Narrative Event Extractor

イベント抽出フレームワーク

1. 有害事象（Adverse Events）抽出

共通有害事象パターン

adverse_event_extraction:
  infusion_related_reaction:
    keywords_ja: ["注入反応", "IRR", "輸注反応", "インフュージョンリアクション"]
    keywords_en: ["infusion reaction", "IRR"]
    context_clues: ["投与中", "投与後", "初回投与"]
    severity_markers:
      grade_1_2: ["軽度", "中等度", "対症療法で改善"]
      grade_3_4: ["重度", "投与中止", "入院", "ステロイド投与"]

  neutropenia:
    keywords_ja: ["好中球減少", "顆粒球減少", "neutropenia"]
    lab_confirmation: "ANC < 1500/μL"
    severity_markers:
      grade_1: "ANC 1500-2000"
      grade_2: "ANC 1000-1500"
      grade_3: "ANC 500-1000"
      grade_4: "ANC < 500"

  infection:
    keywords_ja: ["感染", "肺炎", "発熱", "敗血症", "帯状疱疹"]
    keywords_en: ["infection", "pneumonia", "sepsis", "herpes zoster"]
    pathogen_markers: ["細菌", "ウイルス", "真菌", "COVID"]
    severity_markers:
      mild: ["外来治療", "経口抗菌薬"]
      severe: ["入院", "IV抗菌薬", "ICU"]

  thrombocytopenia:
    keywords_ja: ["血小板減少", "thrombocytopenia"]
    lab_confirmation: "Plt < 100,000/μL"

  peripheral_neuropathy:
    keywords_ja: ["末梢神経障害", "しびれ", "感覚異常", "neuropathy"]
    severity_markers:
      grade_1: ["軽度", "日常生活支障なし"]
      grade_2: ["中等度", "器具的ADL制限"]
      grade_3_4: ["重度", "セルフケアADL制限"]

有害事象メタ情報抽出

ae_metadata:
  onset_date:
    patterns: ["[0-9]{4}/[0-9]{2}/[0-9]{2}", "○月○日", "Day[0-9]+"]
    relative_markers: ["投与後○日", "○クール目", "前回より"]

  outcome:
    resolved: ["改善", "回復", "軽快", "消失"]
    ongoing: ["持続", "継続", "残存"]
    fatal: ["死亡", "永眠"]

  action_taken:
    dose_reduced: ["減量", "dose reduction"]
    dose_delayed: ["延期", "スキップ", "休薬"]
    discontinued: ["中止", "終了", "off"]

  causality:
    related: ["関連あり", "因果関係あり", "薬剤性"]
    possible: ["可能性あり", "否定できず"]
    unlikely: ["関連なし", "因果関係なし"]

2. 効果判定（Response Assessment）抽出

骨髄腫特異的奏効評価

response_extraction:
  imwg_response:
    stringent_cr:
      keywords: ["sCR", "stringent CR", "厳格完全奏効"]
      criteria: "免疫固定法陰性 + FLC比正常 + 骨髄形質細胞<5%"

    complete_response:
      keywords: ["CR", "完全奏効", "complete response"]
      criteria: "免疫固定法陰性"

    very_good_partial:
      keywords: ["VGPR", "非常に良好な部分奏効"]
      criteria: "M蛋白90%以上減少"

    partial_response:
      keywords: ["PR", "部分奏効", "partial response"]
      criteria: "M蛋白50%以上減少"

    stable_disease:
      keywords: ["SD", "安定", "stable disease"]

    progressive_disease:
      keywords: ["PD", "進行", "増悪", "progression"]
      criteria: "M蛋白25%以上増加"

  response_context:
    assessment_timing: ["○クール後", "○ヶ月時点", "best response"]
    assessment_method: ["IMWG基準", "効果判定", "PET-CT"]

3. 症状・状態（Symptoms/Status）抽出

Performance Status

status_extraction:
  ecog_ps:
    pattern: "(?:PS|ECOG|パフォーマンスステータス)[：:\s]*([0-4])"
    keywords:
      ps_0: ["日常生活制限なし", "通常の活動", "無症状"]
      ps_1: ["軽度の症状", "軽作業可", "歩行可能"]
      ps_2: ["歩行可能", "自分の身の回り可", "日中50%以上起床"]
      ps_3: ["限られた自分の身の回り", "日中50%以上臥床"]
      ps_4: ["全く動けない", "完全に臥床", "セルフケア不能"]

  symptoms:
    pain:
      keywords: ["疼痛", "痛み", "骨痛", "腰痛", "NRS"]
      scale_pattern: "NRS[：:\s]*([0-9]|10)"

    fatigue:
      keywords: ["倦怠感", "疲労", "だるさ", "疲れ"]
      severity: ["軽度", "中等度", "重度"]

    appetite:
      keywords: ["食欲", "食事摂取", "経口摂取"]
      status: ["良好", "低下", "不振", "不良"]

4. 治療関連情報抽出

treatment_extraction:
  dose_modification:
    reduction:
      keywords: ["減量", "dose reduction", "○mg→○mg"]
      pattern: "([0-9]+)mg.*(?:から|→).*([0-9]+)mg"

    delay:
      keywords: ["延期", "スキップ", "休薬", "○日遅れ"]

  discontinuation:
    keywords: ["中止", "終了", "off treatment"]
    reasons:
      progression: ["PD", "進行", "増悪"]
      toxicity: ["副作用", "毒性", "有害事象"]
      patient_choice: ["患者希望", "本人希望"]
      death: ["死亡", "永眠"]
      other: ["その他", "転院"]

  line_of_therapy:
    pattern: "([0-9]+)(?:次治療|ライン|line)"
    keywords: ["初回治療", "二次治療", "三次治療", "salvage"]

出力フォーマット

# Narrative Event Extraction Specification: {EVENT_TYPE}

## 1. 概要

| 項目 | 値 |
|------|-----|
| 抽出対象イベント | {EVENT_TYPE} |
| 対象薬剤 | {DRUG} |
| 対象疾患 | {INDICATION} |
| テキストソース | {TEXT_SOURCE} |

## 2. 抽出仕様

### 2.1 対象イベント一覧
| イベント | カテゴリ | 優先度 |
|----------|---------|--------|
| ... | ... | High/Medium/Low |

### 2.2 抽出パターン定義

#### イベント1: [イベント名]
```yaml
event_name: "[イベント名]"