The following is a summary of “Interpretable machine learning models for predicting venous thromboembolism in the intensive care unit: an analysis based on data from 207 centers,” published in the October 2023 issue of Critical Care by Guan et al.

Venous thromboembolism (VTE) is a severe complication in critically ill patients, causing death, disability, and healthcare burden. Researchers started a retrospective study to build an understandable machine learning (ML) model to predict VTE in critically ill patients using clinical features and lab tests.

They obtained data for this study from the eICU Collaborative Research Database (version 2.0). A stepwise logistic regression model was employed to select the predictors included in the model. The model was built using the random forest, extreme gradient boosting (XGBoost), and support vector machine algorithms with fivefold cross-validation. Model performance was analyzed using metrics such as area under the curve (AUC), accuracy, no information rate, balanced accuracy, kappa, sensitivity, specificity, precision, and F1 score. The DALEX package enhanced the interpretability of the final model.

The results showed 109,044 patients, with 1,647 (1.5%) experiencing VTE during their ICU stay. Among the 3 models, the Random Forest model demonstrated superior performance with an AUC of 0.9378, accuracy of 0.9958, kappa of 0.8371, precision of 0.9095, F1 score of 0.8393, sensitivity of 0.7791, and specificity of 0.9989.

Investigators concluded that the Random forest ML model is most effective at predicting VTE in critically ill patients, enabling early intervention.

Source: ccforum.biomedcentral.com/articles/10.1186/s13054-023-04683-4