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

Abstract Background Reporting delays in laboratory confirmation constrain timely situational awareness in respiratory disease surveillance. This study aimed to develop and evaluate a reporting-lag-aware machine learning framework for short-term, age-stratified forecasting of respiratory disease activity under real-world surveillance constraints. Methods We analyzed 1004 days (November 1, 2023 to July 31, 2026) of surveillance data from a national sentinel hospital in Guangzhou, China, including physician-recorded or laboratory-supported influenza, COVID-19 and Mycoplasma pneumoniae cases, fever clinic visits and meteorological variables. Given a routine 2-day reporting lag, an XGBoost model predicted the 7-day-ahead case increment. Most count-derived predictors were normalized, while the latest observable case count was retained as a level anchor. The shrinkage-calibrated XGBoost forecast was combined with a rolling SARIMA forecast. Comparators included naive rules, SARIMA, an oracle-weather SARIMAX benchmark, LSTM and Transformer models. Performance was assessed using RMSE, MAE, R 2 , directional accuracy, alert sensitivity, positive predictive value and prediction-interval coverage over 363 scored target dates within a 371-day calendar evaluation window. Results The framework achieved RMSE values of 16.50, 12.64 and 6.14 in the total population, adults and children, respectively, with corresponding R 2 values of 0.640, 0.694 and 0.353. It significantly outperformed persistence in the total and adult series (Diebold-Mariano p = 0.001 and 0.003), but improvements over a 7-day moving average or rolling SARIMA were not statistically significant. No comparison was statistically significant in children. For the proposed framework, directional accuracy was 67.2% to 73.5%, and sensitivity for identifying days on which activity reached or exceeded twice the development-period median was 58.1% to 78.4%. The absolute-count specification performed worse over the extended evaluation period (RMSE = 20.32; R 2 = 0.453), during which fever clinic attendance declined by 31.5% and the case-to-visit ratio increased by 31.6%. Conclusions A reporting-lag-aware framework based on routine data provided modest supplementary 7-day situational awareness. Its improvement over persistence was approximately one case per day in MAE, with no significant advantage over a 7-day moving average or rolling SARIMA. Performance depended on the stability of count-based predictors under shifts in care-seeking behaviour, supporting prospective multi-site validation before broader use.