Computerized decision support to assess the kidney function in critically ill patients

In critically ill patients, both increased and decreased renal clearance are highly prevalent and may vary during the course of critical illness. Acute kidney injury is a type of organ failure that is highly prevalent in many conditions that cause severe critical illness, and has an association with worse short- and long-term clinical outcomes. Fluctuations in renal clearance will influence the plasma levels and clearance of renally excreted drugs, and may result in unfavorable consequences such as drug toxicity and treatment failure. Predicting these fluctuations in kidney function could allow for more personalized drug dosage, and less treatment failure, potentially leading to better clinical outcomes. In addition, even though the prevention and management of AKI are mainly supportive, early predictions of a reducing or fluctuating kidney function could act as in silico biomarkers, to stratify patients according to their risk, and may aid in developing targeted therapies to prevent or mitigate the course of kidney injury. Despite the importance of kidney function prediction models, the focus of current predictions for kidney function made by machine-learning models has been on predicting the development of AKI or augmented renal clearance. Although some models have been demonstrated with good predictive performance and outperformed the physicians in external validation, they still need more external validations in independent datasets before they can be applied to clinical practice. Furthermore, since kidney function is continuous, predicting the complete spectrum of kidney function is more in accordance with clinical and physiological reality.

The general objective of this thesis is to apply advanced data analysis techniques and machine-learning algorithms to routinely collected clinical data from critically ill patients to develop and validate prediction models for kidney function. This thesis consisted of three primary objectives, where we focused on common medical phenomena and measurements with important clinical implications in the ICU.

In the first part, we performed an external validation of a prediction model for acute kidney injury (AKI), the AKIpredictor, in critically ill adults of the University Hospitals Leuven who were included in the large multicenter M@tric database between 2013 and 2018. This M@tric database contains high-quality and intricately interconnected data from all adult patients annually admitted to the ICU from 2013 to 2018. Even though this external validation dataset was collected 10 years after the original development cohort, the AKIpredictor still demonstrated its robustness. These results verified the AKIpredictor’s potential to be a useful tool for the early detection of AKI patients.

AKI’s progression and recovery are crucial since they are closely related to end-stage kidney disease and progressive renal dysfunction. Nevertheless, large databases frequently lack a good AKI recovery evaluation because the recovery definition is not uniform and baseline serum creatinine is frequently unknown, which presents difficulties for the development of AKI recovery prediction models. Using a large multicenter EPaNIC randomized controlled trial database, where two parenteral nutrition strategies were compared in 4640 critically ill adults between August 2007 and November 2010, we developed and externally validated prediction models for AKI recovery at hospital discharge in critically ill patients with AKI stage 3 during their ICU stay. For the overall ICU population, the developed AKI recovery models only exhibited similarly unsatisfactory discrimination to the plasma neutrophil gelatinase-associated lipocalin (NGAL) measured on the first day of AKI stage 3, which was poorer than the reference that only based on age. For patients with cardiac surgery, the developed models had better performance over NGAL_AKI3 and the reference. The model is of limited clinical utility due to its poor predictive performance, which may be caused by the multiple pathophysiological processes but also by the definition of AKI recovery, which is still debated.

In the second part, we carried out an external validation of a prediction model for augmented renal clearance (ARC), the ARC predictor, in adult COVID-19 pneumonia patients admitted to critical care at the University Hospitals Leuven from February 2020 to January 2021. ARC affects 20-65% of critically ill patients and is associated with decreased exposure to commonly used antibiotics and anticoagulants. Therefore, in this study, we externally validated the ARC predictor in a recent critically ill COVID-19 cohort. Despite the slightly worse calibration, the ARC predictor showed robust performance with good discrimination and a wide clinical usefulness range. The robust performance is noteworthy, given the large differences in patient characteristics between this critically ill COVID-19 cohort and the original ARC predictor development cohort (the cohort in the presented study showed an ICU length of stay almost twice as long, 14 vs. 8 days). The promising performance identified in this study confirmed the ARC predictor’s potential to be a useful tool for the identification of patients with high risk of ARC.

In the third part, we focused on the evaluation of daily kidney function instability and prediction of daily kidney function, based on daily measured creatinine clearance (CrCl). It is well known that kidney function can change rapidly during critical illness, and that this change may have important implications for modifying the dosage of drugs that are excreted through the kidneys. Nevertheless, the actual incidence and degree of fluctuations have never been described systematically. Therefore, in this project, we investigated daily changes in kidney function as defined by the daily differences in CrCl in critically ill adults admitted to the ICUs of the University Hospitals Leuven, included in the EPaNIC RCT database. Using a large amount of daily creatinine clearance measurements, we discovered that on approximately 35–40% of days, critically ill patients may experience potentially clinically significant changes in kidney function on a daily basis. Furthermore, this instability was more pronounced in the first week of ICU admission and at higher CrCl values. Future studies in independent cohorts of critically ill patients are needed to confirm these findings and to examine the factors associated with fluctuations in renal clearance.

In patients with critical illness, potentially clinically significant changes in kidney function were found to happen to 35-40% of days during the ICU stay in the previous study. However, the current methods for kidney function estimation had limited capacity to accurately reflect kidney function in critically ill patients, since they were developed based on healthy people. Additionally, these methods were based on measurements from the past, which may lead to an estimated kidney function lagging behind the actual kidney function. Therefore, we developed and validated models for daily prediction of measured CrCl, named “CrCl predictor”, in critically ill adults admitted to ICUs of the University Hospitals Leuven, included in the EPaNIC RCT database. Three models with progressively more features with increasing data resolution were developed. All models demonstrated good performance when tested on a large external validation dataset of 20930 patients of the University Hospitals Leuven who were included in the large multicenter M@tric database between 2013 and 2018. The same good performance was observed when the model was compared with the reference reflecting current clinical practice, which assumed that CrCl remained unchanged.

To understand the external validity in independent datasets and the added value of these models to physicians’ predictions, we conducted a prospective observational study in 197 critically ill adults admitted to surgical ICUs of the University Hospitals Leuven between January 2022 and April 2022. Treating physicians of the surgical ICU team were asked to predict CrCl and to report their corresponding levels of confidence via a well-designed questionnaire survey. The predictions made by ICU physicians were compared with the ones made by the developed CrCl predictor models, where the CrCl predictor showed robust performance with comparable accuracy to that observed in the original model development study. Furthermore, the CrCl predictor demonstrated slightly smaller prediction errors than the staff members and senior residents, and a much smaller prediction error than the junior residents. The lack of statistically significant differences implied that the CrCl predictor could perform at least as well as the ICU physicians. These findings suggested the potential added value of the CrCl predictor to physicians’ predictions, especially for junior residents, and the potential of the CrCl predictor as a covariate to be integrated in the PK model to help optimize the renally cleared drug exposure. 

Despite the good performance in the comparison with the reference reflecting current clinical practice and the ICU physicians, whether the CrCl predictor can help improve the patient care and outcome remains to be investigated in future interventional studies. For the purpose of evaluating the model performance in a real clinical setting, we developed a prototype software that integrated the developed CrCl predictor models and visualized the prediction results along with prediction explanations in a user-friendly manner. Prediction explanation is crucial since physicians prefer sufficient evidence-based scientific support, in order to provide reasonable intervention and to modify their treatment strategies for vulnerable ICU patients. The developed software was designed to function normally in a real-time setting, which was able to calculate the predictions within less than 30 seconds for the patients staying less than 9 days in the ICU, consisting of 75% of the study cohort in Chapter 7. Nevertheless, a prospective study is still needed to technically validate its functionality. Once a technical validation is finished, the developed software will be ready for an interventional study to investigate whether patient kidney function management and/or patient outcome can be improved with the additional patient kidney function information provided by the developed software.

In conclusion, this thesis focused on the application of advanced data analysis techniques and machine-learning algorithms to routinely collected clinical information from critically ill patients, in order to gain new insights into kidney function of critically ill patients. Many research questions have been thoroughly studied, encompassing the development and validation of prediction models for kidney function, examination of daily kidney function fluctuations, prospective comparison of models with ICU physicians, and integration of the developed kidney function prediction models into a user-friendly software. The developed prototype software provides the possibility for interventional study and assessment of the research findings’ efficacy and safety, as well as their impact on kidney function management, patient care, and outcome.