「医師はCNNに、なぜそのように学んだのか尋ねませんでした。しかし、彼らはなぜ彼がこの決断を下したのかを知る必要があります。」 — 説明可能性は医療 AI 導入の必須条件です。
1. 医療において説明可能性がより重要なのはなぜですか?
| 状況 | XAI が不足している場合の結果 |
|---|---|
| AI ががん陽性 | 医師も理由がわからない → 信用できない |
| AI陰性がん | 医師は何かが欠けているのではないかと心配していましたが、さらに検査を行いました |
| ショートカットだからAIは間違っている | AI は病気のパターンではなくアーチファクトを学習します |
| 規制への提出 | FDA は透明性の証拠を要求 |
医療 AI における賢いハンス効果: モデルは偽の相関を学習します:
- 胸部 X 線モデルは病院のラベルであり、病理学ではありません
- 皮膚病変モデルは定規マーク = 黒色腫を学習します
- 肺炎モデル研究病院（ピッツバーグ = 「肺炎なし」）
2. Grad-CAM (レッスン 4 で学習) — 概要
画像モデルの Grad-CAM (クラス アクティベーション マッピング): 1.前方パス→予測 2. ターゲットクラスの逆伝播勾配 → 最後の変換層に移動 3. グローバル平均プール勾配 → 重要度の重み 4. 特徴マップの加重合計 → ヒートマップ
制限: CNN のみ、特徴マップ解像度でのみ説明可能。
3. SHAP — 表形式データの統合フレームワーク
import shap
import numpy as np
import pandas as pd
from sklearn.ensemble import GradientBoostingClassifier
# Ví dụ: ICU mortality prediction (readmission risk)
# Features: vitals, labs, demographics, diagnosis codes
class ICUMortalityExplainer:
"""
SHAP TreeExplainer cho tree-based models (XGBoost, GBM, RandomForest).
SHAP = unified measure of feature importance dựa trên game theory.
SHAP value của feature j = marginal contribution của j với tất cả subsets.
Tính chất quan trọng:
- Additive: sum(SHAP values) = prediction - expected value
- Local: explain từng prediction cụ thể (không phải global average)
- Consistent: nếu model phụ thuộc nhiều hơn vào feature → SHAP lớn hơn
"""
def __init__(self, model: GradientBoostingClassifier, X_train: pd.DataFrame):
self.model = model
self.feature_names = X_train.columns.tolist()
self.explainer = shap.TreeExplainer(model)
self.background_data = X_train
def explain_patient(self, patient_features: pd.DataFrame) -> dict:
"""
Explain prediction cho 1 bệnh nhân cụ thể.
"""
shap_values = self.explainer.shap_values(patient_features)
# Với binary classifier: shap_values có thể là list [class0, class1]
if isinstance(shap_values, list):
sv = shap_values[1] # Class 1 (mortality)
else:
sv = shap_values
prediction = self.model.predict_proba(patient_features)[0, 1]
expected = self.explainer.expected_value
if isinstance(expected, list):
expected = expected[1]
# Sort features by |SHAP| value
feature_impacts = sorted(
zip(self.feature_names, sv[0]),
key=lambda x: abs(x[1]),
reverse=True
)
return {
"prediction_probability": round(float(prediction), 4),
"expected_value": round(float(expected), 4),
"top_contributing_features": [
{
"feature": f,
"value": round(float(patient_features[f].iloc[0]), 3),
"shap_value": round(float(sv_val), 4),
"direction": "increases risk" if sv_val > 0 else "decreases risk"
}
for f, sv_val in feature_impacts[:10]
],
"waterfall_data": {f: sv_val for f, sv_val in feature_impacts}
}
def global_feature_importance(self, X_test: pd.DataFrame) -> pd.DataFrame:
"""Global importance = mean|SHAP| across test patients."""
shap_values = self.explainer.shap_values(X_test)
if isinstance(shap_values, list):
sv = shap_values[1]
else:
sv = shap_values
importance = pd.DataFrame({
"feature": self.feature_names,
"mean_abs_shap": np.abs(sv).mean(axis=0)
}).sort_values("mean_abs_shap", ascending=False)
return importance
def visualize_shap_explanation(explanation: dict) -> str:
"""
Generate text explanation suitable cho clinical display.
(Trong production: render as interactive chart với Plotly/D3)
"""
pred = explanation["prediction_probability"]
risk_level = "HIGH" if pred > 0.5 else "LOW"
lines = [
f"Mortality Risk: {pred:.1%} ({risk_level})",
"\nTop factors increasing risk:",
]
for f in explanation["top_contributing_features"]:
if f["shap_value"] > 0:
lines.append(
f" ↑ {f['feature']} = {f['value']} "
f"(+{f['shap_value']:.3f} risk contribution)"
)
lines.append("\nTop factors decreasing risk:")
for f in explanation["top_contributing_features"]:
if f["shap_value"] < 0:
lines.append(
f" ↓ {f['feature']} = {f['value']} "
f"({f['shap_value']:.3f} risk contribution)"
)
return "\n".join(lines)
4. ブラックボックス モデルの LIME
from lime.lime_tabular import LimeTabularExplainer
from lime.lime_image import LimeImageExplainer
def explain_with_lime_tabular(
model_predict_fn,
X_train: np.ndarray,
feature_names: list[str],
patient_instance: np.ndarray,
num_features: int = 10,
) -> dict:
"""
LIME: Local Interpretable Model-agnostic Explanations.
Ý tưởng: fit một linear model (interpretable) xung quanh 1 prediction.
1. Perturbate instance: tạo nhiều samples gần patient_instance
2. Get predictions từ black-box model cho perturbed samples
3. Fit weighted linear model (weight = similarity đến original)
4. Coefficients của linear model = local explanations
Ưu điểm: model-agnostic (works với bất kỳ model nào)
Nhược điểm: less stable hơn SHAP, linear approximation chỉ local
"""
explainer = LimeTabularExplainer(
training_data=X_train,
feature_names=feature_names,
mode="classification",
discretize_continuous=True,
random_state=42
)
explanation = explainer.explain_instance(
data_row=patient_instance,
predict_fn=model_predict_fn,
num_features=num_features,
)
return {
"local_prediction": explanation.predict_proba,
"feature_weights": explanation.as_list(),
"score": explanation.score, # R² of local linear model
}
5. 説明の臨床的検証
XAI を構築した後、臨床検証が必要です。
def evaluate_explanation_quality(
explanations: list[dict],
clinical_experts: list[str] = None
) -> dict:
"""
Metrics đánh giá chất lượng explanations:
1. Faithfulness: explanation có phản ánh đúng model behavior không?
- Occlusion test: mask top-K features → prediction thay đổi nhiều không?
2. Plausibility: explanation có hợp lý về mặt clinical không?
- Expert survey: bác sĩ có đồng ý không?
3. Stability: cùng patient, cùng prediction → similar explanation?
4. Completeness: sum(SHAP values) ≈ prediction - baseline?
"""
results = {}
# Faithfulness: monotonicity test
faithfulness_scores = []
for exp in explanations:
# Sort features by |SHAP|
ranked_features = sorted(
exp["top_contributing_features"],
key=lambda x: abs(x["shap_value"]),
reverse=True
)
faithfulness_scores.append(len(ranked_features) > 0)
results["faithfulness"] = sum(faithfulness_scores) / len(faithfulness_scores)
# Completeness check (SHAP property)
completeness_errors = []
for exp in explanations:
base = exp["expected_value"]
pred = exp["prediction_probability"]
shap_sum = sum(f["shap_value"] for f in exp["top_contributing_features"])
completeness_errors.append(abs((base + shap_sum) - pred))
results["mean_completeness_error"] = round(
float(np.mean(completeness_errors)), 4
)
return results
6. 演習
-
MIMIC-III 敗血症予測で XGBoost をトレーニングします。 10 人の患者の SHAP ウォーターフォール プロットを作成します。医師2人に聞く：その説明は合理的ですか？ (妥当性スコア)。
-
忠実度テストを実装します。トップ 3 の SHAP 特徴を使用してマスクし (平均値に置き換え)、予測の信頼性の低下を測定します。
-
同じ患者について SHAP と LIME を比較します。上位の特徴は一貫していますか?安定性テスト: LIME を 10 回実行し、特徴の重みの分散を測定します。
レッスン 13: FDA 規制 — AI を臨床現場に導入する。