Abstract 106: Predicting Pulmonary Hypertension Readmission Using Machine Learning Models

Abstract

Objectives: Pulmonary hypertension (PH) is increasingly recognized but understudied. We sought to determine predictors of 30-day readmission after PH-related hospitalization using Driverless Artificial Intelligence (DAI) algorithm. Methods: We utilized the 2012-2014 Healthcare Cost and Utilization Project Nationwide Readmission Database (NRD) that accounts for weighted estimates of roughly 35 million discharges in the United States. Adult patients with primary ICD-9-CM diagnosis codes of 416.0 and 416.8 for primary and secondary PH with an index admission and any readmission within 30-days of the index event were identified. We assessed predictors of all cause 30-day readmission from clinical, hospital and 29 AHRQ comorbidity measures using predictive modeling to provide interpretable risk factors globally at population level and locally associated with each discharge. The risk prediction model was developed using DAI algorithm and results compared to models based on LIME GLM, Random Forest, and Decision Tree. Overall model performance was assessed using concordance statistic. Results: Data from patients with 14,659 admissions and 2797 30-day readmissions were used to train the model. After data processing, the final model included 3674 variables. The Random Forest model had the best performance with c-statistic of 0.70 and surrogate models compared at AUCs of 0.51(Decision Tree), 0.62 (LIME) and (DAI) 0.59. Global features that mainly important to the overall prediction of DAI model were deficiency anemia, APR-DRG severity risk, APR-DRG mortality risk, renal failure and CHF. Conclusion: Using predictive models with high predictive power, interpretable risk factors and prediction accuracy may enable health care systems to accurately target high-risk PH patients and prevent recurrent readmissions.