Research on Risk Prediction of In-Hospital Death in Sepsis Patients Based on
Random Forest Algorithm
LI Lijuan1a, CAO Xiaojun1b, CHEN Feiyan1c, FAN Huifeng1d, LIU Guangjian2
1. a. Department of Hospital Quality Control; b. Information Center; c. Pediatric Intensive Care Unit; d. Department of Pneumology,
Guangzhou Women and Children’s Medical Center, Guangzhou Guangdong 510623, China;
2. Big Data Center, Shenzhen Dimai Bio-Technology Co., Ltd., Shenzhen Guangdong 518107, China
Abstract:Objective To build a new prediction model based on the balanced random forest (BRF) algorithm to address the issue
of low sensitivity caused by imbalanced sample categories in predicting in-hospital mortality of sepsis patients. Methods Data of
17 time-series variables for patients meeting the criteria of Sepsis-3.0 were obtained from publicly available Medical Information
Mart for Intensive Care-Ⅲ database. Then, the data from the initial 48 h after ICU admission was extracted, 714 statistical features
for the 17 variables were calculated, and they were used for model construction and performance evaluation. Two metrics, the area
under curve (AUC) and adjusted geometric-mean (AGM) were used for hyperparameter tuning. Besides BRF, the performance
was compared with the traditional logistic regression and random forest (RF) algorithm models. Results A total of 10270 sepsis
patients with ICU hospitalization experience were selected, and the overall in-hospital mortality rate was 18.04%. The performance
comparison results showed that the prediction sensitivity of machine learning models based on imbalanced sample classes
significantly improved, from the lowest of 0.1826 (95%CI: 0.1351-0.2322) in the RF-AGM model to the highest of 0.7110 (95%CI:
0.6537-0.7677) in the BRF-AGM model. With the use of the new predictive model, more patients facing mortality could be identified
and given timely treatment. The AUC, AGM and specificity of BRF-AGM model were 0.7994 (95%CI: 0.7696-0.8288), 0.7282
(95%CI: 0.7046-0.7519) and 0.7349 (95%CI: 0.7101-0.7590), respectively. Conclusion The BRF-AGM model has tremendous
potential for predicting mortality in ICU sepsis patients. It can help avoid treatment delays by clinicians, which is significant
importance in improving patient prognosis. However, the clinical utility of the BRF-AGM model still needs further evaluation
through prospective multicenter studies.
李丽娟1a,曹晓均1b,陈飞燕1c,樊慧峰1d,刘广建2. 基于随机森林算法的脓毒症患者
院内死亡风险预测研究[J]. 中国医疗设备, 2023, 38(12): 29-34.
LI Lijuan1a, CAO Xiaojun1b, CHEN Feiyan1c, FAN Huifeng1d, LIU Guangjian2. Research on Risk Prediction of In-Hospital Death in Sepsis Patients Based on
Random Forest Algorithm. China Medical Devices, 2023, 38(12): 29-34.