Forecasting SARS-CoV-2 transmission and clinical risk at small spatial scales by the application of machine learning architectures to syndromic surveillance data

Abstract

Timely and well-informed syndromic surveillance is essential for effective public health policy. The monitoring of traditional epidemiological indicators can be lagged and misleading, which hampers efforts to identify hotspot locations. The increasing predominance of digitalized healthcare-seeking behaviour necessitates that it is fully exploited for the public benefit of effective pandemic management. Using the highest-resolution spatial data for Google Trends relative search volumes, Google mobility, telecoms mobility, National Health Service Pathways calls and website testing journeys, we have developed a machine learning early indicator modelling approach of SARS-CoV-2 transmission and clinical risk at small geographic scales. We trained shallow learning algorithms as the baseline against a geospatial neural network architecture that we termed the spatio-integrated long short-term memory (SI-LSTM) algorithm. The SI-LSTM algorithm was able to—for the assessed temporal periods—accurately identify hotspot locations over time horizons of a month or more with an accuracy in excess of 99%, and an improved performance of up to 15% against the shallow learning algorithms. Furthermore, in public health operational use, this model highlighted the localized exponential growth of the Alpha variant in late 2020, the Delta variant in April 2021 and the Omicron variant in November 2021 within the United Kingdom prior to their spatial dispersion and growth being confirmed by clinical data.