Machine learning models for suspected pulmonary embolism in emergency department patients: a multicentre diagnostic study

Background 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. Objectives To evaluate ML models for ruling out PE without additional testing in ED patients with suspected PE, while maintaining a failure rate <2%. Methods We retrospectively analysed pooled data from four prospective European studies of ED patients with suspected PE (2000-2013). A total of 5038 patients were randomly split into training (n=2400) and validation (n=2638) datasets. ML models classified patients into low, intermediate, and high pre-test probability groups. In intermediate-risk patients, post-test probability was refined using either a fixed D-dimer threshold or a second ML model. The primary outcome was the failure rate assessed in the validation cohort and defined as the proportion of missed PE diagnoses. Results The best-performing algorithm combined weighted ridge regression (pre-test) with weighted random forest (post-test), achieving a failure rate of 1.44% (95% CI 1.04-1.99) and a Matthews correlation coefficient of 0.470 (95%CI 0.440-0.500). This model ruled out PE without any testing in 9.7% of patients, with an estimated imaging proportion of 54.1%, comparable to 4PEPS (52.9%) and lower than PERC (66.8%). Conclusion A two-step ML algorithm performed comparably to the best existing clinical decision rules, with a failure rate whose confidence interval remained below the prespecified 2% criterion and an estimated imaging proportion similar to 4PEPS. Prospective external validation is warranted.