AKI affects up to a quarter of pediatric critical care patients [
1], and is independently associated with higher mortality, longer lengths of stay, and subsequent development of chronic kidney disease [
2‐
4]. Currently, AKI is diagnosed using Kidney Disease Improving Global Outcomes (KDIGO) clinical practice guidelines, based on serum creatinine and urine output [
5]. However, since renal impairment typically precedes increases in creatinine, staging guidelines only detect AKI after renal injury or impairment has already set in. Whilst in pediatric intensive care units (PICU) there are often no specific treatments to reverse AKI after it has developed [
6], some studies have shown that early improvements in renal function after AKI may lead to better outcomes [
1,
5,
7]. Therefore, early prediction of AKI is important for identifying patients at risk of developing AKI and intervening early to improve outcomes. While AKI is multifactorial in PICU patients, it most commonly occurs following a period of renal hypoperfusion due to hypotension. Simple interventions which might improve renal function include ensuring adequate renal perfusion with intravascular filling or inotropes and avoiding or reducing nephrotoxic drugs. The Acute Dialysis Quality Initiative (ADQI) group recommended developing machine learning models for early prediction of moderate to severe AKI (Stage 2/3) between 48 and 72 h before diagnosis, and suggested that the prediction model should present information about patient measurements contributing to these risks and provide feedback to practitioners regarding potential actionable items [
6]. Many research groups have tackled early prediction of AKI using electronic health records (EHR) data [
8‐
11], but no model so far explains the rationale behind specific predictions despite a clear need for explainable and actionable predictions [
6,
12]. In pediatric patients where physiology differs greatly with age, developing a predictive model that learns age-appropriate signs of early AKI remains an additional challenge. This study aims to develop a prediction model of AKI for general pediatric critical care patients, running in real-time, that can detect subtle ongoing changes in patient physiology and alert caregivers about patients at high risk of AKI and provide interpretable context and suggested actions. The primary outcome is the ability to predict the onset of moderate to severe AKI 6 to 48 h before it develops. The same model is also assessed on secondary AKI-related outcome measures, including development of any AKI (Stage 1/2/3) and requirement of renal replacement therapy (RRT). To our knowledge, this is the first AKI prediction model built to explain each prediction, and the first multi-center validated model for general pediatric critical care AKI prediction.