Development and validation of an explainable machine learning model for mortality prediction in ICU patients with...

Lung cancer is the leading cause of cancer-related death worldwide, with a particularly high mortality risk among patients admitted to the intensive care unit (ICU). Accurate and timely prediction of in-hospital mortality in this population is critical for clinical decision-making and resource allocation. However, existing severity scoring systems have demonstrated limited discriminative performance in this specific context. This study aimed to develop and validate an interpretable machine learning (ML) model for in-hospital mortality prediction in ICU patients with lung cancer using real-world clinical data from the MIMIC-IV(Medical Information Mart for Intensive Care IV) database.

Retrospective cohort data were extracted from the MIMIC-IV database. A total of 1,120 ICU admissions of lung cancer patients were included (239 deaths, 21.3%). Candidate variables were systematically screened through univariate analysis, variance inflation factor (VIF) assessment, and bidirectional stepwise logistic regression-all performed exclusively on the training set (70%, n = 785) to prevent data leakage. Eight machine learning algorithms were developed: Logistic Regression (LR), Random Forest (RF), Neural Network (NN), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Adaptive Boosting (AdaBoost), and Decision Tree (DT). Model performance was evaluated on an independent test set (30%, n = 335) using six metrics-Brier score, calibration slope, F1 score, Youden index, area under the receiver operating characteristic curve (AUC), and accuracy (ACC)-with a composite scoring system to rank models. Interpretability was assessed using SHapley Additive exPlanations (SHAP). Clinical utility was evaluated through decision curve analysis (DCA) and comparison with traditional severity scores (Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), and Oxford Acute Severity of Illness Score (OASIS)). Subgroup and robustness analyses were also conducted.

Five variables were selected as final predictors: SOFA score, SAPS II score, OASIS score, Charlson Comorbidity Index (CCI), and minimum peripheral oxygen saturation (SpO₂). The Logistic Regression model achieved the highest comprehensive score (42/48), with an AUC of 0.791 (95% CI: 0.731-0.850), Brier score of 0.137, F1 score of 0.550, Youden index of 0.482, accuracy of 0.746, and calibration slope of 0.801. The LR model significantly outperformed SOFA (AUC: 0.749, P = 0.034), SAPS II (AUC: 0.719, P = 0.011), and OASIS (AUC: 0.678, P < 0.001). SHAP analysis identified SOFA score as the dominant predictor, with SpO₂contributing a negative influence. The model demonstrated stable and consistent performance across subgroups (mechanical ventilation, age, sex) and across 50 repeated random splits. DCA confirmed the model's net clinical benefit at threshold probabilities of 0.10-0.70.

We developed an explainable, LR-based mortality prediction model for ICU lung cancer patients, which outperforms traditional severity scoring systems and provides individual-level clinical interpretability through SHAP values. This model offers a practical and transparent tool to support critical care decision-making in this high-risk population.