A machine learning framework integrating echocardiographic, biomarker, and clinical data predicted in-hospital mortality in acute pulmonary embolism with an AUC of 0.95.
Observational (n=497)
Does a machine learning framework integrating echocardiographic and biomarker data predict in-hospital mortality in adult patients with acute pulmonary embolism?
A machine learning framework integrating echocardiographic, biomarker, and clinical data demonstrated high discriminative ability (AUC 0.95) for predicting in-hospital mortality in acute pulmonary embolism.
Effect estimate: AUC 0.95
Abstract Introduction Risk stratification is an important part of management in acute pulmonary embolism. Currently, Pulmonary Embolism Severity Index (PESI) is used to stratify mortality risk but does not incorporate objective physiological data such as biomarkers or echocardiographic findings. Although management decisions often depend on evidence of right ventricular dysfunction from echocardiography or BNP and troponin, no existing predictive model integrates these multimodal data points. This project aims to develop and evaluate a machine learning-based framework to identify echocardiographic and biochemical predictors of in-hospital mortality among patients with acute PE. Methods We conducted a retrospective analysis of adult patients in the MIMIC-III dataset and ECHO-NOTE2NUM who were admitted with acute PE and underwent transthoracic echocardiography. Structured echocardiographic data reports, including right and left ventricular cavity size, wall motion abnormalities, valvular dysfunction and estimates of right heart pressure, were merged with laboratory (BNP, troponin) and demographic variables.Data preprocessing included cleaning, imputation, and feature harmonization. Univariate analyses were performed using Mann-Whitney U and chi-squared tests to assess associations between clinical and imaging parameters and the primary endpoint of in-hospital mortality. A Random Forest classifier was implemented to explore integrated feature-based mortality prediction. Results 497 adult patients were included in the analysis. The overall in-hospital mortality rate was 12%. Right and left ventricular cavity abnormalities were strongly associated with in-hospital mortality (p 0.01). Wall thickness, diastolic function and valvular pathology were not independently associated with mortality. The Random Forest model achieved an area under the receiver operating curve (AUC) of 0.95, a sensitivity of 0.67, a specificity of 0.94, and an F1-score of 0.64. Feature importance analysis identified elevated lactate, creatinine, troponin and white blood cell count as key predictors. Echocardiographic indicators such as right ventricular dilation, elevated pulmonary pressures and tricuspid regurgitation also contributed to model discrimination to a lesser degree. Conclusions This framework demonstrates the feasibility of using machine learning to integrate echocardiographic, biomarker and clinical data to predict in-hospital mortality in acute PE, offering a more physiologically grounded alternative to traditional risk scores. Future work will focus on expanding model validation, comparing algorithmic performance and developing an interpretable predictive tool. A key limitation is that the model’s inputs are derived from structured echo reports and clinical interpretations rather than raw echocardiographic data. Further exploration from a causal standpoint, including individualized treatment effects of various modalities (i.e., systemic thrombolysis, mechanical thrombectomy) would lead to more informed decisions that provide a predicted impact on various treatment modalities. This abstract is funded by: None
Advano et al. (Fri,) conducted a observational in Acute pulmonary embolism (n=497). Machine learning-based framework (Random Forest classifier) was evaluated on In-hospital mortality (AUC 0.95). A machine learning framework integrating echocardiographic, biomarker, and clinical data predicted in-hospital mortality in acute pulmonary embolism with an AUC of 0.95.