What question did this study set out to answer?

The study aims to develop and evaluate machine-learning models for accurate short-term influenza forecasting in Hunan Province, China.

July 10, 2026Open Access

Multi-source machine learning for short-term influenza forecasting in Hunan Province, China: a spatiotemporal analysis using surveillance, search, meteorological and air-quality data

QSQianlai SunHunan Provincial Center for Disease Control and Prevention JBJiangshaya BahatiXiamen University HZHongyun ZhaoHarbin University of Science and Technology

Puntos clave

The study aims to develop and evaluate machine-learning models for accurate short-term influenza forecasting in Hunan Province, China.
Utilized weekly influenza-like illness data from 54 hospitals across 14 cities from 2015 to 2024.
Integrated ILI+ data with Baidu Search Index, meteorological variables, and air quality data for forecasting 2020-2024.
Compared five machine-learning models for 1–3-week-ahead forecasting using multiple performance metrics.
The random forest model achieved the lowest RMSE (0.00882) and MAE (0.00389) and an out-of-sample R² of 0.723.
Machine-learning models outperformed the SeasonalNaive52 baseline in terms of RMSE, MAE, and SMAPE.
Model suitability varied across cities, with random forest performing best in 11 of 14 cities.

Resumen

Accurate short-term influenza forecasting is important for early warning and public health preparedness, but routine sentinel surveillance may be delayed and may not fully capture behavioural and environmental signals related to influenza activity. This study aimed to develop and evaluate multi-source machine-learning models for short-term influenza forecasting in Hunan Province, China. Weekly influenza-like illness plus (ILI+) data from 54 sentinel hospitals across 14 prefecture-level cities in Hunan Province were used to characterise influenza activity from 2015 to 2024. For forecasting during 2020–2024, ILI+ data were integrated with Baidu Search Index indicators, meteorological variables and air quality index data. ILI + was calculated as the product of the weekly ILI proportion and weekly influenza virus test positivity rate. Five models, including LSTM, GRU, stacked autoencoder, random forest and LightGBM, were compared for 1–3-week-ahead forecasting under four feature settings. Performance was evaluated using RMSE, MAE, MAPE, SMAPE and R², with SeasonalNaive52 as the baseline comparator. Influenza activity showed pronounced winter–spring seasonality with substantial interannual and city-level heterogeneity. At the provincial level, the all-features RF model achieved the lowest absolute errors among the machine-learning models (RMSE = 0.00882, MAE = 0.00389) and an out-of-sample R² of 0.723. Compared with the SeasonalNaive52 baseline, the machine-learning models showed lower RMSE, MAE and SMAPE and higher out-of-sample R². In city-level analyses, RF ranked highest in 11 of 14 cities based on the descriptive summary score, whereas GRU ranked highest in Yueyang, Zhangjiajie and Hengyang, indicating that model suitability may vary across local epidemic settings. We developed a city-level multi-source framework for short-term influenza forecasting in Hunan Province by integrating surveillance, Baidu Search Index, meteorological and air-quality data. Multi-source predictors improved forecasting beyond a seasonal naive baseline, with RF providing a robust option for reducing absolute weekly errors. City-level performance differences highlight the need for local validation before operational deployment. With prospective validation, this framework could support 1–2-week-ahead early warning and targeted influenza preparedness.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo