Machine learning model for early prediction of acute kidney injury (AKI) in pediatric critical care

Abstract

Background: Acute kidney injury (AKI) in pediatric critical care patients is diagnosed using elevated serum creatinine, which occurs only after kidney impairment. There are no treatments other than supportive care for AKI once it has developed, so it is important to identify patients at risk to prevent injury. This study develops a machine learning model to learn pre-disease patterns of physiological measurements and predict pediatric AKI up to 48 h earlier than the currently established diagnostic guidelines.

Methods: EHR data from 16,863 pediatric critical care patients between 1 month to 21 years of age from three independent institutions were used to develop a single machine learning model for early prediction of creatinine-based AKI using intelligently engineered predictors, such as creatinine rate of change, to automatically assess real-time AKI risk. The primary outcome is prediction of moderate to severe AKI (Stage 2/3), and secondary outcomes are prediction of any AKI (Stage 1/2/3) and requirement of renal replacement therapy (RRT). Predictions generate alerts allowing fast assessment and reduction of AKI risk, such as: "patient has 90% risk of developing AKI in the next 48 h" along with contextual information and suggested response such as "patient on aminoglycosides, suggest check level and review dose and indication".

Results: The model was successful in predicting Stage 2/3 AKI prior to detection by conventional criteria with a median lead-time of 30 h at AUROC of 0.89. The model predicted 70% of subsequent RRT episodes, 58% of Stage 2/3 episodes, and 41% of any AKI episodes. The ratio of false to true alerts of any AKI episodes was approximately one-to-one (PPV 47%). Among patients predicted, 79% received potentially nephrotoxic medication after being identified by the model but before development of AKI.

Conclusions: As the first multi-center validated AKI prediction model for all pediatric critical care patients, the machine learning model described in this study accurately predicts moderate to severe AKI up to 48 h in advance of AKI onset. The model may improve outcome of pediatric AKI by providing early alerting and actionable feedback, potentially preventing or reducing AKI by implementing early measures such as medication adjustment.

Keywords: AKI; Acute kidney injury; Machine learning; Pediatric critical care; Predictive model.

Fig. 1

shows an example patient’s AKI disease trajectory and uses the prediction made 36 h before KDIGO Stage 2/3 AKI onset to demonstrate the inner-workings of the model. The top plot shows the patient’s measured serum creatinine values, with AKI onset time referenced as Time 0. The middle plot shows the predicted AKI risk up to prediction time—36 h before AKI onset. It also shows a mockup of the AKI alert that a user would see, including the patient context and suggested actions. The top three predictors contributing the highest risk to this specific prediction are displayed. The bottom portion demonstrates that the model is made up of age-dependent ‘weak classifiers’ of AKI risk based on single predictor values. The predicted AKI risk is the sum of weak classifier risks of all input predictors. Two example weak classifiers are shown. The first is the classifier for creatinine rate of change (CRoC). In the top plot, the example patient’s serum creatinine increases slowly under the AKI threshold prior to prediction time. The increase results in a positive CRoC value and elevated CRoC weak classifier risk of 0.60, as marked on the CRoC classifier plot. At the same time, the patient continuously received drugs with high nephrotoxic potential, shown by triangular ticks marking times of medication administration in the top plot. This results in the high-nephrotoxic drugs classifier risk being elevated to 0.14 (bottom plot). The ellipses (…) in the figure are placeholders for additional predictors not shown due to room constraints

Fig. 2

a AUROC of the model developed in this study and the renal angina index (RAI). Model AUROC increases closer to onset time, especially inside the training window of 48 to 24 h. b The receiver-operator curve at 30 h before onset time for the model and RAI. H1, H2 and H3 represent results from the holdout test data of the three hospitals. RAI results are shown for holdout test data from all three hospitals (H1-3)