SESSION TITLE: Process Improvement in Obstructive Lung Disease Education, Pneumonia Readmissions and Rapid Response Systems II
SESSION TYPE: Original Investigation Poster
PRESENTED ON: Wednesday, October 28, 2015 at 01:30 PM - 02:30 PM
PURPOSE: Predicting venous thromboembolism (VTE) in hospitalized patients is challenging. Hospital admissions accrue significant data, but currently only limited information is used to predict VTE. The advent of the electronic medical record (EMR) offers volumes of patient data that could be used for complex risk modeling. Machine learning is a discipline of computer science that focuses on predicting outcomes in complex datasets, and potentially could be used to make comprehensive, clinically relevant VTE predictive models using EMR data. We aimed to measure the accuracy of machine learning classifiers in their ability to predict VTE in hospitalized patients.
METHODS: Using the 2012 National Inpatient Sample (NIS), we applied different machine learning classifiers to develop a predictive model of in-hospital VTE and tested its ability to predict VTE. Data features that directly suggested a VTE were removed pre-analysis. Accuracy was measured as the percentage of admissions correctly associated with VTE.
RESULTS: Of the 7,296,968 unweighted admissions in the 2012 NIS, 154,405 admissions were associated with VTE. We randomly sampled 5,000 admissions associated with VTE and 5,000 without. From this sample, we trained classifiers on 80% of the data, and then measured predictive accuracy on the remaining 20%. This train/test process was repeated 100 times. The median predictive accuracy ranged from 74% for naïve Bayes to 79.3% for random forest. The mean sensitivity and specificity for VTE prediction were 79.2% and 79.5% respectively, and the mean area under the curve in receiver operator characteristic (ROC) analysis was 0.786.
CONCLUSIONS: Machine learning classifiers were capable of predicting VTE nearly 80% of the time. This approach is promising to assess the in-hospital VTE risk of admitted patients. With further validation, such classifiers could be included in EMR systems to provide risk assessment for patient admissions.
CLINICAL IMPLICATIONS: Machine learning classifiers are capable of generating accurate VTE predictive algorithms using large comprehensive patient databases. With further development, such algorithms could be incorporated into EMRs to generate real time patient risk assessment as well as guide the use of mechanical or pharmacological VTE prophylaxis. This approach requires further development and validation as well as prospective evaluation.
DISCLOSURE: The following authors have nothing to disclose: Spencer James, Arvind Suguness, Addie Hill, Joseph Shatzel
Our abstract reports novel computer algorithms uses to predict venous thromboembolism in hospitalized patients. This predictive model is not yet approved for any purpose.