Phd in SCHOOL OF STATISTICAL SCIENCES

Thesis title: Prediction of acute kidney injury using the Electronic Medical Records of a pediatric cardiac intensive care unit

Acute Kidney Injury (AKI) is a frequent complication in hospitalized patients significantly associated with mortality, length of stay, and healthcare cost. Management of AKI presents an important challenge and clinicians may be helped by robust prediction models for risk evaluation, foster prevention, and recognition. The advances in clinical informatics and the increasing availability of electronic medical records (EMR) have favored the development of predictive models of risk estimation in AKI. In this dissertation, we analyze the problem of predicting the AKI stage during the patient’s stay in the intensive care unit using retrospectively the Electronic medical records (EMRs) recently introduced in the Pediatric Intensive Care Unit (PCICU) of "Ospedale Pediatrico Bambino Gesù". After the initial phase of data selection, extraction, and management of missing data, we develop a random forest classification model including a variable selection step with the aim of predicting the stage of AKI 48 hours in advance in both binary and multiclass cases. The performance obtained in terms of AUC-ROC for binary cases and accuracy for multiclass cases are always very good compared with other recent attempts in the literature. The list of the most important variables obtained in the various classifications highlights the importance of some of the expected variables (such as creatinine) reported in other studies in the literature but also the presence of variables that are specific to pediatric patients under examination (such as PIM3). Moreover, we develop other classifications using the GAMS and BN models that have the benefit of offering a more interpretable approach. Although these results are inferior to the RF, they are comparable with many outcomes reported in the literature. The plot obtained with GAMs and the structure of DAG achieved with BN are consistent with a possible medical explanation and would present further interpretation hints for the doctors about the onset of AKI. Finally, we observe that all implemented models confirm the possibility of making an accurate prediction of the AKI stage using the PCICU. These models can be potentially included in a web interface and, in perspective, be integrated into the EMR of PCICU. This tool would allow the doctors to predict prospectively the patient’s stage of AKI and evaluate how to intervene if necessary. In order to proceed with this, it would be necessary in the future to implement the export of a larger dataset adding new data acquired in the meantime in PCICU.