Does an explainable AI-based prediction model accurately identify patients at high risk of developing de novo chronic kidney disease after cardiac surgery?
An explainable AI model utilizing baseline eGFR, perioperative creatinine increase, age, and sex can accurately predict the development of de novo chronic kidney disease after cardiac surgery.
Objective Chronic kidney disease (CKD) is a serious long-term complication after cardiac surgery associated with increased morbidity and mortality. As CKD often remains undiagnosed for extended periods, we aimed to develop and externally evaluate an explainable artificial intelligence (XAI)-based model to identify patients at high risk of CKD following cardiac surgery. Methods and analysis For model development, we extracted over 200 clinical variables from cardiac surgery patients at Odense University Hospital, Denmark (2000–2022) from the Western Denmark Heart Registry and merged them with biochemical data from regional laboratory systems. Patients with preoperative kidney dysfunction or missing data required for CKD determination were excluded. The dataset was balanced by age, sex, surgical decade and CKD occurrence and split into training, validation and test samples. We employed an XAI algorithm (QLattice) using symbolic regression to generate prediction models. External evaluation was conducted with cardiac surgery patients from Aarhus University Hospital, Denmark (2008–2024). Model performance was assessed using receiver operating characteristic curves with area under the curve (AUC) and calibration plots. Results Data from 11 156 patients were used for model development. Among these patients, the unadjusted frequency of de novo CKD was 13% at 3 years and 18% at 5 years post-surgery, with 47% of all CKD cases developing within 3 years of discharge. Baseline estimated glomerular filtration rate, perioperative creatinine increase, age and sex were identified as key predictors of CKD development. The model achieved an AUC of 0.86 and demonstrated good mean and moderate calibration. External evaluation on 9479 patients yielded an AUC of 0.88 with comparable calibration after intercept recalibration. Conclusion We developed, evaluated, and updated an XAI-based model able to identify patients at high risk of CKD after cardiac surgery. The model is ready for clinical implementation, enabling improved interdisciplinary follow-up of kidney function after cardiac surgery.
Lindhardt et al. (Thu,) studied this question.