Construction of an early prediction model for post cardiopulmonary resuscitation-acute kidney injury based on machine learning
10.3760/cma.j.cn441217-20240301-00303
- VernacularTitle:基于机器学习构建心肺复苏后急性肾损伤早期预测模型
- Author:
Jinxiang WANG
1
;
Luogang HUA
;
Daming LI
;
Hongbao GUO
;
Heng JIN
;
Guowu XU
Author Information
1. 天津医科大学总医院急诊科，天津 300052
- Keywords:
Acute kidney injury;
Cardiopulmonary resuscitation;
Machine learning;
Heart arrest;
Prediction model
- From:
Chinese Journal of Nephrology
2024;40(11):875-881
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct an early prediction model for post cardiopulmonary resuscitation-acute kidney injury (PCPR-AKI) by machine learning and provide a basis for early identification of acute kidney injury (AKI) high-risk patients and accurate treatment.Methods:It was a single-center retrospective study. The clinical data of patients admitted to Tianjin Medical University General Hospital after cardiopulmonary resuscitation following cardiac arrest from January 1, 2016 to October 31, 2023 were collected. The end-point event of the study was defined as AKI occurring within 48 hours after cardiopulmonary resuscitation. The patients were divided into AKI group and non-AKI group according to the AKI diagnostic criteria, and the differences of baseline clinical data between the two groups were compared. The patients who met the inclusion criteria were randomly (using the train_test_split function, set the random seeds to 1, 2, and 3) divided into training and validation sets at a ratio of 7∶3. Random forest (RF), support vector machine, decision tree, extreme gradient boosting and light gradient boosting machine algorithm were used to develop the early prediction model of PCPR-AKI. The receiver-operating characteristic curve and decision curve analysis were used to evaluate the performance and clinical practicality of the predictive models, and the importance of variables in the optimal model was screened and ranked.Results:A total of 547 patients were enrolled, with age of 66 (59, 70) years old and 282 males (51.6%). There were 238 patients (43.5%) having incidence of AKI within 48 hours after cardiopulmonary resuscitation. In the AKI group, 182 patients (76.5%) were in stage 1, 47 patients (19.7%) were in stage 2, and 9 patients (3.8%) were in stage 3. There were statistically significant differences in the age, time to reach resuscitation of spontaneous circulation, time from cardiac arrest to starting cardiopulmonary resuscitation, proportion of initial defibrillation rhythm, proportion of electric defibrillation, proportion of mechanical ventilation, adrenaline dosage, sodium bicarbonate dosage, proportion of coronary heart disease, proportion of hypertension, proportion of diabetes, serum creatinine, blood urea nitrogen, blood lactic acid, blood potassium, brain natriuretic peptide, troponin, D-dimer, neuron specific enolase, and 24 hours urine volume after cardiopulmonary resuscitation between AKI group and non-AKI group (all P<0.05). Among the five machine learning algorithms, RF model achieved the best performance and clinical practicality, with area under the curve of 0.875, sensitivity of 0.863, specificity of 0.956, and accuracy rate of 90.7%. In the variable importance ranking of RF model, the top 10 variables were as follows: time to reach resuscitation of spontaneous circulation, time from cardiac arrest to starting cardiopulmonary resuscitation, initial defibrillable rhythm, serum creatinine, mechanical ventilation, blood lactate acid, adrenaline dosage, brain natriuretic peptide, D-dimer and age. Conclusions:An early predictive model for PCPR-AKI is successfully constructed based on machine learning. RF model has the best predictive performance. According to the importance of the variables, it can provide clinical strategies for early identification and precise intervention for PCPR-AKI.
