Explainable machine learning for early prediction and anatomical classification of pulmonary embolism in the emergency department.

backgroundPulmonary thromboembolism (PTE) is a life‑threatening condition that requires prompt and accurate evaluation in the emergency department (ED). Standardized clinical scoring systems, including the Wells and revised Geneva scores, form the cornerstone of initial risk stratification but have limited specificity, leading to unnecessary D‑dimer testing and frequent overuse of CT pulmonary angiography (CTPA). This study aimed to develop explainable machine‑learning (XML) models as a complementary decision‑support layer following initial clinical assessment, with the dual goals of improving PTE prediction and providing an early, non‑imaging‑based indication of clot location.

methodsClinical and paraclinical data from 472 ED patients with suspected PTE were collected across three centers of Mashhad University of Medical Sciences (2022-2024) using structured forms aligned with Wells/Geneva criteria. CTPA served as the reference standard for diagnosis and anatomical classification. We developed classical ML models, ensemble algorithms such as Extra Trees, and hybrid stacking pipelines. Explainable AI (XAI) using SHAP values quantified global and patientlevel feature contributions. The XML models were designed to operate after Wells/Geneva-based triage: supporting D-dimer decisions in low-risk patients and refining post-triage risk estimation in intermediate - and high-risk patients to help reduce avoidable CTPA utilization.

resultsIn this intermediate‑ to high‑probability cohort, all patients were classified as PTE‑suspect by Wells/Geneva scoring, whereas only 24% were confirmed positive on CTPA. ML models demonstrated improved discrimination, with Extra Trees achieving the best performance (accuracy 0.82, sensitivity 0.69, specificity 0.86, AUC 0.83). Among PTE‑positive cases, the model achieved an AUC of 0.77 for central emboli and 0.67 for peripheral emboli, with an overall anatomical classification accuracy of 63%.

conclusionsThe proposed XML models offer a transparent, clinically aligned framework that augments existing scoring systems, enhancing diagnostic efficiency and reducing unnecessary imaging.