top_tokens.png

def fig_top_tokens(pipe: Pipeline) -> None:
    """Horizontal signed bar chart of the 15 most predictive tokens."""
    feature_names = pipe.named_steps["tfidf"].get_feature_names_out()
    coef = pipe.named_steps["model"].coef_[0]
    n = 15
    idx = np.argsort(np.abs(coef))[::-1][:n]
    tokens = feature_names[idx]
    values = coef[idx]

    # Sort by coefficient for visual clarity (negative → positive)
    order = np.argsort(values)
    tokens_sorted = tokens[order]
    values_sorted = values[order]

    colors = [EVENT if v > 0 else PRIMARY for v in values_sorted]

    fig, ax = plt.subplots(figsize=(7, 5))
    bars = ax.barh(tokens_sorted, values_sorted, color=colors, height=0.7)
    ax.axvline(0, color=SECONDARY, lw=0.9, ls="--")
    ax.set_xlabel("Log-Odds-Koeffizient (logistische Regression)")
    ax.set_title("Prädiktivste Tokens für klinische Verschlechterung")

    # Legend patches
    from matplotlib.patches import Patch
    legend_elements = [
        Patch(facecolor=EVENT, label="Verschlechterung (positiv)"),
        Patch(facecolor=PRIMARY, label="stabil (negativ)"),
    ]
    ax.legend(handles=legend_elements, loc="lower right")
    save(fig, ASSETS / "top_tokens.png")