Using an evolutionary epidemiological model of pandemics to estimate the infection fatality ratio for humans infected with avian influenza viruses

This paper employs an evolutionary epidemiological model to estimate that thousands of undetected human avian influenza infections occur annually with an infection fatality ratio of approximately 0.32%, highlighting the critical need to prevent animal-to-human spillovers to mitigate severe individual outcomes and delay future pandemics.

The Gist

The Big Picture: Counting the Invisible

Imagine a massive, invisible storm of bird flu viruses swirling around the world. Most of the time, this storm hits humans but doesn't stick; it's like a gentle rain that wets your shirt but doesn't soak through. Sometimes, however, a single drop of rain hits a person, and that person gets sick. Even rarer, that person might pass a "super-charged" version of the virus to others, potentially starting a global pandemic.

The problem is that we can only see the people who get very sick and go to the hospital. We can't see the thousands of people who might have caught the bird flu, felt a little sniffle, or had no symptoms at all, and then went about their day.

This paper is like a detective trying to count the invisible drops of rain. The authors used a mathematical model to figure out:

1. How many people are actually getting infected by bird flu every year (even if they don't know it)?
2. How dangerous is it really for a human to catch it?
3. If we stop some of these infections, how much time do we buy before the next global pandemic?

---

1. The "Hidden Iceberg" of Infections

The authors looked at the history of flu pandemics over the last 245 years. They know that pandemics happen, but they don't know exactly how many "practice runs" (infections that didn't become pandemics) happen before one succeeds.

The Analogy: Imagine a lottery. Every time a bird flu jumps from a chicken to a human, it's like buying a lottery ticket. Most tickets are losers (the virus dies out). But occasionally, a ticket wins (the virus adapts and starts spreading between humans, causing a pandemic).

The authors asked: "If we know how often the jackpot (pandemic) is hit, how many tickets must have been sold (infections) to get there?"

The Discovery: They found that for every one person who gets sick enough to be reported, there are likely thousands of people who are infected but never show up on the radar. They estimate that roughly 6,400 people get infected with bird flu globally every year, even though we only officially report a tiny fraction of that number.

2. How Deadly is the "Invisible" Virus?

Usually, when we hear about bird flu, we hear a scary statistic: "48% of reported cases die." That sounds terrifying. But that number is misleading because it only counts the people who were sick enough to be tested.

The Analogy: Imagine a video game where you only count the players who lose all their lives. If you only look at the "dead" players, the game looks impossible. But if you look at everyone who played, including the ones who just lost a few points and kept going, the game is actually much easier.

The authors calculated the Infection Fatality Ratio (IFR). This is the real risk: If you catch this virus, what are the odds you will die?

The Result:

• Reported Cases: Look like a 48% death rate (very scary).
• Real Risk (IFR): Is about 0.32% (32 deaths per 10,000 infections).

Is 0.32% safe? Not really.

• It is about 5 to 10 times deadlier than the seasonal flu we get every winter.
• It is roughly twice as deadly as the original SARS-CoV-2 virus was during the early days of the pandemic.

So, while it's not a 50/50 coin flip, catching bird flu is still a very dangerous gamble.

3. The "Time Bomb" and How to Defuse It

The paper also asks: "What happens if we stop people from catching bird flu from animals?"

The Analogy: Imagine a giant clock ticking down to a pandemic explosion. Every time a bird flu jumps from a cow or a duck to a human, the clock ticks forward a little bit. The more jumps we have, the faster the clock reaches zero.

The authors simulated what would happen if we stopped some of these jumps:

• If we stop 5% of the jumps, we delay the pandemic by about 2 years.
• If we stop 20% of the jumps, we delay it by 9.4 years.
• If we stop 50% of the jumps, we delay it by 37.5 years.

The Takeaway: Every single time we prevent a farmer, a vet, or a hunter from catching the virus from an animal, we are effectively hitting the "pause" button on the pandemic clock. It buys the world decades of time to prepare, develop vaccines, and improve our defenses.

4. Who Needs to Listen?

This isn't just a problem for scientists in labs. The paper highlights that people who work closely with animals are the ones most at risk.

• Farmers, poultry workers, hunters, trappers, and veterinarians are the "frontline soldiers" in this invisible war.
• Because the virus can be caught without showing symptoms, these workers might be carrying it without knowing.

The Advice:

• Don't touch sick or dead birds/animals.
• If you must handle them, wear gloves and a good mask.
• Report sick animals immediately.

Summary

This paper uses math to look behind the curtain of bird flu. It tells us that:

1. There are many more infections than we think.
2. The virus is dangerous (much worse than seasonal flu), even if it doesn't kill everyone.
3. Stopping animal-to-human jumps is the best defense. It saves individual lives and buys the world decades of safety before the next pandemic hits.

Think of it as a warning system: The virus is already in the room, and we are playing a game of "how many times can we catch it before it evolves to break out?" The more we catch it early and stop it, the longer we keep the door locked.

Technical Summary

1. Problem Statement

The risk posed by Highly Pathogenic Avian Influenza (HPAI) viruses to humans is difficult to quantify due to significant surveillance gaps. Most human infections are undetected because they may be asymptomatic, symptomatic but untested, or difficult to trace given the rarity of human-to-human transmission. Consequently, the Case Fatality Ratio (CFR)—calculated from reported cases (e.g., ~48% for H5N1)—likely overestimates the true risk because it ignores the "dark figure" of undetected infections.

The authors aim to address three specific objectives:

1. Estimate the true annual number of human AIV infections (including undetected cases).
2. Calculate the Infection Fatality Ratio (IFR) (deaths per total infection) rather than the CFR.
3. Determine how preventing zoonotic spillovers (animal-to-human transmission) impacts the timing of future influenza pandemics.

2. Methodology

The study employs a stochastic evolutionary epidemiological model based on a multitype branching process, adapting the framework of Day et al. (2006) but focusing on the mean number of infections rather than just the minimum.

Model Formulation

• Branching Process: The model simulates human infections with different AIV genotypes ($i$) circulating in animal populations.
• Zoonotic Spillovers ($N_z$): The annual number of spillovers is modeled as a random variable with mean $\bar{n}_z$.
• Basic Reproduction Number ($R_{0,i}$): Represents the transmissibility of a non-adapted AIV strain in humans. It is constrained to be $<1$ (no sustained human-to-human transmission prior to adaptation).
• Total Infections ($\bar{n}_h$): The mean annual human infections are derived by summing the initial spillovers and the subsequent limited human-to-human chains before the virus dies out.
  $$\bar{n}_h = \bar{n}_z \times \bar{m}$$
  Where $\bar{m}$ is the mean number of secondary infections generated per spillover across all genotypes.
• Pandemic Probability ($\Lambda$): The model calculates the probability that at least one spillover event leads to a pandemic-capable strain (via mutation or reassortment). This is equated to the historical frequency of pandemics.
  $$\bar{\Lambda} = 1 - \sum P(N_z = n_z) (1 - \bar{\pi})^{n_z}$$
  Where $\bar{\pi}$ is the probability that a single infection triggers a pandemic, dependent on the mutation probability ($a_i$) and the reproductive number of the adapted strain ($R^*_i$).

Parameterization and Data

• Historical Data: The model is calibrated using the inter-pandemic periods of seven influenza pandemics over the last 245 years (since 1781).
• Sampling: The authors used Latin Hypercube Sampling (LHS) to generate 1,000 genotype scenarios, sampling distributions for $R_0$, $R^*$, and mutation probabilities ($a_i$).
• Gamma Distribution: A gamma distribution was fitted to the historical inter-pandemic intervals to estimate the mean annual probability of a pandemic. Two scenarios were tested:
  1. Constant inter-pandemic period (no recency effect).
  2. Decreasing inter-pandemic period (pandemics becoming more frequent due to factors like deforestation and urbanization).
• IFR Calculation: The IFR was calculated by dividing the average annual reported human AIV deaths (20.6 deaths/year, based on 473 deaths over 23 years) by the estimated mean annual infections ($\bar{n}_h$).

3. Key Results

Estimated Annual Infections

• Baseline Scenario (No Recency Effect): The model estimates 6,441 annual human AIV infections worldwide (95% CI: 2,736 – 21,407).
• Recency Effect Scenario (2026): Assuming pandemics are becoming more frequent (inter-pandemic period decreasing to ~18.5 years by 2026), the estimated annual infections rise to 13,136 (95% CI: 6,214 – 36,230).
• Distribution: The distribution of estimated infections is highly right-skewed, indicating that while the median is ~6,000–13,000, there is a non-negligible probability of much higher infection numbers.

Infection Fatality Ratio (IFR)

• Baseline IFR: 32 deaths per 10,000 infections (0.32%; 95% CI: 9.6–75).
• Recency Effect IFR (2026): 16 deaths per 10,000 infections (0.16%; 95% CI: 5.6–33).
• Comparison:
  • SARS-CoV-2 (2020-2023): ~16 deaths per 10,000 infections.
  • Seasonal Influenza: ~2.9–6.5 deaths per 10,000 infections.
  • Reported CFR (H5N1): ~48% (4,800 per 10,000).
  • Conclusion: The true IFR is significantly lower than the reported CFR (due to undetected mild cases) but remains comparable to SARS-CoV-2 and significantly higher than seasonal flu.

Impact of Spillover Prevention

The study modeled the delay in pandemic emergence if zoonotic spillovers were reduced by a fraction $\chi$:

• 5% reduction: Delays pandemic by ~2 years.
• 20% reduction: Delays pandemic by ~9.4 years.
• 50% reduction: Delays pandemic by ~37.5 years.

4. Key Contributions

1. Quantifying Undetected Infections: The study provides a mathematical framework to infer the "dark figure" of AIV infections, estimating that thousands of undetected human infections occur annually, far exceeding the ~986 reported cases in the last two decades.
2. Re-evaluating Risk Metrics: By shifting focus from CFR to IFR, the paper corrects the perception of AIV risk. While the risk of death per reported case is extreme, the risk per total infection is lower but still substantial (comparable to SARS-CoV-2).
3. Evolutionary-Epidemiological Link: The model successfully links the historical rate of pandemics to current spillover rates, demonstrating that preventing spillovers has a dual benefit: protecting individuals from severe disease and delaying the emergence of a pandemic.
4. Validation of Minimum Estimates: The authors' estimates of the minimum number of spillovers align with previous studies (Day et al., 2006), but their estimation of the mean reveals a much higher burden of infection.

5. Significance and Implications

• Public Health Policy: The results justify strict biosecurity measures for high-risk groups (farmers, veterinarians, wildlife workers). Even if the IFR is lower than the CFR, the absolute risk of severe outcomes is high enough to warrant preventative measures (PPE, reporting infected animals).
• Pandemic Preparedness: The finding that preventing just 20% of spillovers could delay a pandemic by nearly a decade highlights the value of "One Health" interventions (preventing animal-to-human transmission) as a primary pandemic mitigation strategy.
• Surveillance Gaps: The discrepancy between reported cases and model estimates underscores the critical need for serological surveillance to detect asymptomatic or mild AIV infections, which are currently invisible to standard diagnostic systems.

Limitations

The authors acknowledge that the model relies on historical pandemic data (a small sample size of 7 events) and assumes constant probabilities for evolutionary change, which may not reflect real-world variability (e.g., co-infection rates). Additionally, the model does not account for seasonal dynamics or spatial heterogeneity, treating the estimates as global averages.