Autonomous system supporting patient-specific transfer and discharge decisions

Period of Performance: 01/01/2017 - 06/30/2017

$348K

Phase 1 SBIR

Recipient Firm

Dascena
HAYWARD, CA 94541
Principal Investigator

Abstract

Significance: In this SBIR project, we propose to improve the utility of AutoTriage, a machine-learning based clinical decision support (CDS) system, by integrating clinician intervention medical information into its predictions. Despite identified needs for CDS systems in patient transfer and discharge decisions, existing tools do not meet high standards for sensitivity and specificity. This is because current CDS methods are unable to distinguish changes in patient health due to clinician intervention from those arising due to an internal homeostatic mechanism. Thus, for example, existing tools may erroneously suggest discharge for a patient currently undergoing a life-sustaining treatment. Research Question: Can machine learning principles be used to create a classifier which incorporates signs of clinical intervention to inform transfer and discharge decision support, ultimately leading to higher quality predictions? In addition, will such a tool be able to maintain its performance when tested on a different patient population or one for which the data quality is much poorer? Prior Work: We have developed AutoTriage, a machine learning-based CDSS for 12-hour mortality prediction. On the publicly available MIMIC-III retrospective data set, this system attains an area under the receiver operating characteristic curve (AUROC) of 0.88, which is superior to commonly used triage scores MEWS (AUROC = 0.75), SOFA (0.71), and SAPS-II (0.72) on the same data set. Specific Aims: To integrate clinician intervention information into existing AutoTriage software (Aim 1), and to test the robustness of this modified tool to changes in patient population and data quality (Aim 2). Methods: We will create gold standards for periods of clinician intervention, using chart events and keywords from clinician notes. Then, we will train a binary classifier for identifying these periods and, ultimately, use the classifier to modify AutoTriage scores. Robustness studies will be performed on the retrospective UC ReX and sparse MIMIC III databases. Successful completion of Aim 1will be demonstrated if 75% of all hours of clinician intervention are correctly classified, if the test-set area under the ROC curve improves by 5% over its current value, and if 30-day readmission predictions are 10% more accurate for patients treated within the last hour. Aim 2 will be completed if AutoTriage ROC area performance is within ± 0.10 of its original value for both UC ReX and sparse MIMIC III sets. Future Directions: Following the proposed work, the AutoTriage system will be deployed at the sites of our ongoing clinical implementations. During this study, we project that AutoTriage will assess mortality risk for 25,000 ICU patients per year, helping clinicians more effectively allocate interventions totaling $15 million.