Data analysis and Forecasting of Tuberculosis and HIV co-infection: Exploring models from classical statistics to machine learning

Abstract Considering the complexity and severity of TB/HIV coinfection, accuracy in forecasting future trends is crucial for the efficient allocation of public health resources and the development of intervention strategies. This study explores the application of predictive models, ranging from classical statistical approaches to machine learning techniques, to analyze the time series of TB/HIV coinfection case notifications stratified for men, women, and the general population. Traditional models using Exponential Smoothing and Autoregressive Integrated Moving Average (ARIMA) methods were initially applied, providing a basis for understanding the overall temporal trend and the presence of seasonality. Subsequently, machine learning models, including Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), and Deep Neural Networks such as Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), were tested to find the best model to capture the complex dynamics of TB/HIV coinfection and the inherent non-linearities in the data. Performance evaluation, based on error metrics such as MSE, MAE, and sMAPE, revealed that Deep Learning models, especially Bidirectional LSTM and CNN combined with LSTM, significantly outperformed the classical statistical methods, demonstrating the effectiveness of these techniques for modeling TB/HIV coinfection time series and, consequently, providing more accurate forecast models.