Application of the SARIMA-LSTM model to evaluate the effectiveness of interventions for Visceral Leishmaniasis

Abstract

Introduction: This study proposes a combined Seasonal Autoregressive Integrated Moving Average and Long Short-Term Memory (SARIMA-LSTM) model to enhance the accuracy of evaluating the effectiveness of visceral leishmaniasis prevention and control efforts in Yangquan, China.

Methodology: Data were obtained from the Yangquan Centre for Disease Control and Prevention. The hybrid model integrates a SARIMA component with a residual-based LSTM neural network.

Results: In the SARIMA-LSTM model, the LSTM component included seven hidden layer nodes, a learning rate of 0.001, 500 training epochs, a batch size of 256, and utilized the Adam optimization algorithm. The SARIMA-LSTM model demonstrated superior performance (MSE = 2.824, MAE = 1.279, RMSE = 1.681). A paired samples t-test revealed a statistically significant difference between predicted and actual case counts (t = -4.058, p < 0.001), indicating that the actual number of cases was lower than predicted.

Conclusions: The combined SARIMA-LSTM model outperformed the individual SARIMA and LSTM models, suggesting that the implemented interventions were generally effective.