Machine learning-extended sPESI for 1-year mortality prediction in pulmonary embolism
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Introduction: The simplified pulmonary embolism severity index (sPESI) is extensively used for bedside risk assessment in acute pulmonary embolism (PE). Nevertheless, its effectiveness is constrained by the dichotomization of physiological variables and a primary validation focus on 30-day outcomes only. We developed a machine learning (ML) extension to enhance prognostic precision and extend risk stratification to 12 months, using only bedside data. Materials and Methods: In a retrospective cohort of 2547 adults with computed tomography pulmonary angiography confirmed PE, we developed logistic regression (LR), XGBoost, and multi-layer perceptron (MLP) models using a 10-item hybrid extension of sPESI that retained continuous age, heart rate, systolic blood pressure, and oxygen saturation alongside the six binary sPESI indicators. Models were trained and optimized using 12-month all-cause mortality as the target and were evaluated against 30, 180, and 365-day mortality on a held-out test set. Result: Discrimination with the conventional sPESI attenuated over time (AUC= 0.727, 0.689, and 0.665 at 30 days, 180 days, and one year, respectively). The 10-item ML extension significantly enhanced prognostic performance compared with the conventional 6-item sPESI across all horizons (all p< 0.05). At the primary 12-month horizon, LR, XGBoost, and MLP achieved AUCs of 0.720, 0.726, and 0.712, respectively (vs. 0.665 for sPESI). At 30 days, the peak discrimination reached 0.796 (XGBoost). Moreover, ML increased 30-day specificity to 0.326-0.395 compared to 0.131 for sPESI, while maintaining high sensitivity (0.930-0.982). SHapley Additive exPlanations analyses demonstrated horizon-linked attribution shifts, with acute hemodynamic markers dominating early mortality predictions and age and malignancy emerging as the primary drivers at 12 months under horizonspecific evaluation. Conclusions: Preserving continuous physiological resolution within a familiar clinical framework enables ML to capture both acute instability and latent long-term vulnerability from the index presentation data alone. This bedsidefeasible pathway enhances long-horizon risk stratification and individualizes the management of patients.
Fields of Study & Subjects
Venous Thromboembolism Diagnosis and Management
Artificial Intelligence in Healthcare
Machine Learning in Healthcare
Publication details
- Journal / source
- Tuberkuloz ve Toraks
- Volume
- 74
- Issue
- 1
- Pages
- 43–57
- Publication date
- 25 March 2026
- Language
- EN