# Predicting Mortality in Pulmonary Embolism: A Machine Learning Approach with External Validation in COVID-19 Patients

## TL;DR

This study uses machine learning to predict mortality in patients with pulmonary embolism, finding that models work better in non-COVID patients and worse in those with COVID-19.

## Contribution

The study introduces machine learning models for predicting mortality in pulmonary embolism and validates their performance in both non-COVID and COVID-19 patients.

## Key findings

- Machine learning models outperformed PESI in predicting mortality for non-COVID pulmonary embolism patients.
- Model performance significantly declined when applied to patients with concomitant COVID-19.
- Sepsis, PESI class V, and biomarkers like NT-proBNP were strongly associated with mortality.

## Abstract

Background and Objectives: Pulmonary embolism (PE) is a frequent thrombotic complication associated with SARS-CoV-2 infection and is linked to significant early mortality. Accurate early risk stratification in the emergency department (ED) remains challenging, and it is unclear how well commonly used PE prognostic tools perform in patients with concomitant COVID-19. Materials and Methods: We conducted a retrospective, single-centre study including 538 consecutive patients with acute PE and with or without confirmed SARS-CoV-2 infection admitted through the ED. Univariate analysis and machine learning models were employed to assess mortality risk. Results: In univariate analysis, mortality was strongly associated with sepsis (OR 11.68) and PESI class V (OR 5.56) and was also linked to higher neutrophil count (OR 1.19), platelet count (OR 1.12), and NT-proBNP (OR 1.20). In the non-COVID cohort, XGBoost and RF showed better discrimination than PESI class (AUC 0.864 and 0.834 vs. 0.725), while Support Vector Machines (SVM) was lower (AUC 0.740). On COVID-19 external validation, discrimination decreased: XGBoost AUC was 0.635, RF 0.614, PESI 0.584, and SVM showed no discrimination. Conclusions: ML models using routinely available ED variables improved in-hospital mortality prediction compared with PESI in non-COVID PE, but performance declined in COVID-19 patients, suggesting limited generalizability and the need for COVID-specific refinement and prospective multicenter validation.