Source-linked AI summary
Tracking COVID-19 using online search
Vasileios Lampos, Maimuna S. Majumder, Elad Yom-Tov, Michael Edelstein, Simon Moura, Yohhei Hamada, Molebogeng X. Rangaka, Rachel A. McKendry, Ingemar J. Cox
TL;DR
The paper asks whether online search data can provide useful COVID-19 prevalence signals despite public-interest bias and limited conventional surveillance data. It develops symptom-based unsupervised and supervised models across countries, adjusts signals using news coverage, and evaluates transfer learning and forecasting. Search signals preceded confirmed cases and deaths by 16.7 and 22.1 days, respectively, while rarer symptoms and generic COVID-19 queries were stronger predictors.
Problem
COVID-19 surveillance needed complementary signals because online searches may reflect public concern as well as infection, while many locations lacked sufficient historical ground-truth data for supervised models.
Method
The study builds symptom-weighted unsupervised search models, adjusts for news coverage, transfers supervised models across countries, analyzes query relationships, and adds searches to autoregressive death forecasts.
Results
Search signals preceded confirmed cases by 16.7 (10.2−23.2) days and deaths by 22.1 (17.4−26.9) days, while rarer symptoms and generic COVID-19 queries were stronger predictors.
Takeaways & Limitations
Online search data can provide complementary national-level COVID-19 surveillance signals alongside established approaches.
Takeaways & Limitations
Evaluation is difficult because no definitive ground truth represents community-level disease rates, and supervised analyses depend on consistent sampling of confirmed cases or deaths.
Abstract
from arXiv · showhide
Previous research has demonstrated that various properties of infectious diseases can be inferred from online search behaviour. In this work we use time series of online search query frequencies to gain insights about the prevalence of COVID-19 in multiple countries. We first develop unsupervised modelling techniques based on associated symptom categories identified by the United Kingdom's National Health Service and Public Health England. We then attempt to minimise an expected bias in these signals caused by public interest --as opposed to infections-- using the proportion of news media coverage devoted to COVID-19 as a proxy indicator. Our analysis indicates that models based on online searches precede the reported confirmed cases and deaths by 16.7 (10.2 - 23.2) and 22.1 (17.4 - 26.9) days, respectively. We also investigate transfer learning techniques for mapping supervised models from countries where the spread of disease has progressed extensively to countries that are in earlier phases of their respective epidemic curves. Furthermore, we compare time series of online search activity against confirmed COVID-19 cases or deaths jointly across multiple countries, uncovering interesting querying patterns, including the finding that rarer symptoms are better predictors than common ones. Finally, we show that web searches improve the short-term forecasting accuracy of autoregressive models for COVID-19 deaths. Our work provides evidence that online search data can be used to develop complementary public health surveillance methods to help inform the COVID-19 response in conjunction with more established approaches.
Introduction
The paper develops online-search models for COVID-19 surveillance across multiple countries, addressing the challenge that search activity reflects both infections and public concern. It combines symptom-based unsupervised modelling, news-coverage adjustment, transfer learning, and supervised analyses to improve interpretation and forecasting.
- Online searches can reflect community disease trends but are also influenced by news coverage, reported statistics, and physical-distancing measures.
- Existing supervised search-based models require sufficiently valid, representative, and long historical surveillance data, which was unavailable for most locations.
- The paper develops unsupervised COVID-19 models using NHS/PHE symptom categories weighted by reported symptom occurrence, while incorporating news coverage to reduce concern-related bias.
- Transfer learning adapts supervised models from locations with available disease data to countries lacking comparable historical ground truth.
- Correlation and regression analyses identify generic COVID-19 terms and less common symptoms as useful predictors of confirmed cases, while adding searches to autoregressive models reduces COVID-19 death-forecasting error.
- The study evaluates multilingual and multicultural data from the United States, United Kingdom, Australia, Canada, France, Italy, Greece, and South Africa.
Results
Across multiple countries, weighted online-search models tracked COVID-19 dynamics, preceded reported clinical outcomes, supported transfer learning, and improved short-term death forecasting.
- Weighted search scores exceeded estimated seasonal averages during peaks and generally declined steeply after physical-distancing or lockdown measures.News-effect minimisation produced more conservative estimates and altered peak-period trends.
- 16.7 days was the average lead of search scores over confirmed cases, increasing to 22.1 days over COVID-19 deaths after excluding South Africa.The corresponding maximised correlations were .826 (.735−.917) for cases and .846 (.702−.990) for deaths.
- Transfer learning from Italy produced estimated case trends whose peaks appeared approximately 2–3 weeks before confirmed cases in most target countries.Australia and South Africa were exceptions; transferred estimates were standardised and derived from ensembles of models.
- 13.76 days was the mean correlation-maximising shift for the US, compared with 1.44 days for Greece, across transferred models.Other target-country shifts ranged from 5.24 days in Australia to 12.67 days in the UK.
- Queries containing “covid” were the most impactful predictors of confirmed cases and deaths, contributing 29.32% and 36.29%, respectively.Among symptoms, blue face, loss of smell, appetite loss, pink eye, and shortness of breath were the top five impactful features.
- 32.65% and 33.77% were the MAE reductions achieved by SAR-F versus AR-F for 7- and 14-days-ahead death forecasting.SAR-F also improved upon the PER-F baseline in the more challenging forecasting task.
Discussion
Online search signals showed practical value as early indicators of COVID-19 prevalence, while comparisons with clinical endpoints supported their temporal lead. Evaluation remains constrained by the lack of definitive community-level ground truth and inconsistent target sampling.
- Rarer symptoms and generic COVID-19 queries correlated better with, and were more predictive of, clinically reported metrics.
- 0.954 correlation was observed between transferred estimates and weekly RCGP swabbing positivity rates, compared with 0.816 for unsupervised and 0.855 for news-adjusted unsupervised scores.
- PHE incorporated the unsupervised models into weekly surveillance reports, using web-search trends alongside other systems to identify changes before conventional surveillance.
- Evaluation lacks a definitive community-level ground truth, while supervised analyses depend on consistent sampling of confirmed cases or deaths.
Methods
The analysis combines symptom-based online-search models, news-adjusted autoregression, cross-country transfer learning, multivariate query analysis, and forecasting models to study COVID-19 signals.
- Unsupervised symptom models: Nineteen FF100 symptom categories plus COVID-19 terminology form weighted query sets, with symptom weights based on conditional occurrence probabilities.The weighted current and historical time series are computed from query-category frequencies.
- News-media adjustment: News coverage is incorporated into autoregressive models to estimate and reduce the concern-related component of symptom-search activity.The method compares forecasting errors from models using search history alone versus search history plus current and lagged news coverage, then smooths the resulting estimates over seven-day windows.
- Transfer learning: Transfer learning maps source-country query features to target-country symptom categories, applies source elastic-net models, and aggregates predictions into target estimates with 95% confidence intervals.Query pairs are selected within symptom categories using correlations over temporal shifts, while an ensemble spans multiple sparsity levels.
- Cross-country analysis and forecasting: Joint country analysis normalises each country separately before correlating and regressing aggregated query frequencies against confirmed cases or deaths.The study also compares forecasting models with a persistence baseline and uses composite covariance functions that separately model searches, deaths, and their joint behaviour.
- Data and preprocessing: Google Health Trends provides daily search-query frequencies for specific areas, defined as target-term sessions divided by all sessions for each day and area.The data source is a non-public research interface.
Author contribution statement
The contribution statement assigns responsibility for the research, data, methods, experiments, writing, translations, and news-coverage inputs across the authors.
- Research and manuscript leadership: V.L. conceived the project, assembled most datasets, developed the methods, ran experiments, drafted the manuscript, and led the subsequent write-up.
- Data and translation contributions: M.S.M. provided news-coverage data for all countries, while several coauthors translated search-query groups into Italian, French, and South African languages.
Additional information about the data sets
The paper obtains epidemiological outcomes from ECDC, publishes its search-query inventory on figshare, and defines COVID-19 news coverage using specified terms.
- Outcome data: Confirmed COVID-19 cases and deaths were obtained from the European Centre for Disease Prevention and Control.
- Search-query data: All search queries used in the experiments are available through the paper’s figshare project.
- News classification: COVID-19-related news articles were identified using coronavirus-related terms, with an expanded Greek term list and a common list for other countries.
Supplementary equations
The supplementary material defines min-max and z-score normalisation for vectors and specifies mean absolute error for estimated versus ground-truth time series.
- Normalisation: Min-max normalisation maps a vector x ∈ R^n to a vector x̂ ∈ [0,1]^n.
- Standardisation: Z-score normalisation maps a vector x ∈ R^n to a standardised vector x̂ ∈ R^n using its mean and standard deviation.
- Error metric: Mean absolute error measures the average absolute difference between an estimated time series and its corresponding ground truth.
Supplementary figures and tables
The supplementary material compares online-search signals, symptom-weighting choices, transfer-learning estimates, query correlations, and forecasting models across countries. It also documents search-volume normalization, news-adjusted scores, temporal alignment with reported outcomes, and forecasting-error evaluation.
- Table S1 evaluates AR-F, SAR-F, and PER-F using mean absolute error for 7- and 14-day forecasts of daily confirmed cases across eight countries.Its final row min-max normalizes averages across countries, methods, and forecasting tasks.
- Figure S1 compares normalized Google-search frequencies for “coronavirus,” “weather,” and “the” in the US and UK using 14-day harmonic-mean smoothing.
- Figures S2 and S3 compare symptom-based COVID-19 search scores under occurrence-probability weighting versus uniform weighting, with news-media effects minimized and long-term trends shown.The scores cover eight countries from September 30, 2019 to May 24, 2020.
- Figure S4 aligns news-adjusted online-search scores with reported COVID-19 deaths and country-specific backward shifts that maximize correlation.The shifts range from 17 days in France to 52 days in South Africa.
- Figure S5 shows Italy-based transfer-learning models for seven countries, comparing ongoing ensemble estimates of confirmed cases with reported cases and marking physical-distancing or lockdown dates.
- Figure S6 presents cross-country correlation and regression analyses of search queries against COVID-19 deaths and confirmed cases in the US, UK, Australia, and Canada.The figure identifies positively and negatively correlated queries for deaths and cases after shifting case series by 19 days.
- Figures S7 and S8 show 14-days-ahead forecasts for deaths and confirmed cases, contrasting autoregressive forecasts with models incorporating online-search information.
- Figure S9 reports the average daily proportion of news articles about COVID-19 across the countries analyzed, with confidence intervals defined as two standard deviations above and below the mean.