(1319) Machine Learning-Guided Early Prediction of Persistent Acute Kidney Injury After Cardiac Surgery

Introduction: Acute Kidney Injury (AKI) after cardiac surgery is common and associated with poor outcomes. When persistent for 48h or more (persistent AKI – pAKI), it is associated with more than double risk of death compared to AKI lasting less than 48 hours. It is thus important to determine strategies to predict pAKI to target preventative strategies. We aimed to develop a digital biomarker to predict development of pAKI on admission to intensive care unit (ICU) after cardiac surgery.

Methods: We included adult patients with coronary artery bypass grafting (CABG) and/or valve surgery from MIMIC-IV database. We identified AKI using both creatinine and urine output based KDIGO criteria. In accordance with ADQI guidelines we defined pAKI as AKI lasting at least 48h. We used features available on admission to ICU to develop 3 supervised machine learning models (XGBoost, Random Forest and Logistic Regression) to predict development of pAKI within the first week after cardiac surgery. Features included demographics, Charlson’s comorbidity score, type of surgery, admission SOFA score, available lab values, and use of mechanical ventilation on admission. We trained machine learning models on 80% cohort while using LASSO regression for feature selection and 10-fold cross-validation to tune hyperparameters. We used remaining 20% sample as test cohort.

Results: Among 6481 patients, 55.3% had CABG, 29.1% had valve surgery and 15.6% had both. Median age of the cohort was 68y, with 28.34% females and 75.5% whites. 24.6% of patients developed pAKI within the first week after cardiac surgery. In the test cohort, XGBoost model had an accuracy of 73.8% with an AUC-ROC of 0.71, 95% CI [0.68–0.75]. Random Forest Model showed an improved accuracy of 74.71%, and AUC-ROC of 0.71, 95% CI [0.68-0.74]. Logistic regression model performed the best with an accuracy of 75.10% with an AUC-ROC of 0.72, 95% CI [0.69–0.75].

Conclusions: Our study shows that machine learning can successfully predict the development of cardiac surgery-associated pAKI using data available on ICU admission after surgery. This will enable early identification of high-risk patients to optimize treatments and enrollment in clinical trials for prevention of pAKI.