Predicting SARS-CoV-2 infection among hemodialysis patients using multimodal data

Area of Science:

• Nephrology
• Infectious Diseases
• Data Science

Background:

• The COVID-19 pandemic disproportionately affected dialysis patients.
• Accurate prediction models for SARS-CoV-2 are vital for managing outbreaks in dialysis settings.
• The effectiveness of existing models over time is uncertain.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting SARS-CoV-2 infection in dialysis patients during the incubation period.
• To assess the model's performance and identify key predictors as the pandemic evolved.

Main Methods:

• Developed an XGBoost machine learning model using diverse data sources: demographic, clinical, laboratory, treatment, vaccination, socioeconomic, and county-level COVID-19 data.
• Evaluated model performance on a rolling basis from April 2020 to February 2022.
• Investigated the evolution of prediction power and risk factors over time.

Main Results:

• The model achieved an AUROC of 0.75 from April to August 2020, outperforming a previous model.
• Prediction performance fluctuated, with the lowest AUROC of 0.6 in late 2021/early 2022.
• At a 20% false-positive rate, the model detected 40% of positive cases over the study period.
• County-level infection data and vaccination status were significant predictors; nursing home residency was less predictive post-vaccination.

Conclusions:

• COVID-19 prediction model dynamics change with pandemic evolution.
• County-level infection rates and vaccination status are critical for effective early prediction.
• The developed model can identify SARS-CoV-2 infections during incubation, with potential for clinical application.