Prediction of hospital-onset COVID-19 using networks of patient contact: an observational study

Abstract

Purpose: Predicting healthcare-acquired infections (HAIs) has the potential to revolutionise the prevention and control of transmissible infections. Existing prediction models for HAIs, however, fail to capture fully the contact-driven nature of infectious diseases. Here, we investigate the epidemiological predictivity of patient contact patterns through a forecasting model for hospital-onset COVID-19 infection (HOCI). Methods & Materials: Our cohort comprises all patient admissions at a large London NHS Trust between 1/04/2020 and 1/04/2021. For patients, we consider (i) their hospital pathway, (ii) patient contacts, and (iii) date of COVID-19 infection. We consider rolling 14-day windows and forecast patient infection over the subsequent 7 days. Over each window, we construct a patient contact network and compute network features that capture contact centrality. We then combine network features, hospital environmental variables and patient clinical data to predict subsequent infections. Results: A total of 51,157 patient admissions/episodes were observed during the study. Across all models, we find that contact-network features showed the highest performance (0.91 AUC-ROC). A reduced model with the six most predictive variables was almost as predictive and contained five features from patient contact (including direct contact with and network proximity to infectious cases) and only one environmental variable (length of stay). Conclusion: Our results reveal that the number of direct contacts and network proximity to infectious patient(s) are highly predictive of HOCI. Such contact-based risk factors are easily extracted from routinely collected electronic health records providing a highly accessible route to improve personalised disease prognostics in future models.