models/metadata_model.json

{
  "project": {
    "name": "GDM Risk Prediction - France",
    "objective": "Prediction of gestational diabetes mellitus risk at the 1st trimester without biological variables",
    "version": "1.0.0",
    "trained_at": "2026-05-26T16:31:03",
    "status": "validated"
  },
  "selected_model": {
    "name": "Logistic Regression with Isotonic Calibration",
    "short_name": "dg_lr_isotonic",
    "bundle_path": "artifacts/models/dg_lr_isotonic_bundle.joblib",
    "rationale": "Best AUC-ROC (0.759) and AUC-PR (0.334) among all candidates; best Brier Score (0.088); interpretable for clinical context"
  },
  "candidate_models": [
    {
      "name": "Logistic Regression",
      "AUC_ROC": 0.759,
      "AUC_PR": 0.334,
      "Brier_Score": 0.088,
      "selected": true
    },
    {
      "name": "XGBoost",
      "AUC_ROC": 0.753,
      "AUC_PR": 0.32,
      "Brier_Score": 0.089,
      "selected": false
    },
    {
      "name": "Random Forest",
      "AUC_ROC": 0.748,
      "AUC_PR": 0.298,
      "Brier_Score": 0.09,
      "selected": false
    },
    {
      "name": "SVM",
      "AUC_ROC": 0.76,
      "AUC_PR": 0.335,
      "Brier_Score": 0.088,
      "selected": false,
      "sensitivity_internal": 0.884,
      "specificity_internal": 0.413,
      "external_detection_rate": 0.947,
      "external_FN": 13,
      "DCA_points_utiles": 99
    }
  ],
  "thresholding": {
    "decision_threshold": 0.06,
    "threshold_file": "artifacts/threshold.json",
    "rationale": "Threshold optimized for high sensitivity (88.7%) and high NPV (96.6%) to minimize false negatives in a screening context"
  },
  "performance": {
    "dataset": "Internal test set (80/20 stratified split, n=30000)",
    "AUC_ROC": 0.759394470526865,
    "AUC_PR": 0.3337401809714899,
    "Brier_Score": 0.088,
    "sensitivity": 0.8874074074074074,
    "specificity": 0.4060093896713615,
    "VPP": 0.15922381711855396,
    "VPN": 0.966041108132261,
    "F1_score": 0.27000225377507325,
    "balanced_accuracy": 0.6467083985393844,
    "MCC": 0.19171550055591624,
    "TP": 599,
    "FP": 3163,
    "TN": 2162,
    "FN": 76,
    "metrics_file": "artifacts/metrics_final.json"
  },
  "external_validation": {
    "cohort_positive_243": {
      "n": 243,
      "detection_rate_LR": 0.9259,
      "false_negatives_LR": 18
    },
    "cohort_mixed_455": {
      "n": 455,
      "description": "243 positive + 212 negative cases"
    }
  },
  "data": {
    "dataset_name": "dataset_dg_france_30000_final",
    "n_rows": 30000,
    "n_features": 18,
    "target": "gdm_label",
    "target_values": [
      "Non",
      "Oui"
    ],
    "prevalence_train": "imbalanced (~22% positive)",
    "class_balancing": "class_weight=balanced"
  },
  "features": {
    "input_schema_path": "artifacts/feature_schema/input_schema.json",
    "n_features": 18,
    "features_used": [
      "age_maternel",
      "parite",
      "niveau_etude",
      "zone_residence",
      "imc",
      "ta_systolique",
      "ta_diastolique",
      "hta_chronique",
      "sedentarite",
      "tabagisme",
      "alcoolisme",
      "atcd_gdm",
      "atcd_macrosomie",
      "atcd_preeclampsie",
      "atcd_familial_diabete_1er_deg",
      "sopk",
      "grossesse_multiple",
      "sa_premiere_consult"
    ],
    "excluded": {
      "identifiers": [
        "patient_id",
        "centre_id",
        "pays"
      ],
      "administrative": [
        "annee_inclusion"
      ],
      "biological_leakage": [
        "glycemie_jeun_1T",
        "ogtt_0min",
        "ogtt_60min",
        "ogtt_120min",
        "critere_dg"
      ],
      "collinear": [
        "poids_kg",
        "taille_cm"
      ]
    }
  },
  "preprocessing": {
    "pipeline": "sklearn.pipeline.Pipeline (embedded in bundle)",
    "categorical_missing": "filled with Non_renseigne before encoding",
    "numeric_missing": "SimpleImputer(strategy=median)",
    "encoding": "OneHotEncoder(handle_unknown=ignore)",
    "scaling": "StandardScaler on numeric features",
    "note": "Full preprocessing pipeline is saved inside the .joblib bundle — no separate preprocessing step needed at inference"
  },
  "inference": {
    "how_to_load": "import joblib; bundle = joblib.load(\"artifacts/models/dg_lr_isotonic_bundle.joblib\")",
    "bundle_keys": [
      "model",
      "preprocessor",
      "feature_names",
      "threshold",
      "calibrator"
    ],
    "prediction_example": "proba = bundle[\"model\"].predict_proba(X_preprocessed)[:,1]; label = int(proba[0] >= bundle[\"threshold\"])"
  },
  "compliance": {
    "disclaimer": "Research prototype only. Not validated as a medical device. Not to be used for autonomous clinical decision-making without appropriate medical, ethical, institutional and regulatory validation.",
    "data_not_included": true,
    "data_reason": "Raw and processed data contain potentially sensitive patient information and are excluded from version control per GDPR principles"
  }
}