its full strength requires re training with retrospective data from each target clinical siteWe will use transfer learning techniques to dramatically reduce the amount of target site training data required by InSightour machine learning based CDS tool for sepsis predictionand empirically evaluate several such methods on a patient data setusing three different sepsis related gold standards