John Martin, Senior Director of Enterprise Analytics
Children's Hospital of Philadelphia
Children’s Hospital of Philadelphia (CHOP) has been determined to improve its quality of care and patient outcomes through the use of data and advanced analytics. Utilization of physician notes to extract medical knowledge to support medical practice is an area of current research interest. It implemented in its operational environment a text analytics solution to assist in the detection of a venous thromboembolism (VTE) based on unstructured data (physician note/report). Hospital-acquired VTE is currently considered the second most common contributor to harm in hospitalized pediatric patients, secondary only to central line associated infection, and is the current focus of national prevention efforts.
Current mechanisms to identify VTE events are limited and depend on manually generated clinical lists as well as post discharge ICD-9/10 review. Both processes are time consuming, error prone and do not provide immediate event identification. For this business challenge, CHOP applied natural language processing (NLP) to radiologists’ reports and found that NLP offers a fully automated solution that quickly analyzes complex batches of physician notes and offers a high level of accuracy in identifying and tracking patients with hospital-acquired VTE. The process runs on a scheduled batch on a daily basis and is available to physicians to consult as a decision support tool.
CHOP found that the use of NLP technology applied to radiologists’ reports can identify inpatients with VTE with a high sensitivity and specificity. In addition, the use of NLP identified VTEs in a rapid, automated fashion that clinicians can potentially use to optimize workflow and generate real-time alerts for VTE related QI efforts. For its implementation from July 2015 to March 2016, it processed a total of 6,373 radiology reports from 3,371 hospital encounters. The hematology VTE surveillance team clinically identified 117 patients during the study period. The inference engine correctly identified 97 of the 117 patients with a sensitivity of 82.91% and specificity of 97.51%.