subgruppen_auc.png

def fig_subgruppen_auc(pipe: Pipeline) -> None:
    """Bar chart of AUC by Geschlecht subgroup, with bootstrap 95%-CI error bars."""
    cohort = load_cohort()
    cohort["geschlecht_clean"] = cohort["geschlecht"].replace({"w": "weiblich"})
    notes = load_notes()
    merged = notes.merge(
        cohort[["patient_id", "geschlecht_clean"]], on="patient_id", how="left"
    )

    _, test_idx = train_test_split(
        merged.index, test_size=0.25, stratify=merged[TARGET], random_state=SEED
    )
    test_data = merged.loc[test_idx].copy()

    proba_all = pipe.predict_proba(test_data["notiz"])[:, 1]
    overall_auc = roc_auc_score(test_data[TARGET], proba_all)

    # Display labels with correct German umlauts — "maennlich" is only the
    # internal data-encoding value, not what should appear on the chart.
    display_name = {"weiblich": "Weiblich", "maennlich": "Männlich"}

    groups, aucs, err_lo, err_hi, cis = [], [], [], [], []
    for group in ["weiblich", "maennlich"]:
        mask = test_data["geschlecht_clean"] == group
        if mask.sum() < 5:
            continue
        y_g = test_data.loc[mask, TARGET].to_numpy()
        p_g = proba_all[mask.values]
        auc_g = roc_auc_score(y_g, p_g)
        lo, hi = bootstrap_auc_ci(y_g, p_g)
        groups.append(display_name[group])
        aucs.append(auc_g)
        err_lo.append(auc_g - lo)
        err_hi.append(hi - auc_g)
        cis.append((lo, hi))

    # Flag a bar only if its CI does NOT overlap the overall AUC — a point
    # estimate more than 0.05 away from the mean can still be pure noise if
    # the CI is wide, so colour-coding must be based on the interval, not
    # the point estimate alone.
    colors = [EVENT if not (lo <= overall_auc <= hi) else PRIMARY for lo, hi in cis]

    fig, ax = plt.subplots(figsize=(5.5, 4.2))
    ax.bar(groups, aucs, color=colors, width=0.5,
           yerr=[err_lo, err_hi], capsize=6, ecolor="#3A3A3A", error_kw={"lw": 1.3})
    ax.axhline(overall_auc, color=SECONDARY, lw=1.2, ls="--",
               label=f"Gesamt-AUC ({overall_auc:.2f})")
    for i, (g, a, (lo, hi)) in enumerate(zip(groups, aucs, cis)):
        ax.text(i, hi + 0.02, f"{a:.2f}\n[{lo:.2f}-{hi:.2f}]", ha="center", fontsize=8.5)
    ax.set_ylim(0, 1.15)
    ax.set_ylabel("ROC-AUC")
    ax.set_title("Subgruppenleistung nach Geschlecht (mit 95%-Bootstrap-CI)")
    ax.legend(loc="lower right")
    save(fig, ASSETS / "subgruppen_auc.png")