Traditional public health surveillance relies principally on clinical diagnoses, laboratory confirmation, statutory notification, sentinel sites, mortality records and field investigation. These sources remain indispensable because they are linked to defined populations, case definitions and verification procedures. Their limitation is often timeliness: information must pass through several stages before a signal becomes visible at district, state or national level. Digital platforms, by contrast, generate large volumes of continuously updated information on symptoms, concerns, information seeking, behaviours and public narratives.

The use of such data has produced a family of overlapping concepts. Digital epidemiology broadly refers to the use of digital data for understanding health and disease. Infodemiology examines the distribution and determinants of health information in electronic media, whereas infoveillance applies these data to surveillance. Social-media surveillance commonly analyses posts, comments, hashtags, geolocation, images or interaction networks from platforms such as X, Facebook, Reddit, YouTube and public messaging channels. Search-trend surveillance uses aggregated search activity, most commonly Google Trends, as an indirect measure of population interest or concern. Open-source epidemic intelligence extends further to online news, official websites, reports and other publicly available sources.

The World Health Organization’s Epidemic Intelligence from Open Sources initiative illustrates the institutional maturation of this field. It combines large-scale automated screening of open-source material with expert assessment and verification rather than treating unverified digital mentions as cases.[1] This distinction is fundamental. A social-media spike is a signal requiring interpretation, not proof of an outbreak.

How Digital Surveillance Generates Public Health Signals

Digital surveillance systems usually begin with keyword selection, topic modelling or machine-learning classification. Posts or searches are filtered by disease names, symptoms, medicines, places or behavioural terms. Natural-language processing may distinguish personal illness reports from news sharing, jokes, advertisements or unrelated uses of the same word. Signals are then summarized by time and geography and compared with a baseline or conventional surveillance source.

Search data and social-media data represent different phenomena. A search query generally reflects information seeking, while a social-media post may reflect personal experience, news dissemination, opinion, advocacy, fear or misinformation. Neither directly measures disease incidence. Their usefulness depends on whether the online behaviour has a stable relationship with the health outcome of interest.

Digital signals can support several functions:

1. Early event detection, such as unusual reports of fever, gastrointestinal illness or unexplained deaths.
2. Nowcasting, or estimating current disease activity before official reports are complete.
3. Forecasting, in which digital indicators are incorporated into predictive models.
4. Situational awareness, including shortages, service disruption, risk perceptions and community concerns.
5. Infodemic surveillance, covering misinformation, information voids, mistrust and emerging questions.
6. Evaluation of public response, such as interest in vaccination, preventive measures or government guidance.

These functions should not be conflated. A system that successfully measures public attention may still perform poorly in predicting case counts. Conversely, a weak correlation with incidence may not eliminate its value for identifying misinformation or unmet communication needs.

Evidence for Timeliness and Predictive Utility

Early enthusiasm for internet-based surveillance was stimulated by evidence that search and social-media activity could track influenza and other infectious diseases. However, the failure of Google Flu Trends became a formative warning. Its estimates substantially over-predicted influenza activity during some periods because the model was vulnerable to changing search behaviour, media influence and algorithmic drift. Lazer and colleagues described this as a combination of “big data hubris” and insufficient integration with conventional data.[2] The lesson was not that digital data are useless, but that volume cannot compensate for weak epidemiological design.

More recent studies have adopted narrower outcomes, transparent models and combined data sources. A 2024 measles analysis across European countries and Japan found that Google Trends sometimes correlated strongly with notified cases, particularly for acute regional outbreaks. In Okinawa, regional correlation was considerably stronger than the corresponding national estimate. Yet the study also found weak signals during prolonged low-level transmission, sensitivity to keyword choice and poorer performance in settings with lower internet penetration.[3] Thus, validity was contextual rather than universal.

Indian evidence similarly demonstrates both potential and ambiguity. Verma and colleagues compared Google search trends with Integrated Disease Surveillance Programme data for malaria, dengue, chikungunya and enteric fever. Search patterns showed useful temporal associations for selected diseases, suggesting potential value as a supplementary indicator.[4] During COVID-19, Satpathy and colleagues found high time-lagged correlations between several search terms and reported testing or cases in India. However, peaks often coincided with intense media coverage or government announcements, making it difficult to separate disease-related information seeking from media-driven curiosity.[5]

This problem affects much of the literature. Retrospective correlations can appear impressive because investigators try multiple keywords, lags, geographical units and statistical models and then report the best-performing combination. Without prespecification and out-of-sample testing, this creates a substantial risk of overfitting.

Validity: Does the Digital Indicator Measure What It Claims to Measure?

Validity in digital surveillance has several dimensions.

Construct validity

Construct validity asks whether the digital measure represents the underlying concept. A search for “dengue symptoms” may indicate illness in the searcher, concern about a family member, examination preparation, media exposure or general curiosity. A post mentioning “depression” may represent a diagnosis, metaphor, humour, advocacy or news commentary. Classification algorithms reduce but cannot eliminate this ambiguity.

Systems should therefore define their intended construct precisely. “Online concern about dengue” is a more defensible construct than “dengue incidence” unless a stable relationship with confirmed cases has been demonstrated.

Criterion validity

Criterion validity concerns agreement with a credible reference standard, such as laboratory-confirmed cases, hospital admissions, prescriptions or mortality. Common metrics include correlation coefficients, sensitivity, specificity, positive predictive value, mean absolute error and lead time.

Correlation alone is inadequate. Two seasonal series may correlate simply because both rise during the monsoon or winter. A model can achieve high correlation while systematically overestimating magnitude or missing the beginning of an outbreak. Validation should therefore assess calibration, outbreak detection, false-alert frequency and incremental value beyond routine surveillance.

External validity

Models developed in one country, language or platform may not generalize elsewhere. Platform populations differ by age, education, occupation and socioeconomic status. Search behaviour also changes according to disease awareness and access to healthcare. In multilingual countries, English-language keywords may disproportionately represent urban and educated populations.

External validation requires testing across regions, time periods, languages and demographic groups. Model performance should be reported separately rather than averaged across heterogeneous settings.

Predictive and operational validity

Predictive validity asks whether the signal accurately anticipates future events. Operational validity goes further: does the signal improve a real public health decision? An alert arriving two days before routine notification has limited value if verification requires several weeks or no response protocol exists. The most meaningful outcomes are earlier investigation, faster risk communication, improved resource allocation or reduced morbidity.

Reliability and Reproducibility

Reliability refers to the consistency of a measure when repeated under comparable conditions. Digital platforms create distinctive reliability problems because investigators do not control the data-generating system.

Google Trends provides normalized relative search volume rather than absolute counts. The value 100 represents the peak popularity within the specified query, geography and period. Changing the time window, comparison terms or geographical level can alter the scale. Moreover, the platform may sample searches. Rovetta repeatedly downloaded identical Google Trends queries and found that values and correlations could vary according to the day of extraction, particularly for smaller geographical areas. The study recommended repeated data collection and averaging rather than reliance on a single download.[6]

Social-media datasets are also unstable. Platforms modify application programming interfaces, moderation policies, user demographics and recommendation algorithms. Historical posts may be deleted, accounts suspended and geolocation removed. Researchers may obtain only an unknown fraction of total content. Consequently, studies conducted at different times can produce different datasets even when nominally using the same search strategy.

Reproducibility is further weakened by inadequate reporting. Studies may omit the exact keywords, language variants, exclusions, data-access level, bot-detection procedures or preprocessing rules. Algorithms are often proprietary or described incompletely. At minimum, investigators should archive query specifications, extraction dates, code, model versions and aggregate datasets where ethically permissible.

Major Sources of Bias and Error

Media amplification

News reports, celebrity illness, policy announcements and viral posts can produce large digital peaks unrelated to changes in incidence. Media influence is not merely “noise” when the aim is communication surveillance, but it is a major confounder when estimating disease occurrence.

Selection and digital-divide bias

People producing digital data are not a random sample of the population. Older adults, poorer households, people with low literacy and residents of areas with weak connectivity may be underrepresented. In India, regional and gender disparities in digital access can lead to systematic blind spots. A high-volume urban signal may overshadow a clinically important rural outbreak.

Platform and algorithmic bias

Recommendation algorithms amplify some content and suppress other material. Engagement-driven systems favour emotionally charged or controversial posts. Apparent prevalence of an opinion may therefore reflect algorithmic visibility rather than population prevalence.

Bots, spam and coordinated manipulation

Automated accounts, advertising and coordinated campaigns can distort counts and sentiment. This is particularly relevant to vaccination, tobacco, political health controversies and misinformation. Bot detection helps but may misclassify legitimate highly active accounts.

Semantic and linguistic ambiguity

Symptoms and diseases may be expressed through slang, spelling variants, transliteration, emojis or local idioms. Indian surveillance requires analysis across many languages and mixed-language text. A model trained primarily on standard English will miss substantial content and may perform differently across communities.

Behavioural and policy changes

Testing availability, healthcare-seeking, case definitions and public awareness can change during an epidemic. The relationship between online activity and reported cases is therefore non-stationary. Models require continuous recalibration rather than one-time validation.

Public Health Significance

The principal advantage of digital surveillance is not necessarily greater accuracy but greater speed and breadth. It can reveal community concerns before individuals seek care, identify events in areas with weak formal reporting and monitor behavioural or informational dimensions that clinical systems do not capture.

During emergencies, social-media surveillance can identify shortages, barriers to services, rumours and questions requiring risk communication. Infodemic surveillance has become particularly important because misinformation and information overload may reduce adherence to public health recommendations and erode institutional trust.  WHO    and    collaborating   researchers   have   therefore

Table 1. Validity, reliability and recommended use of major digital surveillance sources

argued that infodemic surveillance should become a formal public health function supported by multidisciplinary teams.[7]

Digital signals can also help prioritize field investigation. A cluster of posts about vomiting after a community event may justify rapid verification even before laboratory results are available. The correct response, however, is investigation—not automatic classification of the posts as cases.

Global Institutional Developments

WHO’s EIOS framework represents a shift from isolated experimental tools towards collaborative public health intelligence. The system processes large volumes of open-source information, but expert analysts assess relevance, credibility and context before escalation. The 2025 EIOS strategy emphasizes integration of diverse sources, collaboration and fit-for-purpose technology rather than fully automated outbreak declaration.[1]

Public health agencies are also incorporating social listening into emergency communication. The future direction is increasingly multimodal: combining online reports with healthcare encounters, laboratory data, mobility, environmental information and community-based event surveillance. Such triangulation reduces dependence on a single volatile platform.

Large language models and transformer-based classifiers can improve multilingual topic identification and distinguish first-person symptom reports from general discussion. Yet their apparent accuracy may deteriorate when disease terminology, platform culture or user behaviour changes. Automated classification should therefore support rather than replace epidemiological judgment.

Indian Perspective

India presents a strong case for supplementary digital surveillance because of its large population, expanding internet use, high social-media activity and continuing variation in the completeness and timeliness of routine reporting. The Integrated Disease Surveillance Programme and its Integrated Health Information Platform provide the formal backbone for outbreak detection. Search and social-media signals could be incorporated as additional event-based inputs, particularly for unusual clusters, public concern and misinformation.

However, India’s diversity creates challenges rarely addressed in published models. Surveillance systems must recognize Hindi, English, regional scripts, transliterated language and mixed-language communication. Disease terminology may differ substantially between communities. Searches for “dengue,” for example, may coexist with local descriptions of fever or platelet reduction. Models relying only on formal disease names will have low sensitivity.

Digital surveillance may also amplify existing inequities. Internet-generated signals will be strongest in metropolitan and digitally connected groups, whereas outbreaks among tribal, remote or socioeconomically disadvantaged populations may remain invisible. Digital data should therefore never be used to reduce investment in community reporting, laboratory networks or field epidemiology.

A practical Indian model would route digital alerts to district surveillance units, which could triangulate them with outpatient syndromes, laboratory data, pharmacy sales, school absenteeism and community health-worker reports. Prospective evaluation should determine whether this improves detection beyond the existing system.

Challenges and Limitations

The literature remains dominated by retrospective studies. Many report statistically significant correlations but do not assess false alarms, missed outbreaks or real-world decision-making. Positive publication bias is likely because unsuccessful keywords and models are rarely reported.

Data access is becoming more restricted and commercially controlled. Public health agencies may depend on private companies whose priorities, algorithms and pricing can change without notice. This raises concerns about sustainability and accountability.

Privacy requires careful attention even when posts are publicly accessible. Users generally do not expect personal health statements to be collected, classified and mapped by surveillance systems. Reporting at aggregate level, data minimization, restricted retention and avoidance of attempts to re-identify individuals are essential. High-risk uses, including monitoring stigmatized conditions or small communities, require enhanced ethical scrutiny.

There is also a risk of “surveillance without response.” Dashboards may generate large numbers of signals but little action. Systems should be judged by whether they improve investigation and control, not by the volume of data processed.

Future Directions

Future research should adopt prospective, registered protocols with predefined keywords, outcomes, time lags and alert thresholds. Models should be evaluated during unseen outbreaks and compared with simple baselines using sensitivity, specificity, positive predictive value, calibration, lead time and false-alert burden.

Hybrid models are likely to outperform digital-only systems. Search trends, social posts, online news, clinical syndromes, laboratory results and environmental surveillance should be combined with weights that are transparent and periodically recalibrated.

Multilingual research is a priority, especially in India and other linguistically diverse settings. Community participation should guide terminology, interpretation and ethical boundaries. Models should be tested for differential performance by region, gender, age, language and socioeconomic context.

Public health agencies need dedicated social-listening and digital-epidemiology teams comprising epidemiologists, behavioural scientists, communication specialists, data scientists, ethicists and local health officers. Automated alerts should include an explanation of why the signal was generated and what evidence supports escalation.

Finally, digital surveillance requires common reporting standards. Authors should disclose data-access methods, sampling, extraction dates, query syntax, preprocessing, missing-data handling, bot detection, validation sources and model drift. Without such transparency, claims of real-time surveillance cannot be independently verified.

Social-media and search-trend surveillance can add valuable speed, reach and behavioural insight to public health intelligence. It is particularly useful for detecting unusual reports, monitoring public concern, identifying misinformation and supplementing delayed formal data. Its limitations are equally important: online signals are indirect, unrepresentative, algorithmically mediated and vulnerable to media influence, semantic ambiguity and temporal instability.

The available evidence does not support replacing conventional surveillance with digital traces. The strongest model is collaborative surveillance in which digital signals trigger verification and are interpreted alongside clinical, laboratory, environmental and community data. Validity must be established for each disease, location, language and purpose; reliability must be protected through repeated extraction, transparent methods and monitoring of platform changes.

For India, digital surveillance offers substantial promise but must be integrated cautiously with the Integrated Health Information Platform and district response structures. Multilingual capability, equity assessment, privacy protection and local verification are indispensable. The future of real-time public health surveillance lies not in treating social media as a diagnostic instrument, but in using it as one carefully governed sensor within a broader, responsive and evidence-based surveillance ecosystem.

1. World Health Organization. Epidemic Intelligence from Open Sources (EIOS) Strategy 2024–2026 [Internet]. Geneva: World Health Organization; 2025 [cited 2026 Aug 8]. Available from: https://www.who.int/publications/i/item/B09476. doi:10.2471/B09476.
2. Lazer D, Kennedy R, King G, Vespignani A. The parable of Google Flu: traps in big data analysis. Science. 2014;343(6176):1203-1205. doi:10.1126/science.1248506.
3. Wang D, Lang JC, Chen YH. Assessment of using Google Trends for real-time monitoring of infectious disease outbreaks: a measles case study. Sci Rep. 2024;14(1):9470. doi:10.1038/s41598-024-60120-8.
4. Verma M, Kishore K, Kumar M, Sondh AR, Aggarwal G, Kathirvel S. Google Search Trends predicting disease outbreaks: an analysis from India. Healthc Inform Res. 2018;24(4):300-308. doi:10.4258/hir.2018.24.4.300.
5. Satpathy P, Kumar S, Prasad P. Suitability of Google Trends™ for digital surveillance during ongoing COVID-19 epidemic: a case study from India. Disaster Med Public Health Prep. 2023;17:e28. doi:10.1017/dmp.2021.249.
6. Rovetta A. Reliability of Google Trends: analysis of the limits and potential of web infoveillance during COVID-19 pandemic and for future research. Front Res Metr Anal. 2021;6:670226. doi:10.3389/frma.2021.670226.
7. Purnat TD, Vacca P, Czerniak C, Ball S, Burzo S, Zecchin T, et al. Infodemic signal detection during the COVID-19 pandemic: development of a methodology for identifying potential information voids in online conversations. JMIR Infodemiology. 2021;1(1):e30971. doi:10.2196/30971.
8. Mendoza-Herrera K, Valero-Morales I, Ocampo-Granados ME, Reyes-Morales H, Arce-Amaré F, Barquera S. An overview of social media use in the field of public health nutrition: benefits, scope, limitations, and a Latin American experience. Prev Chronic Dis. 2020;17:E76. doi:10.5888/pcd17.200047.
9. Tangcharoensathien V, Calleja N, Nguyen T, Purnat T, D'Agostino M, Garcia-Saiso S, et al. Framework for managing the COVID-19 infodemic: methods and results of an online, crowdsourced WHO technical consultation. J Med Internet Res. 2020;22(6):e19659. doi:10.2196/19659.
10. Bernardo TM, Rajic A, Young I, Robiadek K, Pham MT, Funk JA. Scoping review on search queries and social media for disease surveillance: a chronology of innovation. J Med Internet Res. 2013;15(7):e147. doi:10.2196/jmir.2740