A Machine Learning Model for Acute Kidney Injury Prediction with Novel Kidney Biomarkers

Background: AKI is a serious complication characterized by poor short- and long-term outcomes in the intensive care unit. An increase in serum creatinine leads to impairment in renal function of the kidney and significantly increases the mortality rate. Early detection of AKI could lead to preventive interventions, therefore machine learning systems can detect AKI before its symptoms and consequences appear. Methods: We used a machine learning technique like boosted ensembles of decision trees along with novel kidney biomarkers to predict AKI using retrospective data extracted from the MIMIC-III dataset. Results: Our proposed algorithm achieved high AUROC for diagnosing and predicting AKI with novel kidney biomarkers. For AKI prediction at the point of onset, the algorithm achieved an AUROC of 0.890 (95 percent confidence interval [CI], 0.867-0.878). Also, our proposed algorithm obtained an AUROC of 0.82 (95 percent CI, 0.792-0.809) for predictions made 12-hours before onset and an AUROC of 0.85 (95 percent confidence interval: 0.785-0.804) for 24-hours predictions. Conclusion: The findings suggest that a machine learning-based AKI prediction tool with biomarkers could provide key prognostic skills for predicting which patients are at risk of AKI, eventually enabling physicians to intercede before kidney injuries occur.