#552 A machine learning approach to predict acute kidney injury in hospitalised heart failure patients

Abstract Background and Aims Acute Kidney Injury (AKI) is a critical complication of heart failure. Manifesting within the complex frameworks of cardiorenal syndrome, AKI often progresses to Chronic Kidney Disease (CKD) and other severe diseases, thus warranting early detection and intervention. In this study, AKI is defined by an increase in serum Creatinine (SCr) by more than 0.3 mg/dL within 48 hours or by 1.5 times baseline SCr during hospitalisation according to the kidney disease: improving global outcomes (KDIGO) criteria. This study sought to develop a predictive model to predict whether heart failure patients would develop acute kidney injury within a hospitalised setting. Method Demographic data, past medical history, drug history, laboratory test results and vital signs were recorded for 221 heart failure patients across 498 hospitalisations at the University Malaya Medical Centre (UMMC) in Malaysia from January 2020 to June 2024 . Characteristics between the AKI and non-AKI group were compared using the Student's t test; non-normally distributed data were analysed using Mann-Whitney U test; categorical data were analysed using Chi-squared test. Logistic Regression, Extreme Gradient Boosting (XGBoost), Random Forest and K-Nearest Neighbour (KNN) models were built to predict the development into AKI. Established machine learning (ML) models were assessed based on their Area Under the Receiver Operating Characteristic (AUROC) Curve, precision, recall and F1 score. Results 27.3% of the hospitalized patients developed AKI during hospitalization. Hypertension, CKD, length of stay in hospital, systolic and diastolic blood pressure, mean arterial pressure, respiratory rate, hypotensive episodes during hospitalization, cumulative dose of furosemide during hospitalisation, haemoglobin level, white blood cell count, albumin level and C-reactive protein level were identified as significant predictors of AKI, thus were included within the ML models. Random Forest achieved the best performance with an accuracy of 74.3%, AUROC of 76.5% and an F1 score of 31.6%, balancing precision (54.5%) with recall (22.2%). XGBoost showed comparable AUROC (75%), but lower recall (11.1%). KNN demonstrated moderate performance, however with a higher recall (25.9%) and F1 score (34.1%). Logistic Regression underperformed across all metrics. Conclusion The Random Forest Model performed the best in predicting AKI in hospitalised heart failure patients. However, a low F1 score across all 4 models suggests the ineffectiveness of detecting positive AKI cases, possibly due to the positive AKI class being underrepresented in this study. Further studies should focus on addressing class imbalance through various techniques to improve the detection of AKI.