An XGBoost machine learning model for early prediction of DVT in ICU patients with COPD achieved an AUC of 0.840 in internal validation and demonstrated stable discrimination in two external cohorts, with SHAP analysis identifying prolonged PTT, elevated RDW, reduced SpO2, and increased respiratory rate as important contributors to DVT risk.
DVT occurred in 6.9% of critically ill COPD patients in the study cohort.
XGBoost was the best-performing model among eight machine learning algorithms evaluated.
External validation confirmed stable discrimination of the XGBoost model in two independent cohorts.
SHAP analysis identified prolonged PTT, elevated RDW, reduced SpO2, and increased respiratory rate as the most important predictors of DVT risk.
Decision-curve analysis suggested the XGBoost model provides potential clinical benefit across relevant risk thresholds.
A web-based calculator was developed to facilitate clinical application of the model.
Existing risk assessment tools for DVT are not specifically tailored to critically ill COPD patients, motivating development of a COPD-specific model.
This research suggests that a type of machine learning model called XGBoost can accurately predict which critically ill patients with chronic obstructive pulmonary disease (COPD) are at high risk of developing deep vein thrombosis (DVT) — dangerous blood clots that form in the deep veins, usually in the legs. The researchers developed the model using data from nearly 6,700 ICU patients, of whom about 1 in 14 (6.9%) developed DVT. The model performed well not only in the original dataset but also when tested on two separate patient databases, suggesting it is robust and may generalize across different hospital settings. A key feature of this study is that the model is 'explainable' — meaning clinicians can see which factors are driving the risk prediction for each individual patient. The analysis found that four factors were particularly important: a longer clotting time (PTT), a higher variability in red blood cell size (RDW), lower blood oxygen levels (SpO2), and a faster breathing rate. These are all measurements routinely collected in the ICU, meaning the model could be applied without requiring additional testing. The researchers also built a web-based tool to make the model easy to use in clinical practice. This research suggests that applying this kind of interpretable machine learning tool could help doctors identify COPD patients in the ICU who are most likely to develop DVT earlier than current standard tools allow, potentially enabling more targeted use of blood-thinning medications and imaging tests. Because existing DVT risk scores were not designed specifically for this population, a COPD-tailored model may offer a meaningful improvement in how care is prioritized for these high-risk patients.
Wang G, Liu T, Ji W, Li T, Wang Z, Hu T, et al.. (2026). Explainable Machine Learning for Predicting Deep Vein Thrombosis in Critically Ill Patients with COPD: Development and Multicenter External Validation.. International journal of chronic obstructive pulmonary disease. https://doi.org/10.2147/COPD.S609203