Meta-Learning Enhance the Influenza Surveillance Across Spatio-Temporal Heterogeneous Scenario by Recommending Suitable Statistical Models

Public health events monitoring, e.g. influenza surveillance, is a fundamental public service. Statistical Process Control (SPC) charts are widely adopted methods for analyzing medical time-series data in this domain. Public health monitoring is inherently complex due to spatio-temporal heterogeneous, particularly evident in influenza surveillance. Thus, numerous statistical models have been developed; some prioritize detection sensitivity, while others are tailored to specific scenarios, such as varying data distributions. Selecting and configuring the appropriate statistical model presents a significant challenge, particularly for non-experts. Given that the strong spatio-temporal heterogeneity of influenza data implies no universally optimal chart exists, selection remains a critical and urgently needed computational problem.To address this challenge, we introduce the Pandemic Monitoring Control chart Automated Recommendation Model (PMcharm). This model resolves two primary computational issues: (1) the difficulty in defining an suitable chart selection due to the absence of explicit features or labels; and (2) the implementation of a meta-learning-based recommendation framework while addressing severe class imbalance problem. PMcharm extracts features by integrating classical statistical descriptors with an LSTM encoder and generate meta-targets. Additionally, the SMOTE technique is employed to address the class imbalance issue.Evaluated using U.S. state-level influenza surveillance data, PMcharm significantly outperforms fixed selection strategies, achieving an average recommendation accuracy (RA) of 0.97. Furthermore, real-time application during the 2024–2025 U.S. influenza season confirmed PMcharm's ability to recommend the suitable control chart. For instance, in Alabama, the recommended Shewhart control chart achieved a run length of 9, compared to 12 and 13 for alternative methods. In conclusion, the proposed method significantly enhances monitoring performance under such complex scenarios.