Conjunction of factors triggering waves of seasonal influenza

Ishanu Chattopadhyay, Emre Kıcıman, Joshua W. Elliott, Jeffrey Shaman, Andrey Rzhetsky

TL;DR

Seasonal influenza epidemics in the US aren't triggered by a single "smoking gun" factor, but by a specific "perfect storm" of socio-demographics, weather shifts, and viral evolution. By analyzing data from 150 million people and billions of tweets, researchers found that these waves consistently spark in the Southern US and propagate primarily through short-range land travel rather than long-distance flights.

What Problem Does This Paper Address?

Despite decades of study, predicting exactly when and where a seasonal flu wave will begin remains a challenge. Most research focuses on a single variable—like absolute humidity or school schedules—in isolation. However, infectious diseases operate in a "messy" real-world environment where factors interact.

The authors sought to move beyond the ceteris paribus ("all other things being equal") approach. They aimed to identify the "mechanistic triggers" of pan-continental epidemics by performing an integrative joint analysis of dozens of variables simultaneously, determining which factors actually drive the "spark" that turns local infections into a national wave.

Methods

The researchers employed three "orthogonal" statistical approaches to ensure their findings weren't just artifacts of a specific model:

  1. Non-parametric Granger Causality: A new, non-linear approach used to map "causality streamlines." This allowed them to visualize the directional flow of infection across the US and identify "sources" (where it starts) and "sinks" (where it converges).
  2. Mixed-effect Poisson Regression: A Bayesian model (using MCMC) to rank the predictive power of various factors while accounting for the inherent "memory" of the disease (autocorrelation) in time and space.
  3. County-matching Analysis: A non-parametric method that compared "treated" counties (those meeting specific trigger conditions) with "control" counties that were identical in every other way, effectively simulating a randomized trial from observational data.

Datasets

The study is notable for its massive scale, utilizing several longitudinal datasets:

  • Clinical Data: Insurance claims from the Truven MarketScan database covering over 150 million individuals over a decade.
  • Human Movement: Billions of geo-located tweets used to infer week-to-week travel patterns between every US county.
  • Weather: High-resolution data (PRISM and NLDAS-2) covering humidity, temperature, solar radiation, and precipitation.
  • Viral Genetics: Sequence variations of key influenza proteins (HA, NA, M1, M2) from the NIAID Influenza Research Database to track "antigenic drift."
  • Demographics: US Census data for socioeconomic status, ethnicity, and population density.
  • Vaccination: Electronic health records (EHR) to track cumulative vaccination coverage.

Key Insights and Ideas

  • The Southern Cradle: Epidemics consistently originate in the South and Southeast US, particularly near coastal areas. These regions act as the "spark" for the rest of the country.
  • The "Perfect Storm" Trigger: Initiation isn't just about "cold weather." It’s a conjunction of high average humidity/temperature followed by a sharp drop in both. This stressor, combined with specific socio-demographic profiles, creates the ideal environment for a flare-up.
  • Land Travel Trumps Air Travel: While airports help seed the virus, short-range land-based travel (commuting, regional trips) is the primary engine for the actual propagation of the epidemic wave.
  • The School Schedule Myth: Contrary to popular belief, the timing of children returning to school after summer or winter breaks was not a significant predictor of the initial seasonal trigger in this model.
  • Demographics are Dominant: Surprisingly, the host population’s socio- and ethno-demographic properties were the strongest predictors of flu onset, even outranking weather variables.

Key Results

  • Predictive Power: The models achieved an AUC (Area Under the Curve) of approximately 80% when tested against out-of-sample data, meaning they could accurately identify which counties would experience a flare-up.
  • Factor Ranking: The strongest predictors, in order, were: (1) Socio-demographics, (2) Weather variables (humidity/temp/solar), (3) Viral antigenic drift, (4) Land-based travel, and (5) Spatio-temporal autocorrelation.
  • Viral Protein Dynamics: Diversity in the M2 protein was found to enhance the epidemic trigger, while diversity in HA, NA, and M1 proteins actually appeared to inhibit it.

Limitations and Future Work

The authors acknowledge a few key limitations:

  • Twitter Bias: Movement data from Twitter tends to skew toward younger, more urban, and higher-income populations.
  • Genetic Resolution: Viral sequence data was available at a coarser resolution (state-level and yearly) compared to the highly granular county-level clinical data.
  • Observational Nature: While the three-pronged statistical approach bolsters causal claims, it remains an observational study rather than a controlled experiment.

Future Work: The authors suggest these models could be used to formulate new population-level intervention policies, such as targeted regional mitigation strategies that focus on the Southern "trigger zones" before the virus sweeps northward.

Summary updated

Abstract

Using several longitudinal datasets describing putative factors affecting influenza incidence and clinical data on the disease and health status of over 150 million human subjects observed over a decade, we investigated the source and the mechanistic triggers of influenza epidemics. We conclude that the initiation of a pan-continental influenza wave emerges from the simultaneous realization of a complex set of conditions. The strongest predictor groups are as follows, ranked by importance: (1) the host population's socio- and ethno-demographic properties; (2) weather variables pertaining to specific humidity, temperature, and solar radiation; (3) the virus' antigenic drift over time; (4) the host population'€™s land-based travel habits, and; (5) recent spatio-temporal dynamics, as reflected in the influenza wave auto-correlation. The models we infer are demonstrably predictive (area under the Receiver Operating Characteristic curve 80%) when tested with out-of-sample data, opening the door to the potential formulation of new population-level intervention and mitigation policies.
BibTeX
@inproceedings{chattopadhyay2018conjunction,
  author = {Ishanu Chattopadhyay and Emre Kıcıman and Joshua W. Elliott and Jeffrey Shaman and Andrey Rzhetsky},
  title = {{Conjunction of factors triggering waves of seasonal influenza}},
  booktitle = {eLife},
  year = {2018},
  doi = {10.7554/elife.30756},
}
Figure