Data Mining the Electronic Medical Record with Intelligent Agents to Inform Decision Support Systems

Abstract

An intelligent agent framework is used on an ICU EMR to create prediction models for disease onset. Eleven models are created to inspect 5 diseases: acute respiratory distress syndrome (ARDS); severe acute hypoxemic respiratory failure (SAHRF); acute kidney injury (AKI); sepsis; and disseminated intravascular coagulation (DIC). Four of the models (ARDS, AKI Stage 1, AKI Stage 2, and sepsis) are competitive or superior to the best comparable peer-reviewed models. The other seven are novel, including: SAHRF (AUC=0.952); DIC from ARDS positive patients (AUC=0.722); ARDS from DIC positive patients (AUC=0.675); AKI Stage 3 (AUC=0.983); the progression from AKI Stage 1 to Stage 2 (AUC=0.930); the progression from AKI Stage 2 to Stage 3 (AUC=0.951); and DIC (AUC=0.838). In derivative work: a correlation between pre-DIC patients and metabolic acidosis is shown, a meta-analysis on misclassified patients is given, a disease pathway that demon- strates how ARDS and DIC can interact in a positive feedback loop is presented. DIC is shown to be implicated in 78% of all in-hospital mortality of ARDS patients.