Machine learning improves prediction of pulmonary thromboembolism and reduces unnecessary computed tomography scans in the emergency department

Comprehensive Summary

Yoon et al. investigated whether machine learning (ML) models could more accurately predict pulmonary thromboembolism (PTE) in emergency department (ED) patients than the revised Geneva score. In this single-center study, data from 2,525 patients with suspected PTE who underwent computed tomography pulmonary angiography (CTPA) were analyzed and divided into training and test sets. Six ML models were then compared against the revised Geneva score using the area under the receiver operating characteristic curve (AUC). Permutation was applied to assess the importance of individual variables. The study found that the XGBoost achieved the highest AUC of 0.814 (95% confidence interval [CI]: 0.759–0.862), while all models significantly outperformed the revised Geneva score (AUC: 0.622, 95% CI: 0.563–0.675). The XGBoost model could reduce unnecessary CTPA imaging by up to 33.2 % at 90% sensitivity (p<0.001). In addition, D-dimer and activated partial thromboplastin time (aPTT) were identified as the most critical predictors across all models.

Outcomes and Implications

The study concludes that ML models are an effective tool in predicting PTE risk compared to the standard revised Geneva score, which is important given the condition’s unspecific clinical presentation. Integrating these models into hospital electronic records could be used to flag low-risk patients and reduce unnecessary imaging, thereby minimizing exposure to radiation and healthcare costs. Due to the study’s retrospective, single-center design, further prospective and external validation is warranted to confirm the model’s safety and generalizability across different clinical settings.