ScholarWorks@성균관대학교

초록

Background Early identification of patients with septic shock who may soon require renal replacement therapy (RRT) is clinically important but challenging in the emergency department (ED), where definitive indications for RRT often have not yet developed at the time of presentation. Recognizing these patients in advance is important for timely planning of RRT initiation, including coordination of equipment and personnel at the hospital level. This study aimed to develop and validate machine learning (ML) models that predict the need for RRT within 24 h of septic shock recognition in the ED. Methods We analyzed data from the Korean Shock Society septic shock registry collected from October 2015 to December 2023. Feature selection was performed using least absolute shrinkage and selection operator regression, and five ML models were trained. The best-performing model was selected based on the area under the receiver operating characteristic curve (AUROC). Shapley additive explanations were used to interpret the contribution of each feature. Results In total, 5361 patients were included in the analysis, of whom 728 (13.6%) required RRT within 24 h. Among the evaluated models, categorical boosting (CatBoost) demonstrated the best discrimination with an AUROC of 0.86 (95% CI, 0.833-0.887), outperforming conventional severity scores such as the Sequential Organ Failure Assessment (AUROC, 0.673 [95% CI, 0.628-0.717]) and the Acute Physiology and Chronic Health Evaluation (AUROC, 0.672 [95% CI, 0.623-0.719]). Conclusions The CatBoost model demonstrated moderate discriminative performance for predicting early RRT requirement within 24 h of ED septic shock recognition.

키워드