danielochieng23/nigeria-dhs-maternal-mortality-brf

Balanced Random Forest pipeline for a DHS sibling-history maternal-mortality proxy (Nigeria 2024) # Balanced Random Forest for a Maternal Mortality Proxy — Nigeria 2024 DHS Imbalance-aware pipeline that predicts a **sibling-history maternal-mortality proxy** from the Nigeria DHS Individual Recode (IR). The public-health goal is **high sensitivity** on a rare event (~1.6%), not headline accuracy. > **Critical limitation.** There is no VASA file and no clinical near-miss > indicator. A woman is coded positive if she reported a **sister** death > while pregnant, during delivery, after delivery, or within about six weeks > postpartum (`mm9_*`). This is **not** the respondent's own death and **not** > a Three Delays analysis. Use the model as triage / research support only. The written analysis is in `docs/report.docx`. ## What the pipeline produces | Stage | Module | Key outputs | |------|--------|-------------| | Data prep | `src/data_prep.py` | Analytic file, MM9 proxy, exclusion flow | | Data quality | `src/data_quality.py` | Missingness, imbalance, recommendations | | EDA | `src/eda.py` | Weighted proxy rates by subgroup | | Modeling | `src/modeling.py` | Balanced logistic, Balanced RF, SMOTE-Tomek + BRF | | Threshold | `src/threshold.py` | Sensitivity-first rule (recall ≥ 75%) | | Evaluation | `src/evaluate.py` | AUROC / AUPRC, confusion matrices, cost | | Interpretation | `src/interpret.py` | Tree feature importances | | Report | `src/report.py` | Executive HTML report | | Dashboard | `src/dashboard.py` | Streamlit explorer | ## Setup ```powershell py -m venv .venv .venv\Scripts\python.exe -m pip install -r requirements.txt ``` Place the Nigeria 2024 DHS download so that `NGIR8BDT/NGIR8BFL.dta` is visible, or set `NDHS_2024_DIR` to that folder. Restricted microdata is **not** committed. ## Run ```powershell .venv\Scripts\python.exe src\run_pipeline.py .venv\Scripts\streamlit run src\dashboard.py ``` Open `outputs/reports/maternal_mortality_report.html`. ## Locked analysis choices - **Outcome:** MM9 sibling-history proxy; sister deaths only when `m …