Development and Nested Cross-Validation of Explainable Machine-Learning Models for Predicting the Risk of Propofol-Associated Hypertriglyceridemia in Critically-Ill, Mechanically-Ventilated Patients

Abstract RATIONALE Propofol is the first-line sedative for critically-ill, mechanically-ventilated patients, but it can increase serum triglyceride levels. Machine-learning (ML) offers a promising approach for individualized risk stratification of propofol-associated hypertriglyceridemia, enabling proactive triglyceride monitoring and sedation optimization. This study aims to develop and validate ML models to predict hypertriglyceridemia risks in critically-ill patients receiving propofol. METHODS We analyzed retrospective data from adult ICU patients on mechanical ventilation receiving continuous propofol infusion for ≥24 hours across four Mayo Clinic sites from May 2018 to June 2023. The primary outcome was hypertriglyceridemia, defined as serum triglycerides >400 mg/dL within 10 days of starting propofol. Ten-fold cross-validation and leave-one-site-out nested cross-validation (LOSO-CV) were used for model tuning and external validation of our modeling process, respectively. Feature selection was conducted through a union of Boruta and least absolute shrinkage and selection operator (LASSO). While COVID status was a strong predictor of hypertriglyceridemia, we concurrently developed non-COVID models to ensure future relevance in non-pandemic settings. Six ML algorithms (logistic regression, support vector machines, random forests, LightGBM, XGBoost, and multilayer perceptrons) were tuned using Bayesian optimization to minimize cross-entropy loss. Classification thresholds were set to target sensitivity ≥80%. Model performance was evaluated using AUC-ROC, Brier score, sensitivity, and specificity. Models with the highest Youden's Index were designated the best-performing model. Model explainability was assessed using permutation importance and SHapley Additive exPlanations (SHAP). RESULTS We included 3922 eligible patients with triglyceride measurements. Of these, 3042 patients were COVID-negative and used for non-COVID modeling. The best-performing non-COVID models achieved an average AUC-ROC of 0.69 (95% confidence interval [CI] 0.68-0.70), Brier score of 0.12 (95% CI 0.09-0.14), sensitivity of 0.79 (95% CI 0.75-0.83), and specificity of 0.48 (95% CI 0.44-0.52) during LOSO-CV. The best-performing models that included COVID status achieved an average AUC-ROC of 0.71 (95% CI 0.70-0.72), Brier score of 0.13 (95% CI 0.10-0.16), sensitivity of 0.78 (95% CI 0.71-0.84), and specificity of 0.51 (95% CI 0.45 to 0.56) during LOSO-CV. The permutation importance results of the final non-COVID model are shown in the figure. CONCLUSION We have developed sensitive ML models for predicting hypertriglyceridemia risks in critically-ill patients receiving propofol. Although developed during the COVID-19 pandemic, the non-COVID models are more likely to have future relevance, and they have similar performance as models including COVID status. These models will be integrated into a web-based clinical calculator and undergo further validation in prospective studies.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.