Comparative analysis of traditional and deep learning time series architectures for influenza A infectious disease...

Influenza A remains a major cause of respiratory mortality worldwide, motivating accurate forecasting to support timely preparedness and resource allocation. This study presents a comparative evaluation of two traditional seasonal time series baselines, ARIMA and Holt-Winters exponential smoothing (ETS), and six deep learning (DL) architectures (Simple RNN, LSTM, GRU, BiLSTM, BiGRU, and a Transformer) for forecasting monthly Influenza A case counts in the United States. Data from January 2009 to December 2023 were analyzed, using January 2009 to December 2022 for training and January 2023 to December 2023 for out-of-sample testing. Models were tuned using a validation split and assessed with mean squared error (MSE) and mean absolute error (MAE), with GMRAE and Theil U1 reported to reflect relative performance. Results show consistent gains from DL over ARIMA and ETS, highlighting improved ability to capture nonlinear seasonality, irregular surges, and long-range temporal dependence in Influenza A dynamics. The best performance was obtained by the Transformer-based univariate framework, HistoFluAFormer, which uses only historical incidence sequences and applies self-attention with positional encoding to generate forecasts. On the test set, HistoFluAFormer achieved the lowest average errors (MSE 0.0433±0.0020; MAE 0.1126±0.0016), indicating strong generalization to unseen months. Beyond accuracy, the results highlight the feasibility of deploying a univariate Transformer within surveillance workflows when auxiliary covariates may be unavailable, delayed, or unreliable. These findings support the adoption of attention-based time series architectures to strengthen infectious disease forecasting pipelines and inform real-time surveillance, intervention planning, and epidemic preparedness.