kellychandelle/explainable-mpox-forecasting

Reproducible machine learning pipeline for forecasting Mpox cases in Burundi using statistical and explainable AI models. *Development and Validation of Predictive Models for Forecasting Mpox Cases in Burundi* Overview This repository presents a reproducible machine learning workflow for forecasting daily Mpox cases in Burundi using multivariate time series models. The project compares classical statistical models and modern machine learning algorithms under two validation strategies: Simple Train/Test Split Rolling-Origin (Time-Aware) Validation The workflow also integrates SHAP explainability to improve model transparency and interpretation. Models Linear Regression Random Forest XGBoost Artificial Neural Network (MLP) Poisson Regression Negative Binomial Regression Prophet Feature Engineering The forecasting models were trained using epidemiological and temporal predictors, including: Lagged confirmed cases (1, 2, 3, 7 and 14 days) Moving averages Growth rate First differences Expanding mean Calendar variables Daily testing counts Suspected cases Evaluation Two complementary validation strategies were implemented: Random Train/Test Split (80/20) Rolling-Origin Forecast Evaluation Performance was assessed using: Mean Absolute Error (MAE) Root Mean Squared Error (RMSE) Coefficient of Determination (R²) Explainability An XGBoost surrogate model was trained on the engineered feature space to compute SHAP values, allowing interpretation of the variables driving the forecasting models. Reproducibility This repository contains: Synthetic dataset Complete source code