Dynamic machine learning prediction of persistent AKI after cardiac surgery with modifiable perioperative risk factors

Abstract

Early identification and prevention of persistent acute kidney injury (pAKI) remain challenging due to delayed biochemical markers and limited tools to differentiate between transient and persistent forms. We retrospectively analyzed data of 2,285 patients who underwent cardiac surgery with cardiopulmonary bypass (CPB) and developed 3 machine learning (ML) models for predicting pAKI: model 1 (preoperative data); model 2 (intraoperative and immediate postoperative variables); and model 3 (data up to 48 h post-ICU admission). pAKI occurred in 168 patients. Predictive performance improved across models, reflecting the value of time-updated data. SHapley Additive exPlanations highlighted baseline factors (estimated glomerular filtration rate and hemoglobin) in model 1 and perioperative factors associated with pAKI risk (post-CPB perfusion pressure, transfusion volume, and hemoglobin trends) in models 2 and 3 as dominant contributors. Our dynamic ML model enables early risk stratification and identification of perioperative factors associated with pAKI risk, providing a foundation for hypothesis generation and future investigation.