Cell-Attentive LSTM With Lag-Aware Feature Selection for Multi-Week Epidemic Forecasting

Abstract

Accurate multi-week epidemic forecasting requires exploiting the time-lagged predictive structure of heterogeneous surveillance signals, yet most deep learning approaches treat these inputs as temporally synchronized. We present a two-stage framework that first recovers per-feature temporal offsets through a lag-aware, correlation-based feature selection stage, then forecasts with a Cell-Attentive LSTM (CellAtt-LSTM) that applies attention across the input features and inside the LSTM cell, where it refines the hidden state at each time step. Ablations on the US data show that the feature and cell-level attention components are complementary and that, on lag-aligned inputs, adding cross-temporal attention does not improve accuracy. Across COVID-19 data from the United States and Australia and dengue surveillance in Brazil, the framework significantly outperforms 13 baselines on the high-dimensional US data while remaining competitive on the lower-dimensional Australian and Brazilian data. The same selection pipeline, without dataset-specific tuning, surfaces structurally distinct predictive regimes, from inter-regional case spillover in the US and Brazil to digital search signals in Australia.