ABSTRACT
Introduction: This study investigated the lagged correlation between Baidu Index for influenza-related keywords and influenza-like illness percentage (ILI%) across regions in China. The aim is to establish a scientific foundation for utilizing Baidu Index as an early warning tool for influenza-like illness epidemics. Methods: In this study, data on ILI% and Baidu Index were collected from 30 provincial-level administrative divisions (PLADs) spanning April 2014 to March 2019. The Baidu Index was categorized into Overall Index, Ordinary Index, Prevention Index, Symptom Index, and Treatment Index based on search query themes. The lagged correlation between the Baidu Index and ILI% was examined through the cross-correlation function (CCF) method. Results: Correlating the Baidu Overall Index of 30 PLADs with ILI% revealed CCF values ranging from 0.46 to 0.86, with a median lag of 0.5 days. Subcategory analysis indicated that the Prevention Index and Symptom Index exhibited quicker responses to ILI%, with median lags of -9 and -0.5 days, respectively, compared to 0 and 3 days for the Ordinary and Treatment Indexes. The median lag days between the Baidu Index and the ILI% were earlier in the northern PLADs compared to the southern PLADs. Discussion: The Prevention and Symptom Indexes show promising predictive capabilities for influenza-like illness epidemics.
ABSTRACT
Introduction: Respiratory infectious diseases, such as influenza and coronavirus disease 2019 (COVID-19), present significant global public health challenges. The emergence of artificial intelligence (AI) and big data offers opportunities to improve traditional disease surveillance and early warning systems. Methods: The study analyzed data from January 2020 to May 2023, comprising influenza-like illness (ILI) statistics, Baidu index, and clinical data from Weifang. Three methodologies were evaluated: the adaptive dynamic threshold method (ADTM) for dynamic threshold adjustments, the machine learning supervised method (MLSM), and the machine learning unsupervised method (MLUM) utilizing anomaly detection. The comparison focused on sensitivity, specificity, timeliness, and warning consistency. Results: ADTM issued 37 warnings with a sensitivity of 71% and a specificity of 85%. MLSM generated 35 warnings, with a sensitivity of 82% and a specificity of 87%. MLUM produced 63 warnings with a sensitivity of 100% and specificity of 80%. The initial warnings from ADTM and MLUM preceded those from MLSM by five days. The Kappa coefficient indicated moderate agreement between the methods, with values ranging from 0.52 to 0.62 (P<0.05). Discussion: The study explores the comparison between traditional methods and two machine learning approaches for early warning systems. It emphasizes the validation of machine learning's reliability and underscores the unique advantages of each method. Furthermore, it stresses the significance of integrating machine learning models with various data sources to enhance public health preparedness and response, alongside acknowledging limitations and the need for broader validation.
ABSTRACT
Objective: This study aimed to develop a universally applicable, feedback-informed Self-Excitation Attention Residual Network (SEAR) model. This model dynamically adapts to evolving disease trends and surveillance system changes, accommodating various scenarios. Thereby enhancing the effectiveness of early warning systems. Methods: Surveillance data on influenza-like illness (ILI) was collected from various regions including Northern China, Southern China, Beijing, and Yunnan. The reproduction number (Rt) was estimated to determine the threshold for issuing warnings. The Self-Excitation Attention Residual Network (SEAR) was devised employing deep learning algorithms and was trained, validated, and tested. The SEAR model's efficacy was assessed based on five metrics: accuracy rate, recall rate, F1 score, confusion matrix, and the receiver operating characteristic curve. Results: With an advance warning set at three days, the SEAR model outperformed five primary models - logistic regression, support vector machine, random forest, Extreme Gradient Boosting, and Long Short-Term Memory model - in all five evaluation metrics. Notably, the model's warning performance declined with an increase in the early warning value and the number of warning days, albeit maintaining a ROC value over 0.7 in all scenarios. Conclusion: The SEAR model demonstrated robust early warning performance for influenza in diverse Chinese regions with high accuracy and specificity. This novel model, augmenting traditional systems, supports widespread application for respiratory disease outbreak monitoring. Future evaluations could incorporate alternative indicators, with the model continuously updating through data feedback, thus enhancing its universal applicability. Ongoing optimization, using iterative feedback and expert judgment, heralds a transformative approach to surveillance-based early warning strategies.
