diff --git a/CLAUDE.md b/CLAUDE.md
index dfbc1c5..631bf85 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -127,3 +127,36 @@
   English, translate faithfully — not a loose paraphrase — while still
   producing natural, idiomatic English rather than a stiff word-for-word
   rendering.
+
+## Drafting workflow (subagents)
+Sections get written through a fixed pipeline rather than drafted directly
+in English in one pass. Each stage is a project subagent under
+[.claude/agents/](.claude/agents/); read a stage's `.md` file for its full
+brief before invoking it.
+
+1. **Discussion** — no subagent; this is an ordinary conversation between
+   the user and the main session to settle what a section should say
+   (outline, argument, which facts/numbers apply).
+2. **`ja-drafter`** — writes a Japanese prose draft of the section from
+   that outline into [drafts/](drafts/)`ja/<N>_<name>.md` (mirroring
+   `sections/` numbering). Only uses facts given to it explicitly; never
+   pulls numbers from `Reference/`.
+3. **Review, in parallel, on the Japanese draft:**
+   - **`fact-checker`** — checks every number/statistic against the Notion
+     page (below) and flags anything unsupported or copied from `Reference/`.
+   - **`content-reviewer`** — checks logical coherence, clinical
+     plausibility, and terminology consistency with `Reference/main.tex`.
+   Both are read-only; they report findings rather than editing.
+4. **`en-translator`** — once review findings are resolved, translates the
+   reviewed `drafts/ja/*.md` file into faithful, idiomatic English LaTeX
+   and writes it into the matching `sections/*.tex` file. Never changes a
+   number during translation.
+
+Re-run the review stage (3) after any substantive re-draft, and again on
+the English output of stage 4 if it's not a purely mechanical translation.
+
+## 実験に関するデータの参照場所
+https://app.notion.com/p/AI-in-Medicine-3df9efa583a0803eba3ed15f51bf1b2a?source=copy_link
+
+##注意点
+上記のNotion以外の数字は基本的に信頼しないこと
\ No newline at end of file
diff --git a/drafts/README.md b/drafts/README.md
new file mode 100644
index 0000000..09a6b67
--- /dev/null
+++ b/drafts/README.md
@@ -0,0 +1,11 @@
+# drafts/
+
+Working Japanese drafts produced by the `ja-drafter` subagent, one file per
+manuscript section (`drafts/ja/3_method.md` corresponds to
+`sections/3_method.tex`, etc.).
+
+These are an intermediate step, not the deliverable: `fact-checker` and
+`content-reviewer` review the Japanese draft here, then `en-translator`
+renders the reviewed version into the actual manuscript at
+`sections/*.tex`. See `.claude/agents/*.md` and `CLAUDE.md` for the full
+workflow.
diff --git a/drafts/ja/.gitkeep b/drafts/ja/.gitkeep
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/drafts/ja/.gitkeep
diff --git a/drafts/ja/1_introduction.md b/drafts/ja/1_introduction.md
new file mode 100644
index 0000000..fd66992
--- /dev/null
+++ b/drafts/ja/1_introduction.md
@@ -0,0 +1,30 @@
+# Introduction（ドラフト）
+
+心筋梗塞や狭心症に代表される虚血性心疾患は，冠動脈の狭窄あるいは閉塞により心筋への血流が阻害され，胸痛や心筋障害を引き起こす重大な疾患群である．動脈硬化の進行に伴い血管壁に形成されるプラークはこの狭窄の主因であり，プラークが肥大化することで血管内腔が狭小化し，心筋への血液供給が低下する．こうした患者に対しては冠動脈バイパス術や経皮的冠動脈形成術（percutaneous coronary intervention: PCI）が行われており，このうちPCIはカテーテルを用いた低侵襲治療として広く普及している．
+
+しかしPCIには，No reflowおよびSlow flowと呼ばれる合併症のリスクが伴う．No reflowは，ステント留置などにより狭窄そのものは解消されるものの，病変部の拡張過程で流出したプラーク内容物や血栓が末梢の微小血管を閉塞し（末梢塞栓，distal microembolization），末梢血流が完全に遮断された状態を指し，微小血管攣縮もその一因とされる．Slow flowは，血流が完全には遮断されないものの著しく遅延した状態である．これらの合併症はPCI後の患者の約10〜20%に発症するとされ[babapoor2022]，急性心筋梗塞（AMI）症例では約10%以上の頻度で発症するとの報告もある【要出典】．いずれも心筋への酸素供給が絶たれることから，心不全の新規発症や心筋梗塞の再発，院内死亡率の上昇など，重大な予後不良因子となることが知られている【要出典】．発症後は血管拡張薬の冠動脈内投与などの治療が行われるが，既に末梢の毛細血管を閉塞している脂質や壊死物質などを除去することはできず，また急激な血圧低下や不整脈といった副作用のリスクも伴うため，事後治療のみでは限界がある【要出典】．こうした背景から，合併症の発症後に対処するのではなく，発症そのものを未然に防ぐ予防的戦略の確立が重要な課題となっている．
+
+術前にNo reflowやSlow flowのリスクが高いと判明した場合，臨床現場では病変遠位部にステント留置前から末梢保護デバイス（distal protection device）を留置する，あるいは血管拡張薬をあらかじめ冠動脈内投与するといった予防的対応が取られる【要出典】．末梢保護デバイスは高リスク病変に対しては有効である一方，準備に伴う負担からすべての症例に適用することは現実的でなく，術前の正確なリスク層別化が不可欠となる．血管内超音波（IVUS）は血管壁やプラークの状態を可視化でき，病変部に関する有用な情報を提供する．しかしIVUS画像の読影には専門的な知識を要するうえ，血管径やプラーク組成など複数の因子が複雑に絡み合うため，目視のみによる高精度なリスク予測は困難であり，IVUS画像から定量的にリスクを評価する手法が求められている．
+
+IVUS上で観察される減衰プラーク（attenuated plaque）の量については，合併症を発症した患者群と発症しなかった患者群との間に有意差が認められるとする報告がある[wu2011]．これは，減衰プラークが多いほど血栓形成や微小血管閉塞が生じやすく，合併症のリスクが高まる可能性を示唆するものである．しかしこのスコアリングは医師による手動評価にとどまっており，自動化・客観化された予測手法が求められている．
+
+No reflowやSlow flowの予測に有効な情報は，IVUS画像そのもの，患者の臨床変数，病変の形態の交差点に存在すると考えられる．一方でこれらの合併症は発症自体が比較的稀であり，アウトカムがラベル付けされた症例の中でも陽性例はさらに限られるため，これらの情報をエンドツーエンドに学習しきることは難しい．
+
+以上を踏まえ，本研究ではIVUS画像，患者の臨床情報，病変のプラーク特徴を統合的に扱うアーキテクチャ「KSCNet」を提案する．KSCNetは，もともと腎結石の分類のために開発された3Dアーキテクチャ【要出典】を，IVUSデータの構造に合わせて再設計したものである．IVUSのプルバック画像はCTのような等方的な3Dボリュームではなく，時系列に沿った異方的な2D断面の系列であることから，3D畳み込みの代わりに，重み共有された2Dのフレーム単位エンコーダとフレーム間のTransformerによる系列モデリングを採用する．具体的には，(1) 患者単位の臨床情報とプラーク特徴（減衰プラークの面積など）を統合した特徴ベクトルをフレーム系列への追加トークンとして注入し画像表現と相互作用させる仕組み，(2) 血管腔・血管壁のセグメンテーションを補助タスクとして用い，その特徴を手がかりに症例レベルの特徴をcross-attentionにより集約する仕組み，(3) セグメンテーション損失と分類損失を動的に重み付けするマルチタスク損失，という3つの要素を統合している．これにより，画像のみを用いる手法，臨床情報のみを用いる表形式手法，および既存の減衰プラークスコアリング手法と比較して，IVUS画像・臨床情報・病変特徴を組み合わせることが予測性能の向上に寄与するかを検証する．
+
+---
+
+## 挿入した【要出典】プレースホルダー一覧
+
+1. AMI症例での発症頻度「約10%以上」の出典（¶2）
+2. 重大な予後不良因子（心不全新規発症・心筋梗塞再発・院内死亡率上昇）の出典（¶2）
+3. 血管拡張薬による事後治療の限界（脂質・壊死物質の除去不可・血圧低下や不整脈のリスク）の出典（¶2）
+4. 術前リスク判明時の予防的対応（末梢保護デバイス留置・血管拡張薬事前投与）の出典（¶3）
+5. KSCNetの元となった腎結石分類手法（KidneyStoneSC/USCNet）の文献（¶6）
+
+## Introductionから移動した内容（2. Related Work用メモ、未配置）
+
+以下はユーザー指摘によりIntroductionから削除したが，Related Workで使える可能性があるため残しておく．「多くの研究」という表現の妥当性は要検証．
+
+- これまでNo reflowやSlow flowの予測に関する研究の多くは，IVUS画像から得られる情報を用いつつも従来型の統計モデルを中心とした解析にとどまっており，深層学習を用いた研究はほとんど行われてこなかった【要出典・「多く」の妥当性要検証】．
+- 我々自身も先行検討として（短報として発表済み），IVUS画像に極座標変換を適用したうえで，医師のアノテーションから算出した減衰プラークの角度統計量（平均・最大・最小・標準偏差）をCNN-LSTMアーキテクチャに明示的に統合する手法を提案している．この手法は画像と減衰プラーク由来の統計量のみを用いるものであり，患者の臨床情報や血管・プラーク領域のセグメンテーションは統合されていなかった．
diff --git a/sections/1_introduction.tex b/sections/1_introduction.tex
index 7157fd3..f25cf9b 100644
--- a/sections/1_introduction.tex
+++ b/sections/1_introduction.tex
@@ -1 +1,11 @@
-The introduction should briefly place the study in a broad context and highlight why it is important. It should define the purpose of the work and its significance. The current state of the research field should be reviewed carefully and key publications cited. Please highlight controversial and diverging hypotheses when necessary. Finally, briefly mention the main aim of the work and highlight the principal conclusions. As far as possible, please keep the introduction comprehensible to scientists outside your particular field of research. Citing a journal paper \citep{ref-journal}. Now citing a book reference \citep{ref-book1,ref-book2} or other reference types \citep{ref-unpublish,ref-url}. Please use the command \citep{ref-proceeding,ref-thesis} for the following MDPI journals, which use author--date citation: Administrative Sciences, AI in Education, Arts, Behavioral Sciences, Businesses, Econometrics, Economies, Education Sciences, European Journal of Investigation in Health, Psychology and Education, Family Sciences, Future Collections, Libraries, Archives, and Museums, Games, Genealogy, Histories, Humanities, Humans, International Journal of Cognitive Sciences, International Journal of Financial Studies, Investment, Journal of Intelligence, Journalism and Media, Journal of Risk and Financial Management, Journal of Strategic and Applied Marketing, Languages, Laws, Literature, Operations Management, Peace Studies, Project Management Horizons, Psychology International, Publications, Religions, Risks, Social Sciences, Tourism and Hospitality, Youth.
+Ischemic heart disease, typified by myocardial infarction and angina pectoris, is a serious group of conditions in which stenosis or occlusion of the coronary arteries impedes blood flow to the myocardium, causing chest pain and myocardial injury. Plaque that forms in the vessel wall as atherosclerosis progresses is the principal cause of this stenosis: as plaque enlarges, the vessel lumen narrows and the blood supply to the myocardium is reduced. Such patients are treated with coronary artery bypass grafting or percutaneous coronary intervention (PCI); of these, PCI---a catheter-based, minimally invasive treatment---has become widely established.
+
+However, PCI carries the risk of complications known as the no-reflow phenomenon and slow flow. No-reflow refers to a state in which, although the stenosis itself is resolved by stent placement or similar procedures, peripheral blood flow is completely obstructed because plaque contents or thrombus released during dilation of the lesion occlude the distal microvasculature (distal microembolization), with microvascular spasm also considered a contributing factor. Slow flow is a state in which blood flow is not completely obstructed but is markedly delayed. These complications are reported to occur in approximately 10--20\% of patients after PCI \citep{babapoor2022}, and some reports indicate an incidence of 10\% or more in cases of acute myocardial infarction (AMI) \textbf{[citation needed]}. Because both cut off the myocardial oxygen supply, they are known to be significant predictors of poor prognosis, including new-onset heart failure, recurrent myocardial infarction, and increased in-hospital mortality \textbf{[citation needed]}. After onset, treatment such as intracoronary administration of vasodilators is performed; however, this cannot remove the lipid and necrotic material that have already occluded the peripheral capillaries, and it also carries the risk of adverse effects such as acute hypotension and arrhythmia, so post-hoc treatment alone has its limits \textbf{[citation needed]}. Against this background, establishing a preventive strategy that forestalls the onset of these complications altogether---rather than addressing them after they occur---has become a critical challenge.
+
+When the risk of no-reflow or slow flow is found to be high before the procedure, clinical practice takes preventive measures such as placing a distal protection device distal to the lesion before stent deployment, or administering vasodilators intracoronarily in advance \textbf{[citation needed]}. While distal protection devices are effective for high-risk lesions, applying them to every case is not realistic given the burden of preparation, making accurate pre-procedural risk stratification essential. Intravascular ultrasound (IVUS) can visualize the condition of the vessel wall and plaque, providing useful information about the lesion. However, interpreting IVUS images requires specialized expertise, and because multiple factors---such as vessel diameter and plaque composition---interact in complex ways, high-precision risk prediction by visual assessment alone is difficult, creating a need for methods that quantitatively assess risk from IVUS images.
+
+Regarding the amount of attenuated plaque observed on IVUS, a significant difference has been reported between patients who developed complications and those who did not \citep{wu2011}. This suggests that a greater amount of attenuated plaque may make thrombus formation and microvascular occlusion more likely, thereby increasing the risk of complications. However, this scoring remains a manual assessment performed by physicians, motivating the need for an automated and objective predictive approach.
+
+Information useful for predicting no-reflow and slow flow is thought to lie at the intersection of the IVUS images themselves, the patient's clinical variables, and the morphology of the lesion. At the same time, because these complications are relatively rare in themselves, and positive cases are even more limited among the cases whose outcomes are labeled, it is difficult to learn all of this information in an end-to-end manner.
+
+Building on the above, this study proposes KSCNet, an architecture that integrally handles IVUS images, patient clinical information, and lesion plaque features. KSCNet is a redesign, adapted to the structure of IVUS data, of a 3D architecture originally developed for kidney stone classification \textbf{[citation needed]}. Because an IVUS pullback is not an isotropic 3D volume like a CT scan but rather an anisotropic sequence of 2D cross-sections arranged along a time series, we adopt, in place of 3D convolution, a weight-shared, per-frame 2D encoder together with sequence modeling via a Transformer across frames. Specifically, it integrates three elements: (1) a mechanism that injects a feature vector---combining patient-level clinical information and plaque features (such as attenuated-plaque area)---as an additional token into the frame sequence, allowing it to interact with the image representation; (2) a mechanism that uses segmentation of the vessel lumen and vessel wall as an auxiliary task and aggregates case-level features via cross-attention, guided by those features; and (3) a multi-task loss that dynamically weights the segmentation loss and the classification loss. This allows us to verify whether combining IVUS images, clinical information, and lesion features contributes to improved predictive performance, compared with methods that use only images, tabular methods that use only clinical information, and existing attenuated-plaque scoring methods.