Public health keeps miscounting: the gap between disease data and reality

The 17.5-fold rise nobody measured well

In a medRxiv preprint, researchers estimate that emergency department-treated pickleball injuries in the United States rose 17.5-fold between 2014 and 2023, with 70 percent of adjudicated cases involving adults aged 60 to 79. The absolute burden is real, but the hazard is not: no temporal trend was detected in the injury-to-participation measure (P=0.59), meaning the rise tracks participation growth rather than a sport getting more dangerous. m:…8650

The surveillance lesson is sharper. General internet search interest is not a reliable proxy for injury surveillance, showing low correlation after detrending and seasonal misalignment. This suggests that digital traces are a weak stand-in for clinical injury data, and public health teams that lean on search trends for injury surveillance are likely measuring attention, not harm. m:…8650

The pickleball result is a useful template: when a sport, product, or behavior grows fast, raw injury counts will always look alarming. The right question is the rate, and the right surveillance is the clinical kind. Anything cheaper risks confusing popularity with peril. m:…8650

Sixty counties of missed Valley fever

A second preprint tackles under-ascertainment directly. Using a model that integrates epidemiological, veterinary, and socioecological data, researchers identified 60 counties in non-reporting states where predicted coccidioidomycosis incidence reaches 2 or more cases per 100,000 population, areas where the disease is effectively invisible to official surveillance. m:…9228

Even where reporting is mandatory, the counts fall short: outside of Arizona and California, the model predicted county-level incidence rates 28 percent higher than those captured in national surveillance, estimating about 608 potentially missed cases per year, with a 95 percent bootstrap interval of 340 to 964. The implication is that Valley fever's geographic footprint is larger than the surveillance map admits. m:…9228

Billing codes that do not predict

The third preprint examined 21 US health systems and found that ICD-10 Z59 social-needs coding rates vary from 1 percent to 24 percent even after adjusting for patient mix, a spread that reflects documentation habits more than patient hardship. Z59 codes provide no incremental value to clinical and utilization-based risk prediction models. m:…3735

The financial footprint is not nothing: excess emergency and preventable hospital care associated with Z-coded patients is estimated at $2,000 to $2,500 per patient-year. But because the codes add no predictive power beyond what risk models already capture, this suggests health systems investing in Z-code capture as a risk-stratification strategy should validate whether the signal is the hardship or the coding practice. m:…3735

What held for deaths broke for infections

The fourth preprint, a municipality-level ecological analysis of COVID-19 in Austria from 2020 to 2023, draws a clean line between two outcomes. Structural vulnerabilities predicted mortality stably: the proportion of residents aged 85 or older was an independently significant predictor of municipality-level COVID-19 mortality, and home ownership was an independently significant protective factor with a relative risk of 0.92. m:…9646

For infections, the predictors would not sit still. Employment rates, nursing care allowance, and other drivers of SARS-CoV-2 infection exhibited temporal instability across the pandemic period. This suggests surveillance models built on socioeconomic indicators need to be re-estimated as transmission dynamics shift, because a predictor that worked in 2020 may be noise by 2022. m:…9646

The contrast between mortality and infection predictors is the paper's real contribution. Deaths tracked structure; infections tracked behavior, and behavior moved. Models that treat the two as the same problem will inherit the instability of one while trying to predict the stability of the other. m:…9646