Deep Learning–Based Multivariate Time Series Forecasting of Acute Respiratory Infections Among Rohingya Refugees in Cox's Bazar, Bangladesh

Acute respiratory infections (ARIs) are among the leading causes of morbidity in Rohingya refugee camps in Cox's Bazar, Bangladesh, where overcrowding and environmental factors worsen disease transmission. Accurate forecasting models are necessary for early warning and timely interventions in such humanitarian settings. We compared five deep learning architectures, namely CNN–LSTM, BiLSTM, BiGRU, GRU, and LSTM using weekly ARI surveillance data of WHO's Epidemiological Highlights (January 2018–July 2025) combined with meteorological and environmental covariates obtained from NASA POWER, NASA GEOS-CF, Sentinel-2 and Sentinel-5P datasets. Model performance was evaluated using MAE, RMSE, and R2. SHAP analysis was incorporated to interpret feature contributions. The result showed that CNN–LSTM model attained the best performance (MAE = 286.23, RMSE=526.14, R2=98.63%), outperforming BiLSTM (R2=95.31%), BiGRU (R2=94.55%), GRU (R2=91.04%), and LSTM (R2=87.92%). CNN–LSTM predictions closely tracked observed ARI trends, including peaks and troughs, and provided reliable 52-week forecasts. SHAP analysis identified lagged ARI values as the most influential predictors, followed by year, humidity, and key pollutants$(\text{NO}_{2}, \text{CO}$, aerosols,$\text{PM}_{2.5}, \text{CH}_{4})$. Less influential features included precipitation, temperature, and formaldehyde. Among the evaluated architectures, the CNN–LSTM model demonstrated the best overall predictive performance, consistent with recent epidemic modeling literature. The interpretability analysis highlights the joint importance of temporal dynamics, climate, and pollution in driving ARI incidence. These results support the integration of deep learning–based forecasting into early-warning systems for epidemic preparedness in refugee and resource-limited settings.