The COVID-19 pandemic has placed a significant burden on the UK's health and economy. Accurate short-term prediction of COVID-19 outbreaks is crucial for effective public health interventions and resource allocation. Existing forecasting methods often lack adaptability to rapidly evolving situations. This research focuses on developing a dynamic, data-driven model capable of providing accurate short-term predictions of COVID-19 case growth rates at a local level within the UK. The model leverages readily available digital data sources, such as Google search trends for COVID-19 symptoms, population mobility data from Google Community Mobility Reports, and vaccination coverage data. The integration of these diverse data sources aims to capture a more comprehensive picture of COVID-19 transmission dynamics and improve prediction accuracy compared to existing methods. The study's ultimate goal is to create a tool that assists local authorities in making informed decisions regarding public health measures and resource allocation, ultimately contributing to more effective pandemic management.

Previous research has explored the use of digital data for predicting COVID-19 outbreaks. Studies have shown the value of using population-level mobility data (from sources like Google and Apple) and internet search trends for infection-related symptoms in modelling infectious disease outbreaks. These digital metrics provide near real-time insights into population behavior without compromising individual privacy. However, past studies often used static models with a fixed set of predictors, limiting their adaptability to changing pandemic conditions. The emergence of new variants and the widespread rollout of COVID-19 vaccines have significantly altered transmission dynamics, highlighting the need for dynamic models that can adapt to these changes. While some studies have integrated multiple digital data sources to improve predictions, the dynamic aspect and the inclusion of vaccination data often remain missing. This study aims to address these limitations by developing a flexible algorithm that can adjust its predictors over time, incorporating the most up-to-date information available.

This study employed a dynamic supervised machine learning algorithm based on log-linear regression to predict 1-, 2-, and 3-week ahead growth rates of COVID-19 cases at the LTLA level in the UK. The model incorporated three key data sources: Google Search Trends data for 173 COVID-19 related symptoms, Google Community Mobility Reports providing data on population movement across six locations (workplaces, residential areas, parks, retail and recreational areas, grocery and pharmacy, and transit stations), and COVID-19 vaccination coverage data. Data were aggregated to a weekly level and linked by week and LTLA. Missing values were imputed using linear interpolation. The modelling period spanned from June 1st, 2020, to November 14th, 2021. The model selection process involved several iterative steps. It started with a baseline model including LTLA, mobility metrics, vaccination coverage, and eight base symptoms (cough, fever, fatigue, diarrhea, vomiting, shortness of breath, confusion, and chest pain). The algorithm then optimized for the best time lag combination between predictors and growth rates. A forward data-driven method was used to select additional symptoms that improved model predictability. Finally, the algorithm assessed different predictor combinations, using a 4-week rolling mean squared error (MSE) for retrospective evaluation. Prospective model performance was evaluated at eight checkpoints, five weeks apart, using prospective MSE. Two reference models – a naïve model (assuming no change in growth rate) and a fixed-predictors model (using the optimal model from the first checkpoint for all subsequent checkpoints) – were used for comparison. Sensitivity analyses were conducted to assess the impact of including additional symptoms. Finally, the results were visualized in a publicly accessible web application, COVIDPredLTLA.

The study included data from 367 LTLAs with complete data. The median MSEs for the optimal models were 0.12 (IQR: 0.08–0.22), 0.29 (0.19–0.38), and 0.37 (0.25–0.47) for 1-week, 2-week, and 3-week ahead predictions, respectively. Compared to naïve models, the optimal models showed a 21–35% reduction in MSE across all prediction timeframes. The advantage of dynamic models over the fixed-predictors model became more pronounced after several updates. Geographical variations in MSE were observed, with lower MSEs in central England and higher MSEs in Scotland and southwest England. Sensitivity analyses including additional symptoms did not significantly improve predictive accuracy. The retrospective 4-week MSE and prospective MSE showed similar decreasing trends over time, with the absolute difference diminishing after the first 3-4 checkpoints. The online application, COVIDPredLTLA, provides real-time predictions for the current week and the next two weeks, considering both publication date and specimen collection date.

The study demonstrates the effectiveness of a dynamic supervised machine learning approach for short-term prediction of local COVID-19 growth rates. The use of multiple digital data sources and a flexible algorithm that adapts to changing conditions resulted in improved prediction accuracy compared to simpler models. The consistently superior performance of the dynamic model, particularly during periods of rapid case growth like the Delta variant surge, underscores its value in real-world applications. The findings emphasize the importance of using dynamic models that can adapt to evolving pandemic dynamics. The publicly available web application allows for real-time predictions, providing valuable information to local health authorities for informed decision-making. The ability to predict short-term changes in case growth rates enables timely implementation of public health measures and improved resource allocation.

This study successfully developed and validated a dynamic supervised machine-learning model for predicting short-term COVID-19 growth rates at the local level in the UK. The model's performance, particularly during periods of rapid case increases, highlights its potential for supporting public health decision-making. The online application, COVIDPredLTLA, provides a valuable tool for real-time monitoring and forecasting. Future research could explore incorporating additional data sources, such as data on testing capacity, hospitalization rates, and demographic factors, to further enhance prediction accuracy. Investigating the model's performance in predicting the impact of new variants and different intervention strategies would also be valuable.

Several limitations should be considered. The accuracy of real-time predictions can be affected by reporting delays. The model uses publication date rather than specimen collection date as the primary outcome, which could lead to underestimation of real-time cases. Changes in testing practices and potential underrepresentation of certain age groups in digital data sources might influence model accuracy. The uncertainty intervals in the web application mainly reflect model parameter variability, not input data variability. The model does not account for other prevention measures or climate factors. Finally, due to data limitations, the model's performance in relation to specific SARS-CoV-2 variants could not be assessed.