A Climate-Aware Deep Learning Framework for Generalizable Epidemic Forecasting

TL;DR

This work tackles climate-informed forecasting of endemic RSV under limited real-time surveillance by introducing ForecastNet-XCL, a climate-aware, label-free framework that combines an XGBoost pre-module with a CNN–BiLSTM temporal backbone and an autoregressive component driven by climate signals. The model operates with a strictly recursive forecasting strategy and utilizes Monte Carlo dropout to provide calibrated uncertainty, achieving state-of-the-art performance across 34 US states for both one-week-ahead and multi-step horizons. A key contribution is the multi-state transfer learning approach (ForecastNet-XCL-MS) that leverages diverse climatic contexts to improve timing accuracy and robustness, even when local data are scarce. The work demonstrates practical potential for climate-aware early warning of endemic respiratory diseases and offers a template for adapting to other climate-sensitive pathogens.

Abstract

Precise outbreak forecasting of infectious diseases is essential for effective public health responses and epidemic control. The increased availability of machine learning (ML) methods for time-series forecasting presents an enticing avenue to enhance outbreak forecasting. Though the COVID-19 outbreak demonstrated the value of applying ML models to predict epidemic profiles, using ML models to forecast endemic diseases remains underexplored. In this work, we present ForecastNet-XCL (an ensemble model based on XGBoost+CNN+BiLSTM), a deep learning hybrid framework designed to addresses this gap by creating accurate multi-week RSV forecasts up to 100 weeks in advance based on climate and temporal data, without access to real-time surveillance on RSV. The framework combines high-resolution feature learning with long-range temporal dependency capturing mechanisms, bolstered by an autoregressive module trained on climate-controlled lagged relations. Stochastic inference returns probabilistic intervals to inform decision-making. Evaluated across 34 U.S. states, ForecastNet-XCL reliably outperformed statistical baselines, individual neural nets, and conventional ensemble methods in both within- and cross-state scenarios, sustaining accuracy over extended forecast horizons. Training on climatologically diverse datasets enhanced generalization furthermore, particularly in locations having irregular or biennial RSV patterns. ForecastNet-XCL's efficiency, performance, and uncertainty-aware design make it a deployable early-warning tool amid escalating climate pressures and constrained surveillance resources.

Figures (5)

• Figure 1: Schematic of ForecastNet-XCL. From left to right: inputs combine weekly climate fields, state-level demographics, and RSV surveillance; preprocessing adds calendar features (for example, holidays) and seasonal embeddings. The architecture is two-stage: Stage 1 trains gradient-boosted trees (XGBoost) on recent covariates to predict next-week incidence, whose shifted predictions supply label-free autoregressive lags; Stage 2 is a hybrid CNN-BiLSTM that ingests covariates and generated lags over a 16-week window and produces multi-week trajectories via strictly recursive rollouts. Convolutions capture short-range temporal structure while recurrent units capture seasonal and inter-annual dependence; uncertainty is quantified with Monte Carlo dropout(see Methods for full details).
• Figure 2: One-week-ahead RSV forecasting using state-specific training.a, Observed RSV incidence curves (solid black lines) in Arkansas overlaid with one-week-ahead forecasts (dashed lines) generated by six models: CNN, LSTM, hybrid CNN-LSTM, XGBoost, stacked tree ensemble, and ridge regression. The darker gray shaded region denotes the test period, while light gray shaded portions correspond to model fit on the training data. b, Boxplot quantitative comparison of forecasting accuracy using coefficient of determination ($R^2$, left) and Mean Absolute Relative Error (MARE, right) over 34 states.
• Figure 3: Recursive multi-step RSV forecasting performance across state-level scenarios. a, Forecasted RSV incidence curves (dashed lines) generated by the recursive autoregressive ForecastNet-XCL, plotted against the observed RSV values (solid black line) in various states. Each panel corresponds to the held-out final 30 % of each state’s time series, reflecting the heterogeneous reporting durations across 34 states. Shaded regions denote 95% Monte Carlo dropout-based uncertainty intervals. b, Distribution of $R^2$ (left) and MARE (right) for ForecastNet-XCL across all 34 states under recursive forecasting conditions. c, Comparative performance metrics for all models in the recursive setting, shown as metric range plots and cumulative distribution functions.
• Figure 4: Recursive multi-step RSV forecasting performance across single-state and multi-state trained scenarios. a, State-by-state comparison of forecasting accuracy, with coefficients of determination ($R^2$, left axis, orange) and mean absolute relative errors (MARE, right axis, blue) plotted for single-state (solid markers) and multi-state (hollow markers) training. b, Predicted versus observed center-of-gravity (COG) of the RSV season for each state–season under single-state (circles) and multi-state (squares) training. Points are colored by absolute circular timing error $|\Delta\mathrm{COG}|$ (weeks). c, Per-state timing error shown on a radial axis for single-state (orange) and multi-state (blue) models; values nearer 0 indicate better phase alignment. Bold labels mark the seven pretraining states (AR, CA, CO, SC, NY, WA, WI).
• Figure 5: Multi-state transfer learning (XCL) improves accuracy and signal quality.a, Summary of performance gains. Left, MARE signed–log difference $\mathrm{sgn}(\Delta)\ln|\Delta\mathrm{MARE}|$ with $\Delta\mathrm{MARE}=\mathrm{MARE}_{\text{XCL-MS-1year}}-\mathrm{MARE}_{\text{XCL-}k\text{year}}$, $k\in\{1,2,3\}$ (points: bootstrap means over 10,000 iterations; lines: 95% CIs; right labels: $P(\text{XCL-MS-1year better})$). Middle, $R^2$ difference with the same bootstrap summary. Right, bootstrap win probabilities $P(\text{XCL-MS-1year better})$ for each metric and horizon b, Left, MARE boxplots; right, $R^2$ boxplots for ForecastNet-XCL-1year, ForecastNet-XCL-2year,ForecastNet-XCL-3year, and ForecastNet-XCL-MS-1year Notation:, For brevity in panel labels, ForecastNet-XCL (single-state) and ForecastNet-XCL-MS (multi-state) are abbreviated as $\text{XCL-}k\text{year}$ and $\text{XCL-MS-1year}$, respectively.