Neel Shah

Pulmonary embolism affects over a million patients within the United States every year, with more than 100 000 of these cases being fatal. Current validated mortality scores including PESI and BOVA scores both account for only 30-day all-cause mortality. Unfortunately, due to the wide range of mortality within these patients, intensive care unit utilization for these patients varies from 5-90%. There are large risks for placing a patient with high risk of deterioration on the general hospital floor, and a large potential cost savings if patients can be identified who do not need intensive care services.



We developed a machine learning derived classification/decision tree to predict a 48 hour composite outcome including mortality, cardiopulmonary resuscitation, positive pressure ventilation, vasoactive use and either systematic of catheter directed thrombolysis.



We analyzed over 1650 patients within the University of Texas Southwestern system with known diagnosed pulmonary embolism over the last 6 years. We analyzed vital sign data (including heart rate, pulse oximetry, blood pressure, and respiratory rate), lab data including troponin and BNPs, chest CTs, echocardiograms (evaluating right ventricular strain and dilation), Charlson Comorbidity index, and current cancer status to predict our composite 48-hour outcome. We utilized a decision tree classifier within TensorFlow.



We developed a 48-hour model with an area under the curve of 0.84, accuracy of 92%, and positive predictive value of 64%. With further validation outside of our large quaternary academic single center, we hope to develop a clinical decision support tool that may be implemented at the bedside and eventually within the electronic health record. Overall, we hope to improve our ability to identify patients who are at the highest risk for poor outcomes within their first 48 hours of admission and improve both patient outcomes and resource utilization.