rsf_vs_cox.png

def fig_rsf_vs_cox(X_tr_np, X_te_np, events_train, events_test,
                   times_train, times_test) -> None:
    """Time-dependent AUC comparison (RSF vs. Cox) or fallback bar chart."""
    fig, ax = plt.subplots(figsize=(7, 4))

    cox_scores = cox_risk_scores(X_tr_np, X_te_np, events_train, times_train)

    try:
        from sksurv.ensemble import RandomSurvivalForest
        from sksurv.metrics import cumulative_dynamic_auc

        surv_train = make_survival_array(events_train, times_train)
        surv_test = make_survival_array(events_test, times_test)

        rsf = RandomSurvivalForest(n_estimators=200, min_samples_leaf=10,
                                   random_state=SEED, n_jobs=-1)
        rsf.fit(X_tr_np, surv_train)
        chf_funcs = rsf.predict_cumulative_hazard_function(X_te_np, return_array=False)
        rsf_scores = np.array([fn(28) for fn in chf_funcs])

        # cumulative_dynamic_auc needs the FULL test survival array + full
        # score array in a single call (it builds the risk set per time point
        # internally). Masking surv_test/scores per-t corrupts the array's
        # time range and raises "times must be within follow-up time of test
        # data" for later time points. Requested times must be strictly below
        # the largest observed test follow-up time.
        valid_times = [t for t in TIMES if t < times_test.max()]
        auc_rsf_list, _ = cumulative_dynamic_auc(
            surv_train, surv_test, rsf_scores, valid_times)
        auc_cox_list, _ = cumulative_dynamic_auc(
            surv_train, surv_test, cox_scores, valid_times)

        x = np.arange(len(valid_times))
        width = 0.35
        ax.bar(x - width / 2, auc_rsf_list, width, color=PRIMARY,
               label="Random Survival Forest")
        ax.bar(x + width / 2, auc_cox_list, width, color=SECONDARY,
               label="Cox-Proportional Hazards")
        ax.set_xticks(x)
        ax.set_xticklabels([f"Tag {t}" for t in valid_times])
        ax.set_title("Zeitabhängige AUC: RSF vs. Cox-Baseline")

    except ImportError as exc:
        # scikit-survival is a required dependency for this figure — the
        # RSF-vs-Cox time-dependent AUC comparison is the whole point of the
        # chart, so silently degrading to a different chart (with a mismatched
        # caption) is worse than failing loudly. Install scikit-survival and
        # re-run instead of relying on a fallback here.
        plt.close(fig)
        raise SystemExit(
            "scikit-survival (sksurv) is required to generate rsf_vs_cox.png "
            "— install it with `pip install scikit-survival` and re-run. "
            "This figure specifically compares RSF vs. Cox time-dependent AUC; "
            "there is no meaningful fallback chart for that comparison."
        ) from exc

    ax.set_ylim(0, 1)
    ax.axhline(0.5, color=SECONDARY, lw=0.8, ls="--", label="Zufalls-AUC")
    ax.set_ylabel("AUC")
    # AUC values sit high (~0.8-0.9) in this cohort, so the legend goes in the
    # empty lower-right area instead of the default placement (which overlaps
    # the bars).
    ax.legend(loc="lower right")
    save(fig, ASSETS / "rsf_vs_cox.png")