Prospective Validation of AKI Prediction

Overview

The purpose of this prospective observational study is to implement, deploy, and quantify accuracy of an existing Pediatric Early AKI Risk Score algorithm. The implementation will be facilitated using a Health Level 7 (HL7) Fast Healthcare Interoperability Resource (FHIR)-based architecture. Investigators will deploy this model and store results in a manner not viewable to the clinical team caring for the patient. To determine the accuracy of the implemented prediction model, Investigators will prospectively identify patients with AKI at 72 hours following ICU admission. Investigators hypothesize that this model will prospectively detect AKI with a sensitivity \>70% and a positive predictive value \>20%, both chosen a priori as 10% improvement over the initial Pediatric AKI Risk Score tool.

This is a single-center prospective observational study validating an AKI predictive model. Each model feature will be mapped to an appropriate FHIR-based resource. To mitigate the latency issues seen in other distributed CDS systems, Investigators have developed an asynchronous design where algorithm calculations are performed offline (e.g., not within the EHR) and risk scores are subsequently written back to the EHR. Importantly, in this deployment, model output and resulting clinical risk score will not be communicated to the treating clinicians. During the study period, Investigators will review charts daily for all patients admitted to the Golisano Children's Hospital PICU, a 12-bed facility adjacent to our 15-bed Pediatric Cardiac Intensive Care Unit (PCICU). Using a standard protocol to screen and identify patients by chart review, Investigators will generate a list of patients who meet AKI KDIGO criteria by SCr and urine output, along with recorded clinical information about these patients. At the conclusion of the study period, this list will be used as the "gold standard" and compared to the automated screening tool to determine the tool's test characteristics. Model assessment outcomes include sensitivity, positive predictive value (PPV), and number needed to alert (NNA) to prospectively identify AKI in a population of critically ill children. Additional outcomes include timeliness of identification based on model implementation (e.g., measured timestamps of algorithm prediction compared to manual, prospectively identified AKI development). Additionally, Investigators will report interventions and clinical outcomes of the prospectively identified patients with AKI, stratified by those predicted early by the model (within 24 hours of admission) versus not.