Interpretable machine learning models for early prediction of all-cause mortality in patients with traumatic brain injury combined with venous thromboembolism: a retrospective study

To develop and validate interpretable machine learning (ML) models for predicting all-cause mortality in critically ill patients with traumatic brain injury combined with venous thromboembolism using data from the US-based Medical Information Mart for Intensive Care IV (MIMIC-IV) and eICU databases. Data from 321 patients with TBI and VTE were extracted from the MIMIC-IV database and split into training (70%) and validation (30%) sets. Fifty-five variables, including demographics, vital signs, laboratory results, and medical scores, were analyzed using the Boruta algorithm to identify predictors of in-hospital, 30-day, and 90-day mortality. Nine machine learning algorithms were evaluated, with the best-performing model validated using the eICU Collaborative Research Database (eICU-CRD) database. Model performance was assessed using receiver operating characteristic curve (AUROC), while accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, Kappa value etc. SHapley Additive exPlanations (SHAP) analysis was used to interpret the model. The Boruta algorithm identified 6, 5, and 7 predictors for in‑hospital, 30‑day, and 90‑day mortality, respectively. Among nine ML models, multilayer perceptron (MLP) performed favorable for in‑hospital mortality (AUROC 0.758, accuracy 0.919, specificity 0.948, PPV 0.500, F1 0.545, Kappa 0.50). Light gradient‑boosting machine (LightGBM) was optimal for 30‑day mortality (AUROC 0.794, accuracy 0.801, specificity 0.873, NPV 0.901). Logistic regression performed favorable for 90‑day mortality (AUROC 0.791, sensitivity 0.818, NPV 0.926). Top features: in‑hospital: age, diastolic blood pressure min, potassium max, blood urea nitrogen (BUN) max, sodium min; 30‑day: glucose min, potassium max, glucose max, BUN min, BUN max; 90‑day: age, sodium min, Glasgow Coma Scale (GCS) min, heart failure, gender. In external validation, MLP achieved AUROC 0.717 (0.621–0.813) for in‑hospital mortality, LightGBM 0.529 (0.410–0.649) for 30‑day, and logistic regression 0.691 (0.588–0.795) for 90‑day mortality. The results indicated that the MLP model performed well for predicting in-hospital mortality. The LightGBM model showed favorable performance in predicting 30-day mortality, while logistic regression presented good discriminatory ability for 90-day mortality. Additionally, SHAP analysis identified maximum potassium and maximum BUN as shared key indicators for in-hospital and 30-day mortality, whereas age and low serum sodium were critical factors associated with in-hospital and 90-day mortality risks. Future work should focus on validating the model in diverse populations and incorporating additional clinical data to enhance its applicability. Not applicable.