Derivation and validation of a risk score to predict acute kidney injury in critically ill cirrhotic patients

• Published in: Hepatology research Vol. 53; no. 8; pp. 701 - 712
• Main Authors: Zheng, Luyan, Lin, Yushi, Fang, Kailu, Wu, Jie, Zheng, Min
• Format: Journal Article
• Language: English
• Published: Netherlands Wiley Subscription Services, Inc 01.08.2023
• Subjects: acute kidney injury; Bilirubin; Creatinine; decompensated cirrhosis; Hemoglobin; Kidneys; Lymphocytes; machine learning; Magnesium; Neural networks; Oxygen saturation; Patients; Prothrombin; scoring system; acute kidney injury; model; decompensated cirrhosis; machine learning; scoring system
• ISSN: 1386-6346; 1872-034X
• DOI: 10.1111/hepr.13907

Aim Acute kidney injury (AKI) is a common complication in critically ill cirrhotic patients with substantial mortality. Given AKI can be prevented through early detection, it is urgent to develop an easy model to identify high‐risk patients. Methods A total of 1149 decompensated cirrhotic (DC) patients from the eICU Collaborative Research Database were enrolled for model development and internal validation. The variables used for analysis mainly included laboratory tests. We first built an ensemble model (random forest, gradient boosting machine, K‐nearest neighbor, and artificial neural network) named DC‐AKI using machine learning methods. Based on the Akaike information criterion, we then constructed a risk score, which was externally validated in 789 DC patients from the Medical Information Mart for Intensive Care database. Results AKI developed in 212 (26%) of 804 patients in the derivation cohort, and 355 (45%) of 789 patients in the external validation cohort. DC‐AKI identified the eight variables most strongly associated with the outcome: serum creatinine, total bilirubin, magnesium, shock index, prothrombin time, mean corpuscular hemoglobin, lymphocytes, and arterial oxygen saturation. Based on the smallest Akaike information criterion, a six‐variable model was eventually used to construct the scoring system (serum creatinine, total bilirubin, magnesium, shock index, lymphocytes, and arterial oxygen saturation). The scoring system showed good discrimination, with the area under the receiver operating characteristics curve of 0.805 and 0.772 in two validation cohorts. Conclusions The scoring system using routine laboratory data was able to predict the development of AKI in critically ill cirrhotic patients. The utility of this score in clinical care requires further research. We applied the machine learning method to create a simple risk score for ICU patients with cirrhosis from two cohorts. The risk score only included six commonly used laboratory indicators and performed well. Our risk score might be used to determine the risks of AKI for cirrhotic patients during hospitalization on the first day of hospitalization, which in turn will optimize treatment to improve prognosis.