What is the Friedman H-statistic in this avian influenza study?

The Friedman H-statistic is a number that shows how much of a model's predictions come from interactions between two or more variables rather than from each variable on its own. In this study it ranged from 0.30 for wild birds to 0.53 for HPAI. Values above 0.10 are considered meaningful, and the HPAI score of 0.53 means most of that model's signal is interactive.

What drives avian influenza outbreak risk across continents

Q: What is the difference between HPAI and LPAI?

HPAI means highly pathogenic avian influenza and LPAI means low-pathogenic avian influenza. The difference is how severe the virus is in birds. HPAI causes serious disease and high death rates in poultry, while LPAI usually causes mild signs or none at all. Both move through wild and domestic birds, and the continental study modeled them as separate groups.

Q: Which factors predict avian influenza risk in wild birds?

For wild birds, land cover and climate predicted risk better than bird or human density. The vegetation index called NDVI and closeness to Ramsar wetlands were the top signals. Risk was highest at low elevation, with a small daily temperature swing of about 0 to 10 degrees Celsius, and near several wetlands. Wild bird species richness added risk through interactions with duck density.

Q: Why is poultry density the main driver of farm outbreaks?

Domestic duck and chicken densities were the leading predictors for low-pathogenic and domestic poultry outbreaks because crowded flocks give a virus room to establish and persist. Even small numbers of ducks or people raised the probability. The biggest jumps came from combinations: over 25,000 people with 1.5 million ducks and chickens, or duck counts above 150,000 birds.

Q: What is the strongest predictor of highly pathogenic avian influenza?

The clearest driver of where HPAI appears was closeness to places where low-pathogenic outbreaks had already happened. This does not mean mild strains turn deadly on site. It points to shared reservoir hosts, such as migratory ducks and gulls, and shared landscapes where both virus types have circulated together for two decades, making the LPAI pattern a reliable map of the overlap.

Q: How accurate were the machine learning risk models?

The models performed well across every subset. Accuracy ran from 84 percent for the simplest logistic regression up to 90 to 95 percent for the best random forest or gradient boosting model on each subset. Matthew correlation coefficient stayed above 0.8 and the calibrated area under the curve passed 0.9, which indicates balanced predictions with little geographic bias.

Q: Why do variable interactions matter more than single factors?

Friedman H-statistic measured how much of each model came from variables acting together rather than alone. For HPAI the score reached 0.53, so more than half the signal came from interactions. The all-AIV, LPAI, domestic, and wild models scored between 0.30 and 0.40. Ignoring these joint effects hides the real ecology, since risk spikes only when factors such as high duck and high human density meet near wetlands.

Q: Which regions are most at risk for avian influenza?

High-risk zones stretched across Asia from east to west, through southern and western Europe, and into northern Mediterranean, South African, and West African areas. The models also exposed likely underreporting in northern Algeria, Tunisia, Morocco, Lebanon, Iraq, the United Arab Emirates, and Oman, where predicted risk passed 0.5 yet few outbreaks were recorded.

Q: Do wetlands always increase outbreak risk?

Not for wild birds. Across much of the range, more Ramsar wetlands correlated with fewer reported wild-bird outbreaks, probably because protected remote wetlands keep waterfowl away from poultry and get weaker surveillance. For general and HPAI risk, though, more nearby wetlands combined with high duck density raised risk, so the effect depends on which host group you study.

Q: Can these models forecast future avian influenza outbreaks?

No. The models pinpoint areas whose ecological and host conditions favor virus circulation, not specific future events. They support risk-based surveillance and intervention planning rather than predicting timing or size. They still help by ranking where to sample, where to strengthen early warning, and where to send resources among underreported high-risk zones.

Avian influenza outbreak risk across Eurasia and Africa comes more from how host and environmental variables interact than from any single cause. Machine learning models built on more than 50,000 outbreak records from 2004 to 2025 reached at least 84 percent accuracy. The strongest drivers shift by virus type and host: wetland and vegetation signals lead for wild birds, poultry density leads for farms, and prior low-pathogenic activity is the clearest flag for the deadliest strains.

Why avian influenza risk mapping matters now

Avian influenza viruses keep finding new routes into poultry, wild birds, and mammals, including the recent jumps into dairy cattle reported in the United States. The cost is not abstract. The poultry sector supplies close to a fifth of the world’s animal protein, and a single regional epidemic in Southeast Asia around 2005 produced an estimated 10 billion US dollars in losses. Wild bird die-offs from the same viruses threaten biodiversity through mass mortality and species decline.

Outbreaks cluster unevenly and keep shifting. Some regions report constantly while others stay quiet, not because the virus is absent but because surveillance is thin. Knowing where outbreaks are biologically likely, apart from reporting effort, is the foundation for surveillance that spends less and catches more. A continental risk map tries to deliver exactly that.

How the risk models were built

The analysis combined two decades of officially reported outbreaks with 27 ecological and host features. Those features covered climate, elevation, land cover, vegetation index, human population density, domestic chicken and duck densities, Ramsar wetland locations, and wild bird species richness. Records came from the two main global animal health surveillance systems, then were cross-checked to remove duplicates.

After harmonizing the data, the team thinned outbreak points to a 10 km grid so that clustered reports in one place would not inflate model accuracy. That left 37,272 unique outbreak locations, about 68 percent of the raw 54,463 records. Four machine learning algorithms were compared: random forest, extreme gradient boosting, support vector machines, and logistic regression. Each was trained and tested on four subsets of the data: highly pathogenic avian influenza (HPAI), low-pathogenic avian influenza (LPAI), domestic poultry, and wild birds.

Element	Detail
Outbreak window	January 2004 to October 2025
Regions covered	35 African, 41 Asian, and 42 European nations
Raw records collected	54,463
Unique locations after thinning	37,272 (68.4 percent of raw)
HPAI share of locations	35,094 (93.4 percent)
LPAI share of locations	2,491 (6.6 percent)
Domestic poultry share	25,257 (67.8 percent)
Wild bird share	12,015 (32.2 percent)
Features used	27 ecological and host variables
Algorithms compared	Random forest, XGB, SVM, logistic regression
Best accuracy by subset	90 to 95 percent

Validation used spatial block cross-validation with 150 km blocks, which tests the model on geographically separate territory rather than nearby points. A 5 to 1 ratio of pseudo-absence to presence points gave the most stable results. To check robustness, the team withheld half the outbreaks in Western Europe and measured how closely the reduced model matched the full one.

What the models found

All four algorithms performed well, but the tree-based methods led. Random forest and extreme gradient boosting beat support vector machines and logistic regression on every metric. Accuracy ranged from 84 percent for the simplest model up to between 90 and 95 percent for the best model on each subset. Matthew’s correlation coefficient exceeded 0.8, and the calibrated area under the curve passed 0.9, which signals balanced predictions with little spatial sorting bias.

The main result is that interactions between predictors explained far more than any single variable. Friedman’s H-statistic measures how much of a model’s predictions come from variables acting together rather than alone. For HPAI, that value reached 0.53. In plain terms, more than half the predictive signal for the deadliest strains comes from variable interactions. The all-AIV, LPAI, domestic, and wild-bird models showed interaction strengths between 0.30 and 0.40.

The dominant drivers broke down cleanly by subset. For all-AIV and wild birds, the vegetation index (NDVI) and proximity to Ramsar wetlands dominated, with risk peaking at moderate vegetation values between 0.25 and 0.50 and near clusters of wetlands, roughly 10 sites per 100 square km. For LPAI and domestic poultry, domestic duck and chicken densities were the top predictors, and very low thresholds already raised risk. For HPAI, the single strongest driver was proximity to where low-pathogenic outbreaks had already occurred.

The two-way interactions that mattered most tell the real story. Duck density combined with human density pushed risk sharply upward at sites with more than 10,000 domestic ducks and over 15,000 people per 100 square km, and near more than 5 wetlands within a 10 km radius. For HPAI, locations with more than 40 prior low-pathogenic outbreaks and at least 5 nearby wetlands, or with over 10,000 ducks and more than 200 wild bird species, carried the highest risk. For LPAI, combinations of over 25,000 people with 1.5 million domestic ducks and chickens, or duck densities above 150,000 birds with moderate to high precipitation seasonality (40 to 120), drove risk highest. For domestic birds, isothermality values between 40 and 60 interacted with duck density and wild bird species richness. For wild birds, low elevation, a small daily temperature range (about 0 to 10 degrees Celsius), and proximity to multiple wetlands defined the highest-risk zones.

The resulting maps flagged high-risk areas across Asia from east to west, southern and western Europe, and northern Mediterranean, South African, and West African zones. They also pointed to likely underreporting. Places such as northern Algeria, Tunisia, Morocco, Lebanon, Iraq, the United Arab Emirates, and Oman showed model-predicted risk above 0.5 with few recorded outbreaks. The robustness check held up: after dropping half of Western Europe’s data, the spatial correlation with the full model was 0.91, with accuracy still at 93 percent.

The findings that run against intuition

Three results contradict common assumptions, and they matter for anyone planning surveillance.

First, more wetlands near a site did not always mean more wild-bird outbreaks. Across much of the range, a higher count of Ramsar wetlands was linked to lower reported wild-bird risk. Protected, remote wetlands may keep wild waterfowl inside natural habitat and away from poultry, while also receiving weaker passive surveillance. Both ecology and observation bias point the same way.

Second, the strongest predictor of where the deadly HPAI strain appears is not poultry density but the prior footprint of the milder LPAI strain. This is not evidence that mild strains mutate into deadly ones on the spot. It reflects shared reservoir hosts, such as migratory dabbling ducks and gulls, and shared landscapes where both pathotypes have co-circulated for two decades. The LPAI layer acts as a long-term map of the interface, not a real-time trigger.

Third, wind and elevation, often cited as local risk factors, showed almost no predictive power at the continental scale. Their influence matters farm to farm and within a country, but it washes out when you model an entire continent. Macro-ecological drivers simply dominate at that scale.

Practical recommendations for risk-based surveillance

Prioritize the overlap of duck density, human density, and wetland proximity. Sites with over 10,000 ducks and 15,000 people near multiple wetlands deserve the first round of sampling.
Treat LPAI occurrence as an early-warning signal for HPAI, especially where both have historically co-circulated, rather than waiting for the deadly strain to appear.
Strengthen wild-bird and environmental sampling in underreporting regions such as North Africa and the Gulf, instead of reading low case counts as low risk.
Focus fringe monitoring on moderate vegetation zones, not pristine wilderness or dense city centers, because transitional agricultural-wetland interfaces attract both wild and domestic birds.
Build early-warning tools that fuse outbreak data with meteorological and wetland monitoring under a One Health framework spanning animal, human, and environmental health.

Limitations to keep in mind

The data leans heavily toward HPAI and commercial poultry because passive surveillance catches those best. Underreporting in wildlife and developing regions limits accuracy, and the models identified several high-risk zones with no recorded outbreaks. Some predictor layers, such as duck density estimates, are static baselines that may miss recent change, with African estimates carrying higher uncertainty than Asian or European ones.

Migratory routes and farm biosecurity practices were not included because they are hard to collect comparably across countries. The use of artificial pseudo-absence points is a standard but imperfect choice, and the framework does not quantify prediction uncertainty because it avoids heavy Bayesian computation. Most importantly, the models identify where conditions favor virus circulation. They do not forecast specific future outbreaks.

Bottom line

Continental avian influenza risk is an interaction problem, not a single-cause problem. Wetland and vegetation signals lead for wild birds, poultry density leads for farms, and historical low-pathogenic activity is the clearest flag for the deadliest strain. Surveillance budgets should follow those overlaps, with extra weight on regions where the model says risk is high but the reports say otherwise.

Frequently asked questions

What is the difference between HPAI and LPAI?

HPAI stands for highly pathogenic avian influenza and LPAI for low-pathogenic avian influenza. The distinction is virulence in birds. HPAI causes severe disease and high mortality in poultry, while LPAI usually causes mild or no signs. Both spread through wild and domestic birds, and both were analyzed separately in the continental risk study.

Which factors predict avian influenza risk in wild birds?

For wild birds, the strongest predictors were land cover and climate rather than host density. The vegetation index (NDVI) and proximity to Ramsar wetlands led, with risk highest at low elevation, a small daily temperature range of about 0 to 10 degrees Celsius, and near multiple wetlands. Wild bird species richness also amplified risk through interactions with duck density.

Why is poultry density the main driver of farm outbreaks?

Domestic duck and chicken densities were the top predictors for low-pathogenic and domestic poultry outbreaks because dense flocks let a virus establish and persist. Even small thresholds of duck or human density raised probability. Interactions mattered most: sites with over 25,000 people and 1.5 million ducks and chickens, or duck densities above 150,000 birds, carried the highest risk.

What is the strongest predictor of highly pathogenic avian influenza?

The single strongest driver of HPAI spatial risk was proximity to where low-pathogenic outbreaks had already occurred. This is not evidence of on-site mutation. It reflects shared reservoir hosts and shared landscapes where both pathotypes have co-circulated for two decades, making the LPAI footprint a reliable map of the interface.

How accurate were the machine learning risk models?

The models performed well across all subsets. Accuracy ranged from 84 percent for the simplest logistic regression up to 90 to 95 percent for the best random forest or extreme gradient boosting model on each subset. Matthew’s correlation coefficient exceeded 0.8 and the calibrated area under the curve passed 0.9, indicating balanced, low-bias predictions.

Why do variable interactions matter more than single factors?

Friedman’s H-statistic showed that interactions explained most of the predictive signal. For HPAI, the interaction strength reached 0.53, meaning more than half the signal came from variables acting together rather than alone. Ignoring these joint effects would miss the real ecology, since risk rises sharply only when, for example, high duck density meets high human density near wetlands.

Which regions are most at risk for avian influenza?

High-risk areas spanned Asia from east to west, southern and western Europe, and northern Mediterranean, South African, and West African zones. The models also flagged likely underreporting in northern Algeria, Tunisia, Morocco, Lebanon, Iraq, the United Arab Emirates, and Oman, where predicted risk exceeded 0.5 but few outbreaks were recorded.

Do wetlands always increase outbreak risk?

Not for wild birds. Across much of the range, a higher count of Ramsar wetlands was linked to lower reported wild-bird risk, likely because protected remote wetlands keep waterfowl away from poultry and receive weaker surveillance. For general and HPAI risk, however, more nearby wetlands combined with high duck density raised risk, so the effect depends on the host group.

Can these models forecast future avian influenza outbreaks?

No. The models identify areas with ecological and host conditions that favor virus circulation, not specific future events. They are designed for risk-based surveillance and intervention planning, not prediction of timing or magnitude. They remain useful for prioritizing sampling, strengthening early warning, and directing resources to underreported high-risk zones.

How were the outbreak records prepared for modeling?

Records from two global surveillance systems were cross-checked and deduplicated, then thinned to a 10 km grid to prevent clustered reports from inflating accuracy. Of 54,463 raw records, 37,272 unique locations remained. Pseudo-absence points were generated at a 5 to 1 ratio to presence points, and models were validated with spatial block cross-validation across 150 km blocks.