Emergency Department Prediction of In-Hospital Mortality in Suspected Pulmonary Embolism: An Explainable Machine Learning Approach

TL;DR

This study uses machine learning to predict in-hospital mortality for patients with suspected pulmonary embolism in the emergency department.

Contribution

The novel approach combines explainable machine learning with established risk scores to improve mortality prediction in suspected pulmonary embolism.

Findings

Tree-based models outperformed simpler classifiers in predicting in-hospital mortality.

sPESI, oxygenation indices, and malignancy were key predictors identified through SHAP analysis.

Findings were consistent in the subgroup of confirmed pulmonary embolism cases.

Abstract

Background: Pulmonary embolism (PE) is a significant cause of cardiovascular mortality, and emergency department (ED) management requires early risk assessment to guide monitoring and disposition. Because key decisions are often needed while diagnostic evaluation is ongoing, the simplified Pulmonary Embolism Severity Index (sPESI) may provide limited discrimination for in-hospital outcomes. We evaluated whether explainable machine-learning (ML) models integrating routine ED variables with validated risk scores can predict in-hospital mortality in adults evaluated for suspected acute PE. Methods: A retrospective single-center cohort study was performed, including 220 consecutive adults evaluated for suspected acute PE in the ED between January 2021 and March 2025, comprising both PE-confirmed and PE-excluded cases. Predictors included demographics, vital signs, arterial blood gas indices, available imaging/echocardiographic findings, and Wells, Revised Geneva, and sPESI scores. Seven ML algorithms were trained and internally evaluated using the area under the receiver operating characteristic curve (AUC) and complementary metrics. Model interpretability was assessed using SHAP (SHAPley Additive exPlanations), and a sensitivity analysis was conducted in the PE-confirmed subgroup. Results: Tree-based ensemble models demonstrated higher discrimination for in-hospital all-cause mortality than simpler classifiers. SHAP analyses consistently highlighted sPESI, oxygenation/arterial blood gas indices, and malignancy as key contributors to mortality risk. Findings were similar in the PE-confirmed sensitivity analysis. Conclusions: Explainable ML models combining established risk scores with routinely collected ED variables may complement risk stratification along the suspected-PE pathway. External multicenter validation and prospective impact studies are warranted before clinical implementation.

Figures