Development of a predictive model for weekly severe pneumonia cases among children under 5 years of age in Kampala using machine learning and ARIMA models

Abstract

Background: Pneumonia remains a significant public health concern, particularly among children under five in Kampala. Predicting severe pneumonia cases using climate and environmental factors can enhance early interventions and resource allocation. Objective: This study aimed to develop a predictive model for weekly severe pneumonia cases among children under five years old in Kampala, Uganda, utilizing machine learning approaches (Random Forest and extreme gradient boosting-XGBoost) and statistical modeling (Auto-Regressive Integrated Moving Average Exogenous -ARIMAX). Methods: A time series dataset incorporating severe pneumonia cases and climate variables such as temperature, rainfall, and humidity was used. The ARIMAX model was initially developed to capture temporal dependencies, while machine learning models (Random Forest and XGBoost) were applied for further predictive analysis. Model evaluation was performed using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Squared Error (MSE), with RMSE being the primary metric for model selection. The best-performing model was deployed as a web-based application using Flask for real-time predictions. Results: The XGBoost model achieved the lowest RMSE of 85.2 cases, indicating the highest predictive accuracy, followed by the Random Forest model (RMSE = 92.7), while the ARIMAX model recorded an RMSE of 105.4. An ensemble combining XGBoost and ARIMAX further improved performance, achieving an RMSE of 80.3 cases. From the statistical modeling perspective, ARIMAX analysis revealed that climatic factors such as high humidity and rainfall significantly influenced weekly severe pneumonia cases (p < 0.05). These findings demonstrate that integrating environmental data enhances disease forecasting accuracy and provides critical insights for early public health interventions. Conclusion: This study demonstrates the potential of machine learning in forecasting severe weekly pneumonia cases based on climate variables as well as ARIMAX models that effectively capture temporal patterns, highlighting key climatic drivers. The deployment of a Flask-based web application ensures accessibility for healthcare professionals, allowing real-time predictions for better planning and response. Future research should integrate additional air quality factors and explore deep learning techniques for improved prediction accuracy.