Seasonal Patterns and Short-Term Environmental Associations with AECOPD Hospitalizations in Shijiazhuang, China, 2020-2024.
TL;DR
Overall, environmental exposures were weak predictors of high-incidence days and season-specific models for high-incidence admission days were explored as an exploratory environmental-surveillance analysis rather than a ready-to-deploy clinical early-warning system.
Abstract
Objective
To examine the seasonal distribution of AECOPD hospitalizations in Shijiazhuang (2020-2024), the lag associations of air pollutants (PM2.5, PM10, NO2, O3, SO2, CO) and meteorological factors (mean temperature, diurnal temperature range) across seasons, and to explore the season-specific models for high-incidence admission days as an exploratory environmental-surveillance analysis rather than a ready-to-deploy clinical early-warning system.
Methods
This city-wide retrospective time-series study analyzed 22,468 AECOPD admissions in Shijiazhuang from January 1, 2020 to December 31, 2024, using hospitalization front-page records covering all tertiary, secondary, and community hospitals in the city, obtained from the municipal medical insurance database (inpatient admissions only; ED discharges without admission were not included). Each year was divided by solar terms into spring (3/7-6/7), summer (6/8-9/7), autumn (9/8-12/7), and winter (12/8-3/6), with daily air pollution and meteorological data matched to admissions. Univariate and multivariate Poisson generalized linear regression (GLM) assessed single-day lag associations (lags 0-7 as eight separate lags); multivariable models and prediction features used the 7-day mean of lags 0-6 (admission day + preceding 6 days), not an 8-day lag0-7 average. Using each factor's 7-day cumulative window as input, we built logistic regression, random forest, and gradient boosting models and evaluated high-incidence-day prediction via fivefold time-series cross-validation, treating the many lag × pollutant × season tests as exploratory. Sensitivity analyses added during revision included collinearity diagnostics (VIF/correlation matrix), quasi-Poisson models with smooth time, a distributed-lag nonlinear temperature model (DLNM), COVID-period analyses, humidity/heating checks, alternative lag windows, and a near-term exposure prediction re-run.
Results
AECOPD admissions peaked in winter and were the lowest in summer (winter 6,009; spring 5,854; autumn 5,554; summer 5,051); the male-to-female ratio was ∼3.7:1 and mean age 71.7 ± 9.1 years. After multivariate adjustment, each 10 µg/m³ rise in winter PM10 and NO2 was associated with a 6.9% (IRR 1.069, 95%CI 1.052-1.085) and 8.9% (1.089, 1.050-1.129) higher admission rate, respectively. A 1 °C fall in mean temperature corresponded to a 4.4% increase (IRR 0.958, 0.947-0.969) and a 1 °C wider diurnal range to a 5.5% increase (1.055, 1.034-1.077). O3 was positively associated in summer and autumn (1.035, 1.015-1.056; 1.022, 1.005-1.039). Apparent protective multipollutant IRRs (e.g., winter PM2.5) were interpreted cautiously given collinearity. Prediction discrimination was the modest even when above chance: winter random forest fivefold CV AUC of 0.653 ± 0.112; other seasons ranged 0.50-0.62. Overall, environmental exposures were weak predictors of high-incidence days. Diagnostic re-analyses (Section 3.4), after reconciling the panel to the locked 22,468-admission cohort, showed acceptable collinearity (all VIF < 5, addressed with single-pollutant models), overdispersion handled by quasi-Poisson, and nonlinear temperature effects by DLNM (cold-induced excess, heat-related reduction), alongside a 2022 admission trough (∼50-54% of 2020/2024) and a near-term exposure re-run (same-day and lags 1-3 features) that reproduced the modest winter discrimination (AUC ≈0.66); against a calendar-only baseline the incremental value of the environmental block was small (overall ΔAUC ≈ +0.02 to +0.045), indicating the exposures are primarily explanatory rather than operationally predictive.
Conclusion
AECOPD admissions show high winter incidence, with PM10, NO2, CO, low temperature, and a wide diurnal range as the main winter associated factors, and summer-autumn O3 exposure warranting attention. The 7-day cumulative-exposure model showed only the modest discrimination in winter (AUC ≈0.65) and is best viewed as an exploratory, hypothesis-generating surveillance analysis rather than a preliminary operational early-warning tool.