COMPARATIVE EVALUATION OF MACHINE LEARNING AND TIME-SERIES MODELS FOR EPIDEMIC FORECASTING IN LOW-RESOURCE SETTINGS

Abstract

The forecasting of epidemics, if done accurately, could help in strengthening disease surveillance, outbreak preparedness and public health decision-making, especially in low-resource settings. The current study develops and evaluates the forecasting framework by combining classical time series models and machine learning algorithms. An epidemiological data set of 99840 observations was analyzed for a wide array of epidemiological, environmental, demographic, and health sector variables including weekly cases, deaths, hospitalizations, rainfall, temperature, relative humidity, vaccination, population density and health sector indicators. The experts used Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, and coefficient of determination (R²) to evaluate the forecasting performance.  SARIMA performed best among the traditional time-series models with MAE = 13.51, RMSE = 18.67, MAPE = 12.41%, R² = 0.879, while ARIMA and Holt–Winters did worse. The findings show that the machine learning models performed better when compared to the other models. In fact, LSTM managed to get the highest prediction accuracy, which was found to have an MAE of 8.74, an RMSE of 11.63, an MAPE of 7.54%, and 0.953 R2 XGBoost was next best in prediction accuracy with MAE of 9.26, RMSE of 12.88, MAPE of 8.02% and 0.941 R2.  The results of the statistical analysis indicate a significant outperforming of the machine learning models over the traditional methods, specifically ARIMA versus XGBoost (t = 7.94, p < .001) and SARIMA versus LSTM (t = 5.13, p < .001). Applying the independent samples t-test between the accuracy results of XGBoost and LSTM produced t = 1.33, p = .185 meaning no statistically significant difference existed between the two results. Feature importance analysis clearly shows that Lag-1 Cases, Lag-2 Cases, and Lag-4 Cases are the top three predictors of future cases. LSTM took the longest training time (148.4 s) and memory consumption (1,245 MB) in terms of computational performance while SARIMA has much lower computational requirements of 13.8 s, 118 MB. The multi-criteria evaluation carried out under low-resource conditions ranked SARIMA as the foremost model with a score of 89.8 points. This model was the best balance of accuracy, efficiency, timeliness, robustness, interpretability, and operational feasibility. Findings show that while LSTM provides highest accuracy in prediction, SARIMA offers the most practical and scalable forecasting designs for epidemic surveillance and early warning application in low resource public health settings.