Machine learning exposes hidden coccidioidomycosis burden beyond the Southwest
Coccidioidomycosis, better known as Valley fever, still hides in plain sight across much of the United States. A cross-sectional analysis built from epidemiologic, veterinary, ecological, and sociodemographic data suggests the disease is more widespread than routine surveillance captures, especially outside Arizona and California, where reporting rules differ and clinical suspicion often lags behind exposure risk.
The investigators set out to answer a straightforward but consequential question: can alternative data streams be fused to estimate monthly, county-level incidence when nationwide reporting is incomplete? Their model, trained on surveillance data from Arizona and California between 2013 and 2023, was then applied across the contiguous U.S. and produced estimates for 3,108 counties. The approach matters because spores of Coccidioides are inhaled from soil, and infections may be missed when patients are diagnosed outside endemic settings or when local systems do not require reporting.
One result stands out immediately. The model identified 60 counties in non-reporting states with predicted incidence at or above 2 cases per 100,000 population, the threshold used here to define moderate to high incidence.
County-level incidence was higher than surveillance alone suggested
The analysis did not simply map a few isolated hotspots. In states with mandatory reporting outside Arizona and California, predicted county-level incidence was, on average, 28% higher than what national surveillance recorded. The 95% bootstrap interval ranged from 0.1% to 69.8%, a spread that signals real uncertainty but also a consistent direction of effect: routine reporting appears to miss a substantial share of illness.
The estimated shortfall translated into 608 potentially missed cases per year, with a 95% bootstrap interval of 340 to 964. For a fungal disease that already depends on environmental exposure, geography, and testing practices, under-ascertainment is not a statistical footnote. It shapes where clinicians think the disease exists, which patients get tested, and how public health agencies decide where to focus awareness campaigns.
The biologic logic behind the findings is familiar to infectious disease specialists. Coccidioidomycosis emerges where soil ecology, climate, and human behavior intersect, yet diagnosis depends on recognition at the bedside. If a county looks low risk on paper, patients may never reach the testing pathway that confirms infection.
Veterinary and ecological data filled gaps left by human reporting
A notable strength of the work is its use of multiple information sources rather than a single surveillance stream. Human cases diagnosed in electronic health records were combined with serological testing of dogs and with ecological and sociodemographic predictors. That design reflects a practical reality in zoonotic and environmentally driven infections: when human case reports are sparse, other signals can still reveal where exposure is occurring.
Dogs can function as sentinels for environmental fungal exposure because they share outdoor habitats and can develop antibodies after contact with contaminated soil. Ecological predictors, meanwhile, can capture the environmental conditions that make sporadic infection more likely, while sociodemographic factors may influence healthcare access, testing intensity, and case capture. The machine learning framework pulled those channels together into a county-by-month estimate rather than a static map, allowing the signal to vary over time as exposure patterns and reporting practices changed.
The model was not limited to the well-recognized endemic core of the disease. Its output extended across the contiguous United States, which is clinically important because missed emergence in new regions can delay diagnosis for years. When a clinician in a non-endemic county sees pneumonia, fatigue, or prolonged cough, coccidioidomycosis may not be considered unless local awareness has already caught up with shifting risk.
Under-recognition has direct consequences for diagnosis and prevention
The bedside implications are concrete. Under-counted incidence means under-tested patients, delayed antifungal treatment when it is needed, and more cases misclassified as routine community-acquired respiratory illness. It also means public health authorities may underestimate where workers, travelers, and residents are being exposed to contaminated dust.
The model’s finding that 60 counties in non-reporting states crossed the moderate-to-high incidence threshold is especially relevant for regions that do not routinely view Valley fever as a local disease. A county with an apparently low case count may still have meaningful transmission if clinicians are not ordering the right serologies or if cases are being treated empirically without pathogen confirmation. In that setting, missing the diagnosis does not just distort incidence curves; it shapes treatment decisions and the perceived need for local education.
These estimates also sharpen the contrast between mandatory reporting systems and true disease burden. Arizona and California have long served as the backbone for national understanding of coccidioidomycosis, but this analysis suggests even reporting states outside those two are not capturing the full picture. Where reporting is absent, the undercount is likely larger still.
Public health planning will need broader surveillance tools
The study argues for a broader surveillance strategy built from more than one data source. That includes routine human case reporting, but also veterinary sentinel data and environmental or sociodemographic indicators that can anticipate where exposure is likely to occur. For a disease shaped by climate and soil ecology, waiting for perfect passive reporting is a slow way to find new hot spots.
The findings also have a practical use for physician awareness. Counties classified as low incidence by standard reporting may warrant a different level of suspicion if modeled estimates point higher. In respiratory illness, especially when symptoms linger or routine antibiotics fail, that shift in pretest probability can alter whether fungal testing is pursued at all.
The study was approved by the relevant California and Arizona ethics bodies, and the authors declared no competing interests. Data are available upon reasonable request. What remains unresolved is not whether under-ascertainment exists, but how quickly state and local systems can adapt their surveillance to match the true geography of coccidioidomycosis exposure.