Participatory Surveillance for One-Week-Ahead Local ILI Early Warning: A Prospective Evaluation of Statistical and Machine-Learning Approaches

Timely local early warning of rising influenza-like illness (ILI) activity can support short-term public health planning, including situational review, staffing, and resource allocation. We evaluated whether participatory symptom reports can provide reliable one-week-ahead warning of unusually high local ILI activity across U.S. metropolitan areas. Using Outbreaks Near Me reports from April 2020 through December 2024, we constructed a weekly Core Based Statistical Area panel and defined elevated activity within each target-year fold using a training-only upper-tail threshold on a stabilized symptom rate. We then conducted a strictly prospective year-ahead evaluation for target years 2022 through 2024, comparing regularized logistic models, gradient-boosted trees, deep sequence models, and time-ordered ensembles under a shared feature set and fold-safe preprocessing pipeline. Ensemble predictors achieved the strongest rare-event discrimination, with AUPRC 0.6473, outperforming the best single model, XGBoost, with AUPRC 0.6296. Among base models, XGBoost showed the strongest probability reliability, and post-hoc calibration improved the stacked ensemble without changing its rank performance. Overall, participatory surveillance retained meaningful one-week-ahead predictive signal for local surge-risk ranking under prospective evaluation. These findings suggest that participatory surveillance can support local early warning of unusually high ILI-related symptom activity, and that boosted and ensemble methods offer the strongest practical performance for short-horizon alerting in this setting.