Acute kidney injury (AKI) affects up to 50% of cardiac surgery patients. The definition of AKI is based on changes in serum creatinine relative to a baseline measurement or a decrease in urine output. These monitoring methods lead to a delayed diagnosis. Monitoring the partial pressure of oxygen in urine (PuO2) may provide a method to assess the patient's AKI risk status dynamically. This study aimed to assess the predictive capability of two machine learning algorithms for AKI in cardiac surgery patients. One algorithm incorporated a feature derived from PuO2 monitoring, while the other algorithm solely relied on preoperative risk factors. The hypothesis was that the model incorporating PuO2 information would exhibit a higher area under the receiver operator characteristic curve (AUROC). An automated forward variable selection method was used to identify the best preoperative features. The AUROC for individual features derived from the PuO2 monitor was used to pick the single best PuO2-based feature. The AUROC for the preoperative plus PuO2 model vs. the preoperative-only model was 0.78 vs. 0.66 (p-value < 0.01). In summary, a model that includes an intraoperative PuO2 feature better predicts AKI than one that only includes preoperative patient data.
Keywords: acute kidney injury; cardiac surgery; machine learning; urinary oxygenation.