Predicting acute kidney injury onset with a random forest algorithm using electronic medical records of COVID-19 patients: the CRACoV-AKI model

Abstract

Introduction: Acute kidney injury (AKI) is a serious and common complication of SARS‑CoV‑2 infection. Most risk assessment tools for AKI have been developed in the intensive care unit or in elderly populations. As the COVID‑19 pandemic is transitioning into an endemic phase, there is an unmet need for prognostic scores tailored to the population of patients hospitalized for this disease.

Objectives: We aimed to develop a robust predictive model for the occurrence of AKI in hospitalized patients with COVID‑19.

Patients and methods: Electronic medical records of all adult inpatients admitted between March 2020 and January 2022 were extracted from the database of a large, tertiary care center with a reference status in Lesser Poland. We screened 5806 patients with SARS‑CoV‑2 infection confirmed with a polymerase chain reaction test. After excluding individuals with lacking data on serum creatinine levels and those with a mild disease course (<7 days of inpatient care), a total of 4630 records were considered. Data were randomly split into training (n = 3462) and test (n = 1168) sets. A random forest model was tuned with feature engineering based on expert advice and metrics evaluated in nested cross‑validation to reduce bias.

Results: Nested cross‑validation yielded an area under the curve ranging between 0.793 and 0.807, and an average performance of 0.798. Model explanation techniques from a global perspective suggested that a need for respiratory support, chronic kidney disease, and procalcitonin concentration were among the most important variables in permutation tests.

Conclusions: The CRACoV‑AKI model enables AKI risk stratification among hospitalized patients with COVID‑19. Machine learning–based tools may thus offer additional decision‑making support for specialist providers.