Truace
Subscribe

Machine Learning Models for Suspected Pulmonary Embolism in Emergency Department Patients: A Multicentre Diagnostic Study

Pulmonary embolism (PE) is frequently suspected in emergency departments (EDs). Safely ruling out PE without additional testing could reduce imaging use and ED crowding. Machine learning (ML) may improve accuracy and efficiency.To evaluate ML models for ruling out PE without additional testing in ED patients with suspected PE, while maintaining a failure rate n = 2,400) and validation ( n = 2,638) datasets. ML models classified patients into low, intermediate, and high pre-test probability groups. In intermediat…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1739· cite & verify
Machine Learning Models for Suspected Pulmonary Embolism in Emergency Department Patients: A Multicentre Diagnostic Study
Hospital Universitari Doctor Peset, València 04 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0

In brief

Researchers built machine learning models to help emergency physicians decide which patients with suspected pulmonary embolism can be ruled out safely. The models sorted patients into low, intermediate and high pre-test probability groups. For intermediate-risk patients, a fixed D-dimer threshold or a second ML model refined the estimate. The models were assessed in a separate validation cohort of 2,638 patients.

The result matters because suspected pulmonary embolism is common in emergency departments, and avoiding unnecessary imaging could ease crowding. The ML approach performed about as well as existing rules rather than clearly better. The authors say prospective external validation is still needed before the approach can be trusted in practice.

Main points

  1. The best algorithm paired weighted ridge regression for pre-test probability with weighted random forest for post-test probability.
  2. Failure rate, defined as the proportion of missed PE diagnoses, was 1.44% with a 95% CI of 1.04-1.99.
  3. The Matthews correlation coefficient was 0.470 (95% CI: 0.440-0.500).
  4. Estimated imaging proportion was 54.1%, against 52.9% for 4PEPS and 66.8% for PERC.

The gain

A two-step ML algorithm ruled out pulmonary embolism with a failure rate below the 2% criterion and imaging use similar to the 4PEPS rule, and it did so without any testing for some patients.

The rundown

The study used development and validation datasets, with the validation set of 2,638 patients providing the primary failure-rate result. The prespecified safety target was a failure rate below 2%, and the upper confidence bound of the best model stayed under that value.

Comparison with clinical rules showed a trade-off between imaging and safety. The ML model's estimated imaging proportion was close to 4PEPS and well below PERC. It ruled out PE with no additional testing in 9.7% of patients.

Sources

  1. Peer-reviewedThrombosis and Haemostasis2026-10-08

The debate