wkwarah/HIV_Treatment_Prediction_Ghana

R script of the analysis of Retrospective cohort data to predict HIV treatment Interruption in Ghana. Predict_IIT — Predicting HIV Treatment Interruption in Ghana This repository contains all analysis code for the study: > Kwarah W, da-Costa Vroom FB, Dwomoh D, Bosomprah S.** *Predicting HIV treatment interruption in Ghana: development and evaluation of an individual-level Machine Learning risk model.* BMC Global and Public Health (under review, 2026). --- ## Overview This project develops and validates a machine learning model to predict Treatment Interruption (TI) — defined as a visit gap of ≥28 days beyond the expected return date (PEPFAR COP 2021) — among people living with HIV (PLHIV) in Ghana. Using routine visit-level data from 33,613 PLHIV across 245 health facilities (2019–2023), an XGBoost model achieved an AUC-ROC of 0.978 on the hold-out test set. The best model is deployed as an interactive Shiny risk calculator: Predict_TI. --- Repository Structure ``` HIV_Preatment_Prediction_Ghana/ ├── └── Analysis/ ├── TI_Prediction_Analysis_Code.R ← PRIMARY analysis script ``` --- Analysis Script The script runs end-to-end from raw data to final outputs in 16 sequential sections: | Section | Description | |---|---| | 0 | Libraries and reproducibility seed | | 1 | Data loading and outcome (TI) derivation from visit-level records | | 2 | TI prevalence descriptives by calendar year | | 3 | Feature selection documentation (54 candidates → 33 retained) | | 4 | Imputation and encoding (KNN imputation, dummy variables, scaling) | | 5 | Stratified 70/30 train/test split | | 6 | Model training — 6 algorithms (XGBoost, Random Forest, Logistic Regression, KNN, Naive Bayes, AdaBag) | | 7 | Cross-validation model comparison | | 8 | XGBoost hyperparameter tuning via `xgb.cv` | | 9 | Final model evaluation on hold-out test set (AUC, Sensitivity, Specificity, F1, Kappa with 95% CIs) | | 10 | Calibration metrics (Brier score, calibration slope/intercept, ECE) | | 11 | Calibration plot | | 12 | SHAP global feature importance | | 13 | Individual-level SHAP waterfal …