JMIR Publications

ABSTRACT

Background:

Other infectious diarrhea (OID) remains an important public health concern in China because of its high incidence, marked seasonality, and substantial burden, particularly among children. Accurate short-term forecasting and early warning are important for timely public health response. However, previous OID forecasting studies have mainly relied on reported case data, and the added value of multisource indicators remains insufficiently evaluated.

Objective:

This study aimed to develop and evaluate a multisource CNN-BiLSTM-SE Attention model for short-term forecasting and early warning of reported other infectious diarrhea cases in Chongqing, China.

Methods:

Daily OID case counts in Chongqing from January 2015 to June 2025 were collected, together with meteorological variables and Baidu search indices related to infectious diarrhea. After data normalization, Pearson correlation analysis and random forest variable-importance analysis were used for predictor selection. A CNN-BiLSTM-SE Attention hybrid model was developed to integrate multisource data, extract local temporal patterns, model temporal dependencies, and recalibrate informative feature channels. Forecasting performance was evaluated using RMSE, MAE, MAPE, and R², and compared across different input settings and benchmark models. In addition, 5-day-ahead predictions were converted into binary warning signals using training-set 75th and 90th percentile thresholds, and compared with a persistence baseline.

Results:

Under the full-input setting, the CNN-BiLSTM-SE Attention model achieved the best predictive performance, with an R² of 0.7828, RMSE of 35.418, MAE of 25.411, and MAPE of 17.27%. Compared with the case-only model, R² increased by 0.0326, while RMSE and MAE decreased by 2.560 and 1.643, respectively. The proposed model also outperformed random forest, XGBoost, CNN, and LSTM. In the threshold-based early-warning evaluation, the full-input model showed better overall warning performance than the persistence baseline at both the 75th and 90th percentile thresholds.

Conclusions:

The CNN-BiLSTM-SE Attention hybrid model improved short-term forecasting of reported OID case counts in Chongqing. Integrating epidemiological, meteorological, and internet search data provided complementary information, suggesting potential utility for OID surveillance, forecasting, and early warning.