Information Extraction from EMRs to Predict Readmission following Acute Myocardial Infarction
Dartmouth College, Hanover NH
Investigators
Linked publications & trials
Abstract
Project Summary Hospitals have rapidly adopted the use of electronic medical records (EMR) for routine management and reporting of patient health care utilization. In spite of the comprehensive data collected in EMRs, they have not realized their potential for conducting routine surveillance of quality measures, for measuring hospital performance, or for surveillance of patient safety. The use of EMRs for patient safety surveillance and for predictive analytics has been underutilized especially for acute myocardial infarction (AMI). Reasons for this underuse include fragmentation of data entry and storage, poor compliance in completing structured fields for quality reporting, and the abundance of unstructured information described in narrative notes. We propose to develop a robust automated surveillance toolkit built in two independent EMRs with external validation in multiple EMRs. We will combine the rich information locked in clinical notes with structured data to quantify the risk for readmission after an AMI directly from the EMR, validate, and demonstrate its portability across institutions to other EMRs. Our overall hypothesis is that adding structured variables from the EMR with NLP-derived variables will improve our ability to predict 30-day readmission from AMI. We will evaluate this hypothesis by mapping relevant variables to common information models, developing and validating prediction models for AMI, and creating and validating a portable toolkit for generating predictive models from multiple EMRs in the following specific aims: 1) To evaluate potential AMI risk factors for 30-day readmission from AMI to a common information model using structured EMR variables and novel NLP variables extracted from EMR text;? 2) To develop an optimal prediction model for 30-day readmission from AMI at each site using registry data, structured EMR, and novel social NLP variables extracted from unstructured EMR text and to cross-validate each model at another institution;? 3) To validate an automated surveillance toolkit (ReX) for portability to three other EMRs. This research is significant in that it will improve our ability to identify AMI patients at risk of 30-day readmission, identify risk for causes of readmission for actionable intervention before readmission occurs, and for the first time provide a validated portable surveillance toolkit. Our research is innovative, because it expands the use of NLP tools to novel variables previously only obtained through manual extraction (e.g., social risk factors) and develops a generalizable and portable toolkit built in parallel on two independent EMRs with external validation in multiple EMRs. We will shift the paradigm from current single-center approaches to a 2-center parallel development and cross- validation method allowing for novel information evaluation and systematic differences in data representation between the two institutions and adapting our portable toolkit accordingly. We will significantly advance biomedical informatics tool development and our ability to perform risk assessment for AMI patients, enabling improved clinical care and improved patient outcomes.
View original record on NIH RePORTER →