preeclampsia_ml

Machine learning workflow for benchmarking classical machine‑learning models on a maternal–fetal dataset. See: https://pmc.ncbi.nlm.nih.gov/articles/PMC12383967/

🚀 Quick‑start

# One‑time: clone repository, create env and install dependences
git clone https://github.com/cfarkas/preeclampsia_ml.git
cd preeclampsia_ml
python3 main.py --install_conda

# Run the full pipeline
python3 main.py --input dataframe.csv --output ./out_20_80/
# Re-Train
python3 main.py --input ./out_20_80/subset_25.csv --output ./out_20_80/25_perc_subset/

# Run the full pipeline (using k-fold instead of split)
python3 main_kfold.py --input dataframe.csv --output ./out_10kfold/
# Re-Train (using k-fold instead of split)
python3 main_kfold.py --input ./out_10kfold/subset_25.csv --output ./out_10kfold/25_perc_subset/

# Cochran-Armitage trend test (exact, two‑sided) for ordinal variables
python3 cochran_armitage.py --csv dataframe.csv --meta_xlsx PE_dataset_variables.xlsx

# Exact Fisher–Freeman–Halton test for r × c tables
python fisher_ffh.py --csv dataframe.csv --meta_xlsx PE_dataset_variables.xlsx

---

Key Features

Stage	Highlights
Environment	--install_conda bootstraps ml_preeclampsia env (Python 3.9 + all deps).
Reproducibility	Global seed SEED = 7 → identical splits, weights and importances every run.
Correlation analysis	Generates both a full Pearson matrix and a filtered matrix that keeps any variable showing |ρ| ≥ 0.12 with at least one outcome.
Models	Classification models: LogReg, LDA, GNB, KNN, DecTree, RF, GradBoost, SVM, MLP. Regression models: RFreg, GBreg.
Metrics	Macro‑averaged recall (classification) → primary score. MSE / RMSE / MAE / R² (regression).
Visual reporting	• pdfA_*.pdf confusion‑matrix grids. • pdfB_*.pdf permutation‑importance bar grids. • Fig3_paper.pdf consolidated best‑model bars (18 × 14 in). • importances.pdf supersized dot‑plot across all models/outcomes. • Fig2_paper.pdf recall heat‑map. • Fig1_paper.pdf filtered correlation matrix.
Feature export	Three ready‑to‑use CSV subsets — importance > 0.02, top‑50 %, top‑25 % — for re‑training.

The subsets are obtained by the union of all predictors that fell within the top 25 % (or 50 %) permutation-importance ranks in the best-performing model for any outcome.

Implementation Notes

• Scaling: numeric features → StandardScaler; categoricals → label‑encoded.
• Permutation importance: 5 repeats, seeded, n_jobs=-1.
• Splits: 80 / 20 stratified (classification) or plain (regression), invoked with main.py
• 10 K-fold cross validation, shuffling variables prior cross-validation, invoked with main_kfold.py

---

Outputs

results/
├── Fig1_paper.pdf # filtered correlation (|ρ| ≥ 0.12, 30 pt font)
├── Fig2_paper.pdf # recall heat‑map
├── Fig3_paper.pdf # best‑model permutation‑importances (panels A–F)
├── importances.pdf # dot‑plot: all models × all outcomes
├── pdfA_<outcome>.pdf # confusion‑matrix grids
├── pdfB_<outcome>.pdf # permutation‑importance bar grids
├── full_corr_matrix.pdf # complete Pearson matrix
├── regression_metrics_summary.txt
├── subset_overall_subset.csv
├── subset_top50_percent.csv
├── subset_top25_percent.csv
└── … (additional PDFs for every outcome)