Description
Pulmonary embolism (PE) is a life-threatening clinical problem and CT imaging is the gold standard for diagnosis. However, in the past years, a substantial rise in the number of CT examinations for PE evaluation has been observed with a concomitant decrease in imaging yield. Clinical decision support rules based on PE risk scoring models to inform CT imaging decisions have been developed but are underutilized. The purpose of this study is to develop a machine learning model - PERFORM for predicting PE imaging outcomes based on patient EMR data to generate a patient-specific risk score for PE. We tested our approach on hold-out intra- and extra-institutional patient data and retrospectively compared to existing clinical risk scoring systems. Our model achieved an AUC performance of predicting a positive PE study of 0.9 (CI 0.87-0.91) on intra-institutional holdout data with an AUC of 0.71 (CI 0.69-0.72) on an external dataset from another institution; we also reported superior AUC performance and cross-institutional generalization of our artificial neural network model of 0.81 on a hold-out ED population from both intra- and extra-institution data.
Learning Objective: 1. Learn about a new generalizable machine learning model for predicting CT imaging outcome with structured EHR data.
2. Learn about a open-source EHR feature engineering pipeline that can handle multi-institutional data.
Authors:
Imon Banerjee (Presenter)
Stanford University School of Medicine
Miji Sofela, Duke University
Timothy Amrhein, Duke University
Daniel Rubin, Stanford University School of Medicine
Roham Zamanian, Stanford University
Matthew Lungren, Stanford University School of Medicine
Presentation Materials: