Severe coronavirus disease 2019 (COVID-19) requiring intensive care unit (ICU) admission is often associated with respiratory failure [1] as well as the development of extra-pulmonary organ dysfunctions. According to previous studies, acute kidney injury (AKI) is a frequent complication in 30% to 85% [[2], [3], [4], [5]] of ICU patients with COVID-19. The presence of multi-organ failure, including severe AKI and the use of RRT is often associated with very high mortality rates (ranging from 25 to 80%) [[6], [7], [8], [9], [10]] and requires intense resource allocation [11].

Clinical characteristics and risk factors for developing AKI and using RRT are well described in ICU patients without COVID-19 [12,13]. Moreover, studies have successfully used machine learning (ML) models to predict AKI [[14], [15], [16], [17], [18], [19]] in diverse populations of critically ill such as patients with sepsis [20], post-cardiac transplantation [21], post-liver transplantation [22] and under mechanical ventilation [23]. However, translating the findings of predictive models developed for non-COVID critically ill to COVID-19 patients may be challenging. Studies have described that risk factors and incidence of AKI in COVID-19 patients may differ from those observed in non-COVID-19 ICU patients [24]. A similar situation was already described with general ICU mortality models where the applicability to COVID-19 patients is limited as traditional models tend to underperform in this population due to differences in the disease course and factors associated with outcomes in these patients [25,26]. Therefore, specific models to predict the occurrence of AKI or the use of RRT in COVID-19 patients are needed and to the best of our knowledge, remain scarce.

The present study aimed to develop and validate a model to identify COVID-19 patients using RRT through data obtained in the early stage of ICU admission. The hypothesis that an accurate model could be developed using a small number of clinical and laboratory variables readily available at ICU admission, was tested. In addition, different ML models with logistic regression methods were compared.