tetano
Editor, Senior Moderator
BMC Med Inform Decis Mak
. 2026 Jun 5.
doi: 10.1186/s12911-026-03588-w. Online ahead of print.
Research on influenza surveillance and a prediction model based on multi-source data
Wei Duan[SUP] #[/SUP][SUP] 1 [/SUP], Lizhong Duan[SUP] #[/SUP][SUP] 2 [/SUP], Xiuhong Yang[SUP] 2 [/SUP], Lijuan Zhao[SUP] 2 [/SUP], Kai He[SUP] 2 [/SUP], Jiatian Yang[SUP] 2 [/SUP], Daiwen Cun[SUP] 2 [/SUP], Abeer Teeti[SUP] 1 [/SUP], Dongsheng Huang[SUP] 3 [/SUP], Xiaoqing Fu[SUP] 4 [/SUP], Xiaowen Wang[SUP] 5 [/SUP]
Affiliations
Objective: To develop and validate a multivariate Long Short-Term Memory (LSTM) model that integrates multi-source surveillance data for forecasting influenza activity. This study aimed to identify the most predictive variables and establish an optimized data fusion framework to enhance public health surveillance.
Methods: We collected influenza case data, influenza-like illness (ILI) reports, and symptom monitoring data, along with corresponding meteorological data and Baidu Index data in Baoshan city from January 2022 to June 2025. Spearman correlation analysis was used to verify the relationship between each dataset and influenza case numbers. Furthermore, the SHapley Additive exPlanations (SHAP) was employed to quantify feature importance. A LSTM model was constructed for predictive research, to identify in the optimal multi-source dataset. The prediction model based on this optimal dataset utilized the moving percentile method to determine the best early warning threshold.
Results: Influenza activity in Baoshan City exhibited distinct seasonality, with outbreaks peaking in winter and spring. ILI reports demonstrated the strongest correlation with confirmed cases (r[SUB]s[/SUB] = 0.56, p < 0.001). Among 43 Baidu Index keywords, four, including "H1N1 flu symptoms"(S2) showed higher correlations (r[SUB]s[/SUB] > 0.40, p < 0.001), whereas meteorological and symptom surveillance data were weak correlation. In predictive modeling, the LSTM achieved peak performance using ILI data alone (test set R[SUP]2[/SUP] = 0.79, MSE = 24.82, MAE = 3.14). Moreover, refined feature selection consistently outperformed models using full feature sets. The combination of ILI and the Baidu keyword "Is the Flu Shot Necessary?(P2)" provided the balanced and practical model (R[SUP]2[/SUP] = 0.79, MSE = 25.15, MAE = 3.18), matching the accuracy of the standalone ILI model. Based on the LSTM model for ILI predictions, the optimal early warning threshold was determined as P70, with the optimal threshold value being 8.10.
Conclusions: This study demonstrates that a strategically simplified LSTM model, leveraging refined multi-source data, can achieve high accuracy and robustness, providing solutions for public health surveillance scenarios. The threshold value of influenza epidemic warning in Baoshan city demonstrates reasonable sensitivity and specificity, and can be recommended as an early warning index of the influenza epidemic in Baoshan city.
Clinical trial number: Not applicable.
Keywords: Influenza; Long Short-Term Memory (LSTM); Machine Learning; Multi-source Data; Surveillance and prediction.
. 2026 Jun 5.
doi: 10.1186/s12911-026-03588-w. Online ahead of print.
Research on influenza surveillance and a prediction model based on multi-source data
Wei Duan[SUP] #[/SUP][SUP] 1 [/SUP], Lizhong Duan[SUP] #[/SUP][SUP] 2 [/SUP], Xiuhong Yang[SUP] 2 [/SUP], Lijuan Zhao[SUP] 2 [/SUP], Kai He[SUP] 2 [/SUP], Jiatian Yang[SUP] 2 [/SUP], Daiwen Cun[SUP] 2 [/SUP], Abeer Teeti[SUP] 1 [/SUP], Dongsheng Huang[SUP] 3 [/SUP], Xiaoqing Fu[SUP] 4 [/SUP], Xiaowen Wang[SUP] 5 [/SUP]
Affiliations
- PMID: 42249461
- DOI: 10.1186/s12911-026-03588-w
Objective: To develop and validate a multivariate Long Short-Term Memory (LSTM) model that integrates multi-source surveillance data for forecasting influenza activity. This study aimed to identify the most predictive variables and establish an optimized data fusion framework to enhance public health surveillance.
Methods: We collected influenza case data, influenza-like illness (ILI) reports, and symptom monitoring data, along with corresponding meteorological data and Baidu Index data in Baoshan city from January 2022 to June 2025. Spearman correlation analysis was used to verify the relationship between each dataset and influenza case numbers. Furthermore, the SHapley Additive exPlanations (SHAP) was employed to quantify feature importance. A LSTM model was constructed for predictive research, to identify in the optimal multi-source dataset. The prediction model based on this optimal dataset utilized the moving percentile method to determine the best early warning threshold.
Results: Influenza activity in Baoshan City exhibited distinct seasonality, with outbreaks peaking in winter and spring. ILI reports demonstrated the strongest correlation with confirmed cases (r[SUB]s[/SUB] = 0.56, p < 0.001). Among 43 Baidu Index keywords, four, including "H1N1 flu symptoms"(S2) showed higher correlations (r[SUB]s[/SUB] > 0.40, p < 0.001), whereas meteorological and symptom surveillance data were weak correlation. In predictive modeling, the LSTM achieved peak performance using ILI data alone (test set R[SUP]2[/SUP] = 0.79, MSE = 24.82, MAE = 3.14). Moreover, refined feature selection consistently outperformed models using full feature sets. The combination of ILI and the Baidu keyword "Is the Flu Shot Necessary?(P2)" provided the balanced and practical model (R[SUP]2[/SUP] = 0.79, MSE = 25.15, MAE = 3.18), matching the accuracy of the standalone ILI model. Based on the LSTM model for ILI predictions, the optimal early warning threshold was determined as P70, with the optimal threshold value being 8.10.
Conclusions: This study demonstrates that a strategically simplified LSTM model, leveraging refined multi-source data, can achieve high accuracy and robustness, providing solutions for public health surveillance scenarios. The threshold value of influenza epidemic warning in Baoshan city demonstrates reasonable sensitivity and specificity, and can be recommended as an early warning index of the influenza epidemic in Baoshan city.
Clinical trial number: Not applicable.
Keywords: Influenza; Long Short-Term Memory (LSTM); Machine Learning; Multi-source Data; Surveillance and prediction.