Uplift modeling to predict individual treatment effects of renal replacement therapy in sepsis-associated acute kidney injury patients

Background: Renal replacement therapy (RRT) is a crucial treatment for sepsis-associated acute kidney injury (S-AKI), but it is uncertain which S-AKI patients should receive immediate RRT. Identifying the characteristics of patients who may benefit the most from RRT is an important task. Methods: : This retrospective study utilized a public database and enrolled S-AKI patients, who were divided into RRT and non-RRT groups. Uplift modeling was used to estimate the individual treatment effect (ITE) of RRT. The validity of different models was compared using a qini curve. After labeling the patients in the validation cohort, we characterized the patients who would benefit the most from RRT and created a nomograph. Result: A total of 8,878 patients were assessed, among whom 687 received RRT, and 8,191 did not receive RRT. The RRT group had a higher severity of illness than the non-group, with a Sequential Organ Failure Assessment (SOFA) score of 9 (IQR 6,12) vs. 5 (IQR 3,7). The 28-day mortality rate was higher in the RRT group than the non-RRT group (34.35% vs. 14.67%, p < 0.0001). Propensity score matching (PSM) was used to balance baseline characteristics, 687 RRT patients and an equal number of non-RRT patients were enrolled for further research. After PSM, there was no significant difference in 28-day mortality between the RRT and non-RRT groups (34.35% vs. 36.83%, P = 0.338). Using uplift modeling, we found that urine output, fluid input, SpO2, BUN, and platelet were the top 5 factors that had the most influence on RRT effect. The area under the uplift curve (AUUC) of the class transformation model was 0.064, the AUUC of SOFA was 0.031, and the AUUC of Kdigo-stage was 0.048. The class transformation model was more efficient in predicting individual treatment effect. A logical regression model was developed, and a nomogram was drawn to predict whether an S-AKI patient can benefit from RRT. Nine factors were taken into account (ventilation, urine output, fluid input, temperature, creatinine, chlorine, pH, white blood cell count, and first care unit). Conclusion: Uplift modeling can better predict the ITE of RRT on S-AKI patients than conventional score systems such as Kdigo and SOFA. We also found that some inflammation indicators such as temperature and white blood cell count may influence the benefit of RRT on S-AKI patients.