Machine learning-based prediction of circuit clotting during pediatric continuous kidney replacement therapy sessions

Background: Continuous kidney replacement therapy (CKRT) is commonly used for managing acute kidney injury (AKI) in critically ill pediatric patients. However, unexpected circuit clotting remains a frequent complication, resulting in therapy interruptions, blood loss, and increased clinical workload. Timely prediction of clotting could enhance circuit management and patient outcomes. Methods: We retrospectively analyzed de-identified data from 23 pediatric patients undergoing 101 CKRT sessions at a tertiary PICU between 2012 and 2017. Time-series data were collected from CKRT machines and patient records, including demographic, clinical, and treatment-related variables. A machine learning (ML) classification model was developed to predict clotting events 60 min before occurrence. The dataset was preprocessed and split into training (70%) and validation (30%) sets, preserving class balance. Feature selection was performed using LightGBM, and model performance was evaluated using the Extra Trees classifier with cross-validation. Results: Of 101 CKRT sessions, 59 ended due to clotting. After data cleaning and exclusion of sessions shorter than 60 min, 88 sessions and over 218,000 data points were analyzed. The final model achieved an AUROC of 0.99 in the training set and performed well in validation, predicting clotting events 60 min in advance in 148 instances. The most important predictive features included effluent volume, treatment duration, fluid removal, and dialysate flow. Conclusion: This study demonstrates that ML can effectively predict circuit clotting during pediatric CKRT, offering a valuable tool for early clinical intervention. Integration of predictive algorithms into clinical workflows may help reduce complications, optimize treatment continuity, and improve outcomes for critically ill children.