Predicting persistent acute kidney injury in ICU patients

Acute Kidney Injury (AKI) is a common disease consisting in the loss of functionality of the kidneys. Depending on the length of the episode, it can be classified as transient or persistent. Failures lasting longer than 48-72 hours were proved to be related to higher short- and long-term mortality and morbidities. Whereas the recent literature focused on the usage of biomarker tests to predict persistent AKI, the innovation of the proposed approach consists of processing time-series measurements of clinical parameters of the patients and using machine learning algorithms to predict the onset of a persistent renal injury. The patients satisfying the inclusion criteria belonging to the eICU-CRD database were split into training, validation, and test set. Patients from MIMIC-III database were used as an external test set. Four different machine learning algorithms were evaluated in both test sets: the Random Forest Classifier proved to be the most predictive one, reaching good accuracy results in both the internal and external test sets. Moreover, the advantage of the usage of machine learning methods is the possibility to immediately update the prediction as soon as new measurements are available: by postponing the prediction, the accuracy of the algorithm significantly increases. This thesis exhibits that clinical parameters commonly collected in clinical practices, mainly serum creatinine and urine, can be used as predictors of persistent renal failures outbreaks by machine learning algorithms. This approach constitutes an efficient and responsive alternative to biomarker tests that can help clinicians’ decision making when managing patients affected by Acute Kidney Injuries.