Construction and validation of short term mortality risk prediction model for patients with cancer in the intensive care unit

KONG Jun; XIA Biyun; LIU Ting; DAI Li; LI Li

doi:10.16048/j.issn.2095-5561.2026.02.04

Abstract：Objective To develop a 7-day mortality risk prediction model for patients with cancers in the intensive care unit(ICU)based on logistic regression(LR),support vector machine(SVM),random forest(RF),gradient boosting decision tree(GBDT),extreme gradient boosting(XGBoost)and light gradient boosting machine(LightGBM).Methods A total of 4 102 cancer patients in the ICU were selected from the Medical Information Mart for Intensive Care Ⅳ database,including 2 412 males and 1 690 females,with the age of(68.4±12.8)years old,ranging from 20 to 100 years old.The cohort was randomly divided into a training set(n=3 281)and an internal validation set(n=821)at an 8∶2 ratio.In addition,932 cancer patients in the ICU of the Third Affiliated Hospital of Naval Medical University between December 2016 and December 2024 were enrolled as the external validation set,consisting of 699 males and 233 females,with the age of(61.8±11.4)years old,ranging from 19 to 90 years old.Clinical baseline characteristics,vital signs,laboratory parameters and supportive treatment data of the patients were collected and preprocessed.7-day mortality risk prediction models for ICU cancer patients were established based on LR,SVM,RF,GBDT,XGBoost and LightGBM algorithms,and the model performance was evaluated by area under the receiver operating characteristic curve.Results The LightGBM model exhibited the optimal performance,with the area under curve of 0.730(95%CI:0.679 to 0.785)in the internal validation set and 0.806(95%CI:0.736 to 0.866)in the external validation set.Conclusion The LightGBM model demonstrates satisfactory predictive ability for the 7-day mortality risk of cancer patients in the ICU.