Fusing meteorological factors and baidu search index with an adapted deformtime model to predict influenza-like illness in Beijing
TL;DR
A multimodal deep-learning architecture with epidemic-aware design substantially improves weekly ILI% forecasting in Beijing, particularly at 1–3 week horizons, with meteorological variables gaining prominence at longer lead times.
Abstract
Background Influenza-like illness (ILI) poses a significant burden on public health systems, particularly in densely populated megacities. Accurate and timely forecasting is crucial for early warning and resource allocation. While deep learning models have shown promise, capturing the complex, non-linear coupling between environmental drivers and human behavior remains a challenge. Methods We used weekly ILI% surveillance data from Beijing (2010–2019), complemented by 20 meteorological variables and 322 Baidu Search Keyword Indices. Cross-correlation function (CCF) analysis with seasonal differencing, ADF stationarity testing, and two-stage Benjamini–Hochberg FDR correction (q < 0.05) identified 11 predictive features, 6 meteorological variables and 5 search-derived features. An Adapted DeformTime model (DeformTime-Ad) was developed, incorporating a task-decoupled attention block, epidemic-level sample weighting, per-fold Optuna hyperparameter optimization, and Huber loss. Performance was evaluated against 10 comparators: persistence, seasonal-naïve, calendar-based, SARIMA, ARIMAX, SARIMAX, XGBoost, Random Forest, LSTM, and unadapted DeformTime-Base, under a 3-fold rolling-origin protocol across four horizons (7–28 days). Pairwise Diebold–Mariano tests, stratified evaluation by epidemic intensity, and six ablation variants were conducted. Results Beijing ILI% exhibited a pronounced unimodal winter–spring peak. On common target weeks, DeformTime-Ad achieved the lowest MAE at 7-−14 days (0.39–0.45) and the lowest MSE at 7–21 days, representing up to 62.0% MSE reduction over a reporting-lag-corrected persistence benchmark. In Diebold–Mariano tests, DeformTime-Ad was significantly superior to 8, 6, 2, and 1 of 10 comparators at 7, 14, 21, and 28 days, respectively, with zero significant losses at any horizon. During epidemic peaks (5 common target weeks), DeformTime-Ad attained a peak MAE of 2.09, 31.9% lower than DeformTime-Base, though the persistence benchmark achieved the lowest peak MAE (0.78) on this small sample, reflecting high autocorrelation during sustained peak plateaus. Ablation analysis identified per-fold hyperparameter optimization as the most impactful component (+51.9% MSE when removed), followed among architectural components by the task-decoupled attention block (+7.9%); epidemic-level sample weighting improved peak-period accuracy at a small overall-MSE trade-off. Meteorological and search-index features provided complementary, horizon-dependent predictive signals. Conclusions A multimodal deep-learning architecture with epidemic-aware design substantially improves weekly ILI% forecasting in Beijing, particularly at 1–3 week horizons. Both meteorological and internet-search features contribute predictive signal in a horizon-dependent manner, with meteorological variables gaining prominence at longer lead times. The framework, combining exploratory CCF-based feature screening, leakage-free rolling-origin evaluation on common target weeks, and component-level ablation, offers a methodology warranting further prospective, multi-city validation before operational deployment for infectious disease surveillance.