Estimating the secondary attack rate and serial interval of influenza‐like illnesses using social media

Elad Yom‐Tov, Ingemar J. Cox, Vasileios Lampos, Andrew HaywardInfluenza 2015 · DOI

TL;DR

This paper introduces a novel method for estimating the secondary attack rate (SAR) and serial interval (SI) of influenza-like illnesses (ILI) by analyzing social media data, specifically Twitter. By leveraging public posts, the authors demonstrate a way to monitor disease transmission dynamics more rapidly and cost-effectively than traditional methods, offering a valuable tool for public health surveillance.

What Problem Does This Paper Address?

Understanding how infectious diseases like influenza spread is crucial for effective public health interventions. Two key metrics for this are the secondary attack rate (SAR), which measures how likely infection is to spread within a susceptible population after initial exposure, and the serial interval (SI), the time between symptom onset in an infected individual and symptom onset in a newly infected individual.

Traditionally, estimating SAR and SI requires extensive surveys and contact tracing, which are resource-intensive and often lag behind real-time disease activity. Furthermore, a significant portion of ILI cases don't seek medical attention, making official surveillance data incomplete. This gap highlights the need for more agile and comprehensive methods to track disease transmission dynamics, especially for early detection and response to outbreaks.

Methods

The core of this research lies in a two-stage filtering process applied to Twitter data to identify relevant ILI-related posts.

  1. Identifying Illness Tweets:
    • Initial Filtering: The process begins by creating a large list of terms (n-grams) associated with ILI symptoms by browsing health websites and Wikipedia. This list is refined by analyzing co-occurring terms in a large Twitter corpus.
  2. Predictive Modeling: A linear prediction model (ridge regression) is trained to correlate these terms with official ILI rates from the Royal College of General Practitioners (RCGP) and Public Health England (PHE). The top 20 most predictive phrases are then used to filter tweets.
  3. Second Filter (User/Family Focus): A crucial step involves using a trained Support Vector Machine (SVM) classifier to distinguish tweets that refer to the poster's own ILI symptoms from those that mention a family member's ILI. This is achieved by analyzing tweet features like keywords, emoticons, replies, and hashtags, trained on a manually labeled dataset.
  4. Estimating Familial Secondary Attack Rate (fSAR):
    • The authors define fSAR as the probability that at least one household contact becomes a secondary case given an initial case.
  5. They make key assumptions: ILI sufferers tweet about their illness, and if a family member also gets sick, there's a higher probability the user will tweet about it. They also assume that if two family members are ill, it's more likely due to transmission within the household than independent infection, with a 7-day window for symptom onset.
  6. fSAR is estimated using a weighted linear regression model, where weights are derived from the SVM classifier's confidence in identifying family-related ILI tweets.
  7. Estimating Familial Serial Interval (fSI):
    • fSI is calculated as the average time difference between a user reporting their own ILI symptoms and reporting a family member's ILI symptoms, for cases where both are reported within the 7-day window.

Datasets

  • Twitter Data: The study utilized a massive dataset of tweets originating from England during two flu seasons: October 1, 2012, to April 30, 2013, and October 1, 2013, to April 30, 2014. This included approximately 80.9 million tweets from over 883,000 users for the 2012 season and 133.5 million tweets from over 1.2 million users for the 2013 season.
  • Survey Data: A survey of 93 self-reported Twitter users was conducted via the CrowdFlower platform to validate assumptions about user behavior regarding tweeting about family members' illnesses.
  • Official ILI Data: The study used ILI rates reported by the Royal College of General Practitioners (RCGP) and Public Health England (PHE) for model training and validation. Virologically confirmed influenza data from the PHE DataMart system was also used for comparison.

Key Insights and Ideas

  • Social Media as a Proxy for Disease Dynamics: The paper convincingly demonstrates that public social media posts can serve as a valuable, near real-time indicator of disease transmission dynamics, complementing traditional surveillance.
  • Predictive Power of Social Media: The method successfully correlates social media activity with official ILI rates, suggesting that public discourse on platforms like Twitter reflects actual disease burden.
  • Early Warning Signal: The study found that increases in estimated fSAR often preceded increases in physician-reported ILI rates, suggesting social media surveillance could provide an early warning.
  • Strain-Specific Transmission: The fSAR varied between influenza seasons and appeared to correlate with the dominant circulating strains (e.g., higher fSAR for Influenza B in 2012), hinting at differences in transmissibility or population immunity.
  • Consistent Serial Interval: While fSAR varied, the estimated familial serial interval remained relatively consistent across seasons, suggesting a stable incubation period for ILI in this context.

Key Results

  • Classifier Performance: The SVM classifier used to distinguish personal from family ILI tweets achieved an Area Under the ROC Curve (AUC) of 0.84, indicating good accuracy.
  • fSAR Estimates:
    • 2012 season: 30.5% (SE 1.3%)
    • 2013 season: 25.7% (SE 0.8%)
    • The fSAR for the 2013 season was approximately 23% lower than in 2012, aligning with lower overall ILI rates reported by physicians.
  • fSI Estimates:
    • 2012 season: 2.41 days
    • 2013 season: 2.48 days
    • These estimates are consistent with findings from prospective studies.
  • Correlation with Viral Strains: Peaks in fSAR were observed to coincide with peaks in specific influenza strains (Influenza B and A), suggesting a link between viral characteristics and transmission rates.

Limitations and Future Work

  • Demographic Bias: The authors acknowledge that Twitter users are not representative of the general population, particularly regarding age (fewer users under 18). This could bias fSAR estimates if certain strains disproportionately affect specific age groups.
  • Index Case Identification: The method cannot definitively identify the "index case" within a household, which could lead to overestimation of fSAR if multiple independent infections occur simultaneously.
  • Future Directions:
    • Utilize other Internet data sources (search engines, other social media) to validate and improve fSAR estimates.
  • Implement targeted advertisements to invite users reporting ILI to provide samples for virological testing, thus obtaining ground-truth data.
  • Further research is needed to understand factors influencing fSI and their interactions with virological profiles.

Summary updated

Abstract

The proposed method can assist health authorities by providing near-real-time estimation of SAR and SI, and especially in alerting to sudden increases thereof.
BibTeX
@inproceedings{yomtov2015estimating,
  author = {Elad Yom‐Tov and Ingemar J. Cox and Vasileios Lampos and Andrew Hayward},
  title = {{Estimating the secondary attack rate and serial interval of influenza‐like illnesses using social media}},
  booktitle = {Influenza},
  year = {2015},
  doi = {10.1111/irv.12321},
}
Figure