Deep learning for subphenotype identification in COVID-19-associated AKI

Background: Acute kidney injury (AKI) is common in COVID-19 and associated with increased adverse outcomes. COVID-associated AKI (COVID-AKI) pathophysiology is heterogenous, and deep learning may discover subphenotypes. Methods: We used data from 5 New York City hospitals from adults admitted between March '20-March '21 with COVID and AKI, excluding patients with kidney failure. An autoencoder compressed 58 features containing comorbidities, the first laboratory values and vital signs within 48 hours of admission for unsupervised K-means clustering. Outcomes were mortality, dialysis, mechanical ventilation, and ICU admission. Results: We identified 1634 patients with COVID-AKI and discovered 3 subphenotypes. Subphenotype one had 576 patients (35%);two had 635 patients (39%), and three had 423 patients (26%) (Table 1). Subphenotype three had the lowest median blood pressures, highest median BMI, and highest rates of all outcomes. (Figure 1) Conclusions: There are distinct subphenotypes in COVID-AKI indicating the heterogeneity of this condition.