Li Pei-Long, Huang Rong-Wei, Xie Rong-Man, Xie Juan, Liu Kai
Department of pulmonary and critical care medicine, Yunnan Key laboratory of Children's Major Disease Research, Kunming Children's Hospital, Kunming, China.
Kunming Children's Hospital & Children's Hospital Affiliated to Kunming Medical University, Kunming Medical University, Kunming, China.
BMC Public Health. 2025 Aug 16;25(1):2796. doi: 10.1186/s12889-025-23710-3.
The global incidence of Influenza-Like Illnesses (ILI) has demonstrated an overall increasing trend. In the context of climate change, it is imperative to conduct research on the impact of meteorological factors on epidemic prediction.
To assess the potential of meteorological factors with Long Short-Term Memory (LSTM) models for improving ILI incidence prediction accuracy, providing a reference for the future development of related public health applicability.
Data on ILI incidence from November 2017 to January 2022, along with corresponding meteorological data over the same period. Pearson correlation analysis was employed to validate the relationship between the meteorological data and ILI incidence. Various LSTM architectures to forecast ILI incidence. These models were tested both with and without incorporating the the meteorological data as an additional feature. Additionally, Kernel Attention Network (KAN) was introduced into the LSTM models to enhance their nonlinear learning capability.
The description of ILI incidence and meteorological show that all the related variables are characterized by certain periodic changes. After incorporating the meteorological data into the analysis, the Mean Absolute Percentage Error (MAPE) for predicting ILI incidence using LSTM and attention-based stacked LSTM was 46.31% and 30.74%. Additionally, the application of KAN to these models further enhanced their performance.
The study demonstrates that stacking layers within LSTM models and incorporating KAN can further enhance the representational capabilities of these models. These improvements suggest that by leveraging meteorological data and utilizing advanced LSTM architectures, those can achieve more accurate and reliable predictions of ILI incidence.
全球流感样疾病（ILI）的发病率呈总体上升趋势。在气候变化的背景下，开展气象因素对疫情预测影响的研究势在必行。
评估气象因素结合长短期记忆（LSTM）模型提高ILI发病率预测准确性的潜力，为相关公共卫生应用的未来发展提供参考。
收集2017年11月至2022年1月ILI发病率数据以及同期相应的气象数据。采用Pearson相关性分析来验证气象数据与ILI发病率之间的关系。使用各种LSTM架构预测ILI发病率。这些模型在纳入和不纳入气象数据作为附加特征的情况下都进行了测试。此外，将核注意力网络（KAN）引入LSTM模型以增强其非线性学习能力。
ILI发病率和气象情况的描述表明，所有相关变量都具有一定的周期性变化。将气象数据纳入分析后，使用LSTM和基于注意力的堆叠LSTM预测ILI发病率的平均绝对百分比误差（MAPE）分别为46.31%和30.74%。此外，将KAN应用于这些模型进一步提高了它们的性能。
该研究表明，在LSTM模型中堆叠层并纳入KAN可以进一步增强这些模型的表征能力。这些改进表明，通过利用气象数据并采用先进的LSTM架构，可以实现对ILI发病率更准确可靠的预测。
