Analysis of influenza-like illness trends in Saudi Arabia: a comparative study of statistical and deep learning...

To develop and evaluate forecasting models using the Holt-Winters statistical approach and the long short-term memory (LSTM) deep learning method for weekly seasonal influenza-like illness (ILI) incidences in Saudi Arabia. The study compares model performance and assesses the predictive value added by incorporating region-specific exogenous variables within Middle Eastern epidemiological modeling.

This study compared the performance of Holt-Winters and LSTM models in forecasting weekly ILI cases in Saudi Arabia, using data collected from 2017 to 2022. Time series analysis integrated exogenous variables including climatic conditions and population mobility trends. The Holt-Winters model employed both additive and multiplicative seasonal components. Model performance was evaluated using root mean squared error (RMSE), mean absolute percentage error, and R2.

The best-performing model, LSTM with exogenous variables, achieved an RMSE of 28.55, mean absolute error (MAE) of 0.14, R2 of 0.96, and percent bias (PBIAS) of +2.1%, indicating negligible systematic error. The LSTM model without exogenous variables demonstrated slightly lower accuracy (RMSE of 34.07, MAE of 0.18, R2 of 0.93, PBIAS of +5.8%), indicating strong predictive capability but less precision in determining peak ILI cases. The Holt-Winters model effectively captured seasonal and long-term trends, but showed a moderate performance with an RMSE of 82.57, MAE of 0.38, R2 of 0.58, and a high PBIAS of +14.2%, revealing significant unexplained variability during periods of high incidence fluctuation.

This study highlights the respective strengths and limitations of statistical and machine learning approaches for ILI forecasting.