Key result
LGBM model beats RCRI predicting 30-day mortality in non-cardiac surgical ICU patients, achieving ~0.98 AUROC.
Why the study?
Does the LGBM machine learning model improve the prediction of 30-day postoperative mortality compared to the traditional RCRI score in non-cardiac surgical patients admitted to the SICU?
Cohort (n=4,843)
No
Does the LGBM machine learning model improve the prediction of 30-day postoperative mortality compared to the traditional RCRI score in non-cardiac surgical patients admitted to the SICU?
Answers are grounded in this paper and cite their sources.
The LGBM machine learning model significantly outperforms the traditional RCRI score in predicting 30-day mortality among non-cardiac surgical patients in the ICU, offering a promising tool for early risk stratification.
Timely and accurate mortality risk stratification for patients admitted to surgical intensive care units (SICUs) following non-cardiac surgery facilitates rapid risk triage and rational allocation of medical resources. In this retrospective analysis, multiple machine learning (ML) and deep learning (DL) models were constructed to quantify 30-day postoperative mortality risk in this patient population. Clinical laboratory indicators and electrocardiogram (ECG) signals were extracted from non-cardiac surgical patients recorded in the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Five classical ML models were developed to predict 30-day mortality: logistic regression (LR), decision tree (DT), light gradient boosting machine (LGBM), extreme gradient boosting (XGBoost), and support vector machine (SVM). Two additional DL architectures, backpropagation neural network (BPNN) and recurrent neural network (RNN), were established for comparative performance testing. Shapley additive explanations (SHAP) analysis was applied to the optimal predictive model to identify dominant mortality-associated risk factors from both global and local interpretability perspectives. A total of 4 843 eligible patients were included, among whom 526 (10.9%) died within 30 days after non-cardiac surgery. The LGBM model outperformed all alternative machine learning and deep learning architectures, achieving the highest F1 score (0.934, 95% CI 0.913–0.942), and area under the receiver operating characteristic curve (AUROC, 0.977, 95% CI 0.972–0.982). Compared with the traditional Revised Cardiac Risk Index (RCRI), the LGBM framework delivered substantially improved classification performance. Global SHAP analysis identified preoperative international normalized ratio (INR), serum bicarbonate, blood urea nitrogen (BUN), and creatinine as the four most predictive variables for postoperative 30-day mortality. The LGBM model offers a novel and convenient method for preliminary prognostic risk stratification among non-cardiac surgical patients. This model can assist clinicians with early risk evaluation, though it cannot independently guide targeted clinical treatment decisions.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) conducted a cohort in Non-cardiac surgery (n=4,843). Light gradient boosting machine (LGBM) model vs. Revised Cardiac Risk Index (RCRI) was evaluated on 30-day postoperative mortality prediction (AUROC) (AUROC 0.977, 95% CI 0.972-0.982). The LGBM model accurately predicted 30-day mortality in non-cardiac surgical ICU patients, achieving an AUROC of 0.977 compared to 0.585 for the traditional Revised Cardiac Risk Index.
Synapse has enriched 10 related papers. Consider them for comparative context: