Abstract

Accurate forecasting of tuberculosis cases is essential for effective public health planning and intervention in the Philippines. Machine learning models offer promising solutions for capturing complex temporal patterns in disease trends and environmental relationships. This study aimed to evaluate and compare the performance of LSTM and MLP models for forecasting tuberculosis case counts using time series data from Zamboanga Sibugay spanning January 2020 to December 2024, integrating monthly precipitation data as an environmental factor. Multiple configurations of both models were trained on historical tuberculosis records and corresponding precipitation measurements, with performance assessed using RMSE and MAPE to examine the effects of model architecture. The MLP model with 64 hidden units, ReLU activation, and 0.2 dropout rate achieved performance with the lowest RMSE (46.99) and MAPE (17.30%), significantly outperforming all LSTM configurations. Among LSTM models, variable performance was observed, with the 20-unit model achieving the best RMSE (50.69) and the 100-unit model delivering the best MAPE (22.48%). The MLP architecture demonstrated approximately 7–8% better accuracy across both metrics, effectively handling the dataset’s complexity, including pandemic-related disruptions, high temporal variability, and climate-health associations. These findings suggest that deep MLP networks with appropriate activation functions and moderate regularization are more effective than recurrent architectures for tuberculosis forecasting when incorporating environmental variables. The results demonstrate significant potential for improving tuberculosis surveillance and early intervention systems through advanced machine learning approaches that integrate climate factors.