hamid0272/measles-mortality-predictive-modeling-ethiopia

Machine learning models and data cleaning scripts for my thesis on measles surveillance. # measles-surveillance-cleaning Machine learning models and data cleaning scripts for my thesis on measles surveillance. Predictive Modeling of Measles Patient Mortality in Ethiopia Overview This repository contains the source code, data preprocessing pipelines, and machine learning models developed for my MSc thesis in Public Health Data Science. The project addresses the critical challenge of identifying high-risk measles patients in Ethiopia by integrating national surveillance data with health facility readiness indicators. Problem Statement Despite a robust national surveillance system, Ethiopia experienced a fivefold increase in confirmed measles cases between 2021 and 2023. Current clinical assessment methods struggle to predict mortality due to the complex interplay between patient-level characteristics and systemic health facility limitations. Key Technical Approach To address the extreme rarity of mortality events (0.59% prevalence), this study implemented: Advanced Resampling: Applied Synthetic Minority Over-sampling Technique (SMOTE) to mitigate severe class imbalance. Model Optimization: Trained and compared six supervised learning algorithms: Ensemble Models: Random Forest (RF), XGBoost (XGB), Easy Ensemble (EE). Deep Learning: Multi-Layer Perceptron (MLP) and 1D-Convolutional Neural Networks (CNN). Interpretability: Used SHAP (SHapley Additive exPlanations) values to identify key predictors of patient mortality. Evaluation: Utilized Bayesian optimization for hyperparameter tuning and assessed performance using PR-AUC and F1-score to account for the minority class. Tech Stack Language: Python Data Analysis: Pandas, NumPy, Scikit-learn Machine Learning: Imbalanced-learn (SMOTE), XGBoost, Keras/TensorFlow (for CNN/MLP) Visualization: Matplotlib, Seaborn Explainability: SHAP Key Findings Ensemble-based models, particularly Random Forest and XGBoost, achieved the highest discriminatory power in identifying high-risk mortality outcomes. Key pr …