Source-linked AI summary
Global disease monitoring and forecasting with Wikipedia
Nicholas Generous, Geoffrey Fairchild, Alina Deshpande, Sara Y. Del Valle, Reid Priedhorsky
TL;DR
The paper addresses the need for disease surveillance that is accurate and timely despite the cost and delay of traditional monitoring and limitations of existing internet-based methods. It uses freely available Wikipedia access logs with language as a location proxy and simple statistical models, finding promising present estimates, forecasts up to 28 days, and possible model transfer across locations.
Problem
Existing surveillance is accurate but costly and slow, while internet-based methods face gaps in scientific peer review, disease and country coverage, and forecasting.
Method
The paper analyzes freely available Wikipedia article access logs using language as a proxy for location and a systematic article-selection procedure with simple statistical models.
Results
Models were successful in 8 of 14 disease-location contexts, with r2 up to 0.92 and same-disease models showing similarity up to 0.81 across locations.
Takeaways & Limitations
Wikipedia data appear useful for disease information and may support forecasting and model use where traditional surveillance is poor or nonexistent.
Takeaways & Limitations
Article selection can omit unavailable target-language articles and mishandle redirects, while improvements such as nonlinear models and noise suppression remain future work.
Abstract
from arXiv · showhide
Infectious disease is a leading threat to public health, economic stability, and other key social structures. Efforts to mitigate these impacts depend on accurate and timely monitoring to measure the risk and progress of disease. Traditional, biologically-focused monitoring techniques are accurate but costly and slow; in response, new techniques based on social internet data such as social media and search queries are emerging. These efforts are promising, but important challenges in the areas of scientific peer review, breadth of diseases and countries, and forecasting hamper their operational usefulness. We examine a freely available, open data source for this use: access logs from the online encyclopedia Wikipedia. Using linear models, language as a proxy for location, and a systematic yet simple article selection procedure, we tested 14 location-disease combinations and demonstrate that these data feasibly support an approach that overcomes these challenges. Specifically, our proof-of-concept yields models with $r^2$ up to 0.92, forecasting value up to the 28 days tested, and several pairs of models similar enough to suggest that transferring models from one location to another without re-training is feasible. Based on these preliminary results, we close with a research agenda designed to overcome these challenges and produce a disease monitoring and forecasting system that is significantly more effective, robust, and globally comprehensive than the current state of the art.
2 Author Summary
The paper evaluates Wikipedia access logs as an open source for disease surveillance, addressing limitations in traditional and emerging internet-based approaches. Its proof of concept supports present estimation and forecasting while suggesting broader geographic applicability.
- Wikipedia access logs provide a freely available data source for disease surveillance models.The approach is intended to address gaps in scientific peer review, disease and country coverage, and forecasting capabilities.
- 28 days was the maximum forecasting horizon tested in the proof of concept.
- Models could potentially support surveillance in places without official data for model training.
- The paper outlines a path toward a reliable, scientifically sound, operational global disease surveillance system.
3 Introduction
The paper proposes Wikipedia access logs as an open basis for disease surveillance and forecasting, addressing gaps in openness, breadth, transferability, and forecasting evaluation. Proof-of-concept experiments across 14 location-disease contexts support present estimation, short-term forecasting, and possible model translation between locations.
- Motivation: Traditional surveillance is accurate but costly and delayed, while internet-based approaches remain limited by weak review, narrow coverage, and underdeveloped forecasting.These limitations motivate an open, operationally useful alternative.
- Approach: Wikipedia access logs provide a freely available data source for building open disease-surveillance models.The paper emphasizes that complete data and source code support review, replication, improvement, and deployment.
- Transferability: Models can be adapted to new contexts by supplying a reliable incidence time series and selecting input articles.The authors report effective models across several contexts despite using a simple article-selection procedure.
- Transferability: Similar models across locations suggest that inter-language article mappings could enable translation without retraining.The proposed transfer mechanism uses Wikipedia’s inter-language article links.
- Forecasting: Simple models retain usefully high r2 for forecasts extending a few days or weeks into the future.The paper reports forecasting value up to the 28-day limit of its tests and suggests reasonably tight confidence intervals.
- Approach: The proof of concept evaluates 14 location-disease contexts, uses language as a location proxy, and applies a systematic article-selection procedure.The study also tests daily shifts, including forecasting value, rather than restricting evaluation to weekly granularity.
4 Methods
The study combines Wikipedia access logs with official disease-incidence reports to build linear models across 14 disease-location contexts. It uses language as a location proxy, systematic article selection, and aligned time series while accounting for data-availability constraints and redirect complications.
- Data sources and study contexts: The analysis uses Wikipedia article access logs and official disease-incidence reports to model 14 disease-location contexts.The data cover approximately three years, and Table 1 records the diseases, locations, language proxies, dates, incidence resolution, and data sources.
- Data sources and study contexts: The selected contexts were constrained by the availability of reliable, high-resolution official incidence data and mature, sufficiently trafficked disease articles.The researchers selected 14 contexts using professional judgement, making the list informative rather than comprehensive.
- Data sources and study contexts: Country-specific access counts are approximated through language editions because Wikipedia logs provide global request totals rather than geographic counts.This proxy works best when a language is concentrated in one country and is less suitable for globally distributed languages such as Spanish.
- Article selection: Article inputs begin with a disease article and selected linked pages on symptoms, pathogens, treatments, biological processes, and epidemiology, then map them to target-language counterparts.Missing target-language articles are omitted, and article titles are encoded into the form used by the access logs.
- Article selection: Redirects complicate article selection because access logs record requests under the redirect name rather than its target, and redirect mappings change over time.The study therefore leaves reliable redirect aggregation for future work, while its procedure uses the target article when a redirect is encountered.
- Model construction: Hourly article requests are aggregated to the incidence-data frequency and aligned into comparable article and disease time series for model evaluation.The procedure ignores time zones, introducing a possible temporal offset of up to 23 hours.
5 Results
Across 14 disease-location contexts, 8 models produced useful nowcasting and forecasting estimates, while failures reflected subtle official-data patterns or unfavorable Wikipedia signal-to-noise ratios. Successful models showed forecasting value, possible cross-location transferability, and performance limits that require complementary evaluation and improved modeling.
- Evaluation: Model quality was not fully captured by r2: tuberculosis in China had r2 = 0.66 yet captured its baseline and three peaks, while subjective evaluation judged its timescale utility unclear.The authors therefore identify a need for additional metrics such as RMSE or more complex analyses.
- Forecasting: Models for dengue and influenza retained significant forecast value through the 28-day analysis limit, whereas tuberculosis in China became less effective as the forecast interval increased.The authors relate the contrast to seasonality and, for tuberculosis, a 2–12 week incubation period.
- Failure modes: Failures arose when official patterns were flat, noisy, too slow, or obscured by Wikipedia accesses unrelated to infection observations.The authors suggest longer analysis periods, nonlinear or improved regression methods, noise suppression, and disease-specific tuning as future directions.
- Performance summary: 8 of 14 contexts produced useful estimates for both nowcasting and forecasting, while 6 were unsuccessful.Three failures involved patterns too subtle to capture, and three involved unfavorable signal-to-noise ratios.
- Performance summary: All influenza and dengue models succeeded, while two of three tuberculosis models succeeded; cholera, Ebola, HIV/AIDS, and plague models failed.The authors found no relationship between r2 and either the number of articles or total traffic in the corresponding Wikipedia language edition.
- Transferability: The influenza Japan and Thailand models yielded a transferability score of rt = 0.81, suggesting that carefully selected source models may support contexts without training data.Transferability was evaluated by correlating article-level scores for shared articles across language editions.
6 Discussion
Wikipedia access logs are presented as an open data source that addresses key gaps in disease surveillance, with proof-of-concept evidence for breadth, transferability, and forecasting. The discussion also identifies substantial limitations and research directions before operational global deployment.
- C2. Breadth: 14 disease-location contexts were tested, with successful models in 8 contexts and r2 up to 0.92.These results support the feasibility of using Wikipedia data across multiple contexts.
- C3. Transferability: Model similarity reached 0.81 for some same-disease, different-location pairs, suggesting trained models may be transferable across locations.The proposed mechanism uses Wikipedia’s mappings between languages.
- C4. Forecasting: Forecasting value extended through the 28-day maximum tested horizon.Simple statistical models were used because traditional mechanistic models may require unavailable parameters.
- Limitations: The study requires testing in many more contexts before determining when and why these methods work.The authors identify broader evaluation as a central future research need.
- Limitations: The manual article-selection procedure used at most a few dozen candidates, whereas operational systems should evaluate thousands, millions, or all plausible articles.A comprehensive procedure could also support analysis of which article types predict disease incidence.
- Limitations: Language is an inherently weak location proxy and is implausible below country-level resolution.The authors propose hierarchical geographic aggregation of article traffic as a stronger alternative.
8 Supplementary information
Supplementary Data S1 contains input data, raw results, and additional figures supporting the paper’s analyses.
- Data S1 includes article mappings, figures, epidemiological inputs, correlation scores, Wikipedia input data, and file-format documentation.