Artificial Intelligence for COVID-19 Risk Classification in Kidney Disease: Can Technology Unmask an Unseen Disease?

Abstract

Background

We developed two unique machine learning (ML) models that predict risk of: 1) a major COVID-19 outbreak in the service county of a local HD population within following week, and 2) a hemodialysis (HD) patient having an undetected SARS-CoV-2 infection that is identified after following 3 or more days.

Methods

We used county-level data from United States population (March 2020) and HD patient data from a network of clinics (February-May 2020) to develop two ML models. First was a county-level model that used data from general and HD populations (21 variables); outcome of a COVID-19 outbreak in a dialysis service area was defined as a clinic being located in one of the national counties with the highest growth in COVID-19 positive cases (number and people per million (ppm)) in general population during 22-28 Mar 2020. Second was a patient-level model that used HD patient data (82 variables) to predict an individual having an undetected SARS-CoV-2 infection that is identified in subsequent ≥3 days.

Results

Among 1682 counties with dialysis clinics, 82 (4.9%) had a COVID-19 outbreak during 22-28 Mar 2020. Area under the receiver operating characteristic curve (AUROC) for the county-level model was 0.86 in testing dataset. Top predictor of a county experiencing an outbreak was the COVID-19 positive ppm in the general population in the prior week. In a select group (n=11,664) used to build the patient-level model, 28% of patients had COVID-19; prevalence was by design 10% in the testing dataset. AUROC for the patient-level model was 0.71 in the testing dataset. Top predictor of an HD patient having a SARS-CoV-2 infection was mean pre-HD body temperature in the prior week.

Conclusions

Developed ML models appear suitable for predicting counties at risk of a COVID-19 outbreak and HD patients at risk of having an undetected SARS-CoV-2 infection.