Predictors of acute pulmonary embolism-related hospitalizations – an artificial neural network analysis using a nationwide cohort in the United States

Abstract

Abstract Background Considering a paucity of large-scale data on predictors of pulmonary embolism (PE) and its higher association with complications and worse outcomes, we aimed to determine the predictors of PE in this United States population-based analysis using Artificial Neural Network (ANN) Model in a nationally representative cohort. Methods We identified PE-related hospitalizations using 2018's National Inpatient Sample database. The relevant predictive factors for ANN were selected for this cohort. Of all admissions (unweighted n=7,105,498, weighted n=35,527,481), PE cohort (weighted n=387805) consisted of 1.1% of all admissions in 2018. The 2018 cohort was randomly split into training data (unweighted n=4716132, 70.0%) which were used to calibrate ANN and testing data (unweighted n=2019290, 30%) which were used to assess the accurateness of the algorithm. We equated the rate of incorrect prediction between training and testing data and measured the Area under Receiver Operator Curve (AUC) to determine ANN's efficacy in predicting PE hospitalizations. Results Patients hospitalized with PE often consisted of older (mean age 62.5±17.1 years), female (51.3%), white (70.5%) patients, and patients from lower-income quartile (0–25% income quartile: 28.8%%), often admitted non-electively (93.7%) with higher rates of cardiovascular disease risk factors. PE admissions revealed significantly higher (6.5% vs. 1.9%, p<0.001) in-hospital mortality, less frequent routine discharges (51.4% vs. 68.1%) and more frequent other facility transfers and requirement of home health care. Normalized Predictors of PE admissions are displayed in Fig. 1. Our ANN model had AUC 0.873 which correlates with an excellent prediction model. Our data demonstrated low levels (0.8%) error in both testing and training models. Conclusion Our ANN model showed high performance to predict risk factors for PE admissions in the US population. It will enable clinicians to screen patients at high-risk for PE admissions, curtail complication rate, improve survival and lower the healthcare cost. Funding Acknowledgement Type of funding sources: None.