The researchers, led by Zhiheng Luo and Yunting Liu of the Affiliated Guangdong Second Provincial General Hospital of Jinan University, confronted a constraint familiar to every public health agency: a routine two-day reporting delay between when a case occurs and when it appears in the surveillance record. Rather than pretending this lag does not exist, they designed their forecasting pipeline around it. Their XGBoost model, an extreme gradient boosting algorithm, was tasked with predicting the increment in cases seven days ahead, using only information that would actually be observable at forecast time. This reporting-lag-aware design is the study’s central technical contribution, because most published forecasting models implicitly assume the forecaster can see the latest true case counts, an assumption that collapses in real-world operations.

The data spanned November 1, 2023 to July 31, 2026 and included physician-recorded or laboratory-supported cases of influenza, COVID-19 and Mycoplasma pneumoniae, alongside fever clinic visit counts and meteorological variables. The team engineered their predictors with care: most count-derived features were normalized to ratios, which stabilizes them when overall attendance fluctuates, while the latest observable case count was deliberately retained as a level anchor. This hybrid specification lets the model distinguish between a genuine rise in infections and a mere surge in care-seeking behavior, a distinction that absolute counts alone cannot make. The XGBoost forecast was then shrinkage-calibrated and combined with a rolling seasonal autoregressive integrated moving average, or SARIMA, forecast, blending machine learning flexibility with classical time series stability.

Performance was assessed over 363 scored target dates within a 371-day calendar evaluation window, using a battery of metrics including root mean square error, mean absolute error, R-squared, directional accuracy, alert sensitivity, positive predictive value and prediction-interval coverage. The framework achieved RMSE values of 16.50, 12.64 and 6.14 cases in the total population, adults and children respectively, with corresponding R-squared values of 0.640, 0.694 and 0.353. Directional accuracy, the proportion of days on which the model correctly predicted whether activity would rise or fall, ranged from 67.2 percent to 73.5 percent across the three series. For identifying days on which activity reached or exceeded twice the development-period median, a proxy for alert-worthy surges, sensitivity ranged from 58.1 percent to 78.4 percent.

The statistical comparisons were revealing. Using Diebold-Mariano tests, which compare forecast accuracy between two methods on the same target dates, the framework significantly outperformed persistence, the naive rule that tomorrow’s increment equals today’s, in the total and adult series, with p-values of 0.001 and 0.003. However, improvements over a simple seven-day moving average or a rolling SARIMA forecast were not statistically significant. In children, no comparison reached statistical significance at all, likely reflecting the smaller and noisier case counts in that age group. The practical improvement over persistence amounted to roughly one case per day in mean absolute error, a gain the authors describe as modest supplementary situational awareness rather than a transformative forecasting capability.

Deep learning comparators, including long short-term memory networks and Transformer models, were also evaluated, alongside an oracle-weather SARIMAX benchmark that used future meteorological information as an upper reference. The fact that gradient boosting on carefully engineered routine features held its own against these architectures underscores a recurring lesson in applied forecasting: with limited training data and strong autoregressive structure, simpler models with thoughtful feature design often match or beat more complex ones. The study followed the TRIPOD+AI reporting guidelines and used tree-structured Parzen estimator optimization for hyperparameter tuning, with SHAP values available for interpreting feature contributions.

Perhaps the most important finding emerged from the extended evaluation period, during which the absolute-count specification of the model performed noticeably worse, with RMSE rising to 20.32 and R-squared falling to 0.453. Over that same interval, fever clinic attendance declined by 31.5 percent while the case-to-visit ratio increased by 31.6 percent. This divergence in care-seeking behavior destabilized the count-based predictors the model relied upon, demonstrating that forecasting performance depends critically on the stability of the underlying data-generating process. A model trained when patients flock to fever clinics may misjudge activity when they stay home, even if the infection itself behaves similarly. This vulnerability to distributional shift is a fundamental challenge for all surveillance forecasting systems, not just this one.

The authors are candid about the limits of their work. The study used data from a single sentinel hospital, and the retrospective design means the model was not prospectively registered or tested before analysis. They explicitly recommend prospective multi-site validation before broader use, noting that performance under shifts in care-seeking behavior remains the key uncertainty. The framework’s value proposition is also framed carefully: it provides supplementary seven-day situational awareness from routine data, not a replacement for the judgment of epidemiologists or for laboratory-based confirmation. In an era when machine learning forecasts are often oversold, this kind of measured evaluation is itself a public service.

For the science of epidemic forecasting, the Guangzhou study offers a template worth copying. It shows how to build a forecasting pipeline that respects the operational reality of reporting delays, how to specify predictors that are robust to attendance fluctuations, and how to evaluate performance with statistical rigor against meaningful baselines rather than straw men. It also demonstrates the discipline of publishing null or modest results: the finding that a sophisticated model barely beats a seven-day moving average is as informative as any headline-grabbing accuracy claim. As respiratory pathogens continue to circulate in overlapping waves of influenza, COVID-19 and Mycoplasma pneumoniae, tools like this one, honest about their lag-aware design and their limits, may become quiet but useful fixtures in the surveillance toolkit, provided they are validated across sites and monitored as care-seeking behavior evolves.

Subject of Research: Machine learning forecasting of respiratory disease surveillance under reporting delays

Article Title: Reporting-lag-aware machine learning framework for short-term forecasting of respiratory disease activity: a sentinel hospital study in Guangzhou

Article References: Luo, Z., Duan, W., Lin, Y., Wang, H., Liu, H., & Liu, Y. (2026). Reporting-lag-aware machine learning framework for short-term forecasting of respiratory disease activity: a sentinel hospital study in Guangzhou. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14653-0

Image Credits: AI Generated

DOI: 10.1186/s12879-026-14653-0

Keywords: machine learning, disease surveillance, forecasting, reporting delay, XGBoost, respiratory infections, Guangzhou, sentinel hospital, Reporting-lag-aware, machine, learning, framework