An Acute Kidney Injury Prediction Model Based on Ensemble Learning Algorithm

Acute Kidney Injury (AKI), a common disease in Intensive Care Unit (ICU) patients, is related to high cost, morbidity and mortality. The early prediction of AKI is critical for improving patients' outcome. However, sparse clinical data and highly imbalanced dataset bring great challenges to AKI prediction. Among existed machine learning algorithms, ensemble learning often stands out with its good performance on complex classification problems. Boosting algorithm is one of the best ensemble learning algorithms. Therefore, we develop a prediction model based on it aiming to forecast AKI ahead 24 hours and 48 hours. We also adopt the way that is analogous to text modeling transforming heterogeneous time series reflecting patients' medication information into multidimensional vector to overcome problem brought by sparse data. Additionally, since the imbalanced dataset would affect predictive performance, we artificially construct a more balanced dataset based on the original dataset to initialize the model. According to the experimental results, our model works well on the ICU patients dataset (AUC 24h ahead: 0.80 48h ahead 0.77). We also verify that medication information improves model performance (24h ahead: AUC 0.75 to 0.80 48h ahead: AUC 0.75 to 0.77) and find the optimal ratio of number between classes when initializing model for AKI prediction.