Prediction of Cisplatin‐Induced Acute Kidney Injury Using an Interpretable Machine Learning Model and Electronic Medical Record Information

摘要

Predicting cisplatin‐induced acute kidney injury (Cis‐AKI) before its onset is important. We aimed to develop a predictive model for Cis‐AKI using patient clinical information based on an interpretable machine learning algorithm. This single‐center retrospective study included hospitalized patients aged ≥ 18 years who received the first course of cisplatin chemotherapy from January 1, 2011, to December 31, 2020, at Nagoya City University Hospital. Cis‐AKI‐positive patients were defined using the serum creatinine criteria of the Kidney Disease Improving Global Outcomes guideline within 14 days of the last day of cisplatin administration in the first course. Patients who received cisplatin but did not develop AKI were considered negative. The CatBoost classification model was constructed with 29 explanatory variables, including laboratory values, concomitant medications, medical history, and cisplatin administration information. In total, 1253 patients were included, of whom 119 developed Cis‐AKI (9.5%). The median time of AKI onset was 7 days, and the interquartile range was 5–8 days. The mean ± standard deviation of the total cisplatin dose in the initial treatment was 77.9 ± 27.1 mg/m2 in Cis‐AKI‐positive patients and 69.3 ± 22.6 mg/m2 in Cis‐AKI‐negative patients. The predictive performance was an ROC‐AUC of 0.78. Model interpretation using SHapley Additive exPlanations showed that concomitant use of intravenous magnesium preparations was negatively correlated with Cis‐AKI, whereas loop diuretics were positively correlated. This suggests the need for magnesium preparations to prevent AKI, although the effects of diuretics may be small. Our model can predict Cis‐AKI early and may be helpful for its avoidance.