Prediction of radiologically confirmed hospital-acquired venous thromboembolism: machine learning versus Padua and IMPROVE, with external validation
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
The guideline-endorsed Padua and IMPROVE scores have only moderate discrimination for hospital-acquired venous thromboembolism (VTE). We developed and externally validated machine learning models for radiologically confirmed in-hospital VTE, benchmarked against electronic health record (EHR)-computed scores.
Approach and Results
We analyzed 379,720 adult MIMIC-IV admissions (radiology-report-covered 2008–2019), retaining prior VTE. The outcome was imaging-confirmed pulmonary embolism or deep-vein thrombosis diagnosed >24 hours after admission. Static and dynamic landmark models were compared with the scores using patient-grouped cross-validation, cluster-bootstrap inference and decision curve analysis; frozen models were externally validated in 29,144 high-risk admissions in Beijing, China. VTE occurred in 674 admissions (0.178%); ICD-based ascertainment of the same population yields 2.7%, roughly 15-fold higher. Because discharge-coded flags are unavailable at prediction time, the primary deployment-ready model used 42 admission-window features: out-of-fold AUC 0.7864 (ΔAUC +0.0795 versus Padua; 95% CI +0.054 to +0.104). The information-complete 57-feature models reached out-of-fold AUCs of 0.8090 (XGBoost) and 0.7893 (logistic regression) versus 0.7073 (Padua) and 0.6884 (IMPROVE); ΔAUC +0.1021 and +0.1211 for XGBoost versus Padua and IMPROVE (95% CIs excluding zero). Dynamic updating added +0.0953 AUC; a sequence model did not (ΔAUC −0.026). Externally, the admission-safe model transported better (AUC 0.706 vs. 0.678), exceeding both scores (95% CIs excluding zero).
Conclusion
Machine learning significantly improves on EHR-computed Padua and IMPROVE scores for in-hospital VTE prediction; the advantage persists on transport to a separate tertiary hospital’s high-risk admissions at lower absolute discrimination. Deployment requires admission-window features, local recalibration, and prospective validation.
