MizanHM/Predicting-NCD-Multimorbidity

Predict NCD multimorbidity in Ethiopia using machine learning # Predicting NCD Multimorbidity **Predicting NCD multimorbidity Risk in Ethiopia, A Machine Learning Approach** Ethiopia is undergoing an epidemiological transition. Diabetes mellitus (DM) and cardiovascular diseases (CVDs) are among the most common NCDs often occur together and share major risk factors.NCDs account for ~30% of national deaths, with CVD alone responsible for ~9%. The rise is driven by low awareness of risk factors, leading to late diagnoses and complications. **Workflow** *Data source*: Ethiopian STEPS survey (secondary, nationally representative). *Preprocessing*: feature engineering, imputation, transformation. *Labeling*: clinical guidelines + risk model + SSL. *Models tested*: Logistic Regression, Random Forest, CatBoost, GBDT, XGBoost, etc. *Multi-label strategies*: Binary Relevance, Classifier Chains, Label Powerset, Multi-Output. *Evaluation metrics*: Macro F1, Hamming Loss, and Subset Accuracy. *Class imbalance*: Multiple balancing techniques experimented *Hyperparameter tuning*: Bayesian optimization *Interpretability*: SHAP values for feature importance **Findings** Initial findings: Multi-Output CatBoost, GBDT, CC_Logistic Regression best (F1 ≈ 0.65). Final model: Multi-Output GBDT Macro F1 = 0.67 Subset accuracy = 0.85 Hamming loss = 0.07 AUC: 0.88 (DM), 0.95 (CVD) The model shows a good performance in the simultaneous prediction of DM & CVD risk. Class imbalance → persistent challenge in minority class performance, limited improvements from balancing. Key drivers: Age, cholesterol, hypertension, family history, sex, Regional variation: Somali (+), Amhara (–). Developed a robust multi-label ML model with good generalizability. Supports early identification of NCD risks in Ethiopia. Provides actionable insights for public health interventions in resource-limited settings.