med-health
寫醫療健康新聞,守住相對與絕對風險並列、證據等級和自殺報導守則
中文翻譯
--- name: "med-health" description: "寫醫療或健康新聞時使用,包含臨床研究突破、公共衛生警訊、藥品核准、流行病學、健康政策、病患故事、風險溝通,素材來源可以是研究論文、新聞稿、衛生主管機關聲明或訪談。這是 med-news-reporter 工作流程的健康線特化版,重點紀律有:相對風險的呈現方式、一定要附絕對基礎值、證據等級查核、去識別化規範,以及遵守 WHO 自殺報導守則。觸發語句例如 '寫一篇醫學新研究'、'draft a health news piece'、'整理流行病新聞'、'幫我把這份臨床試驗結果寫成新聞'、'健康新聞報導'、'write up this drug approval'、'health story from this study'。不要用在:醫療建議(→ 請諮詢醫療人員)、藥品行銷(→ mkt-pharma)、以機構自家口吻發的醫院公關稿或新聞稿(→ pr-press-release)。" metadata: category: "WP-50 大眾傳播" tags: ["news", "journalism", "health-news", "medical-journalism", "public-health", "epidemiology", "media-ethics"] --- # 醫療與健康新聞報導 > **這個 skill 是 med-news-reporter 的醫療/健康線特化版。** 先讀 med-news-reporter,了解通用的 6 步工作流程(類型選擇、素材盤點、事實查核、平衡、倫理、識讀)。這份檔案是在那之上再加**健康線專屬的紀律**。 ## 概述 內容提煉自 Stanford Medicine+Muse、Johns Hopkins SFDH、AHCJ(美國健康照護記者協會)、Columbia Mailman、台大公衛,以及台灣健康媒體倫理規範的健康新聞教材。涵蓋五種子類型:**研究突破/公共衛生警訊/藥品核准/健康政策/病患故事**。核心難題是:把統計證據翻成大眾看得懂的話,同時不能扭曲風險,也不能給出虛假的確定感。 ``` IRON LAW: Relative Risk Without Absolute Risk Is Misleading Every medical claim in the form "X% increase/decrease in risk" MUST cite absolute baseline numbers: baseline incidence, NNT (Number Needed to Treat), absolute risk reduction, or absolute risk change. "50% reduction in risk of heart attack" is meaningless without "from 4 in 1000 to 2 in 1000 per year". LLM default: lead with the relative risk (sounds dramatic), omit baseline. Readers then overestimate the clinical significance. Override that default by naming the denominator first, then the percentage. ``` 為什麼這件事不直覺:「降低 50%」*聽起來*比「每年每 1000 人少 2 次心臟病發作」有力得多,但兩者講的是同一個結果。從研究到媒體的轉譯,常常就是把這個順序搞反:新聞稿寫「降低 50%」,媒體直接用這個數字,讀者就以為臨床效果比證據實際支持的更大。這是健康新聞誇大報導最常見的單一來源。 **合理化對照表:以下這些說法都「不能」推翻鐵律:** | Claude 可能會想… | 為什麼這樣還是違規 | |---|---| | 「『降低 50%』就是研究結果,我照引就好」 | 只引相對風險、不附絕對基礎值,等於轉述一個不完整的事實。期刊論文裡有基礎值,新聞稿通常沒有。兩個都引,或兩個都不引並加上標記。 | | 「基礎值在方法章節,讀者可以自己查」 | 讀者不會查。這篇報導就是他們唯一會讀的脈絡。省略它是用遺漏來誤導,不只是不完整。 | | 「加上絕對數字會讓新聞比較不戲劇化」 | 那正是重點。準確不是 bug。如果絕對效果很小,讀者有權知道。 | | 「NNT 對一般大眾太專業」 | 沒錯,但它也是呈現臨床意義最清楚的方式。用旁邊一句話帶出 NNT(『意思是醫師大約要治療 500 人,才能預防 1 個病例』)。這不是選配。 | | 「研究者說『統計顯著』,這就是主要新聞」 | 統計顯著不等於臨床顯著。10 萬人的研究,只要效果是真的,連 0.5% 的差異都可以算「顯著」。p 值和效果量都要報。 | --- ## 何時使用 **觸發條件:** - 使用者提供健康/醫療素材(期刊摘要、臨床試驗結果、藥品核准公告、公衛警示、流行病學數據、健康政策聲明、病患訪談),並要求寫成新聞稿。 - 使用者提到「醫學新聞」、「健康新聞」、「health story」、「research reporting」、「drug news」、「epidemic coverage」、「health policy piece」、「clinical breakthrough」。 - 使用者換句話說:「寫一篇醫學新研究」、「整理流行病新聞」、「幫我把這份臨床試驗結果寫成新聞」、「turn this NIH press release into a story」、「draft a piece on this WHO alert」。 **輸入訊號:** - 有具名的疾病、藥物、治療、研究者、機構、研究名稱、臨床試驗編號或流行病學數據。 - 有臨床醫師、研究者、衛生主管機關(CDC、衛福部、WHO 等)的直接引述。 - 有統計宣稱(相對風險、發生率、盛行率、死亡率、療效、信賴區間)。 - 有法規狀態(FDA 核准、食藥署上市許可、Phase III 試驗完成)。 **何時不要用:** - 「我得了什麼病?」/個人症狀的醫療建議 → 引導使用者去找醫療人員,不要用新聞 skill。 - 藥廠以自家口吻發的新聞稿 → 用 `pr-press-release`。 - 醫院行銷/機構公關(「本院進階手術中心榮獲…」)→ 用 `pr-*`。 - 把未經證實的偏方或保健食品當成科學事實來推廣 → 拒絕,並建議使用者先檢查來源的可信度。 --- ## 方法 ### Step 0:通用流程交給 med-news-reporter 讀過或已載入 `med-news-reporter`,裡面有:素材盤點、事實查核、消息來源強度標記、平衡原則、媒體倫理檢核、媒體識讀自我檢查。**這裡不要重做那些步驟。** 這份檔案特化的是 Step 1、2、3,並新增健康線專屬的 Step 7(證據等級與風險呈現稽核)。 ### Step 1:判斷健康新聞的子類型 | 子類型 | 訊號 | 子範本重點 | |----------|---------|--------------------| | **研究突破** | 期刊論文、預印本、大學/NIH 的新聞稿 | 證據等級檢查;RR + AR 呈現;重複驗證狀態 | | **公共衛生警訊** | CDC 警示、衛福部公告、WHO 聲明、疾病爆發 | 絕對數字(病例、死亡);傳播風險;高風險族群;因應指引 | | **藥品核准** | FDA/食藥署核准、Phase III 完成、臨床試驗結果 | 試驗設計嚴謹度;療效與副作用發生率;NNT;費用/可近性;替代療法 | | **健康政策** | 給付決定、疫苗建議、篩檢指引、法規 | 政策理由;受影響族群;證據基礎;專家共識;不同意見 | | **病患故事** | 訪談、見證、個案敘事 | 去識別化規範;可推論性的限制;出處標註;專家脈絡 | 如果素材跨了好幾種子類型(例如某項研究觸發的政策改變),就依*主要新聞驅動力*來分類。 ### Step 2:來源審查與證據等級標記 每一項健康宣稱,在**第一次提到時**都要帶**證據等級標記**: ``` Evidence Hierarchy (strongest → weakest): 1. Meta-analysis / systematic review of RCTs 2. Large RCT (n > 500) 3. Small RCT (n < 500) 4. Cohort study / case-control study 5. Case series / case report 6. Expert opinion / editorials 7. Anecdote / single patient story ``` **壞的標記方式:**「新研究表示...」(哪一項研究?強度如何?) **好的標記方式:**「今年發表在 Lancet 的一項 1,200 人隨機對照試驗表示...」或「基於個案報告(證據等級 5)...但尚未進行人體試驗」 **消息來源層級(延伸自 med-news-reporter):** | 層級 | 例子 | 處理方式 | |------|----------|-----------| | **政府衛生主管機關** | CDC、衛福部、疾管署、食藥署、WHO | 直接引用;可信度最高的一層 | | **同儕評審期刊** | Lancet、JAMA、BMJ、Nature Medicine、台灣醫學會期刊 | 一律附期刊名稱 + DOI;附上發表日期 | | **預印本/尚未同儕評審** | medRxiv、bioRxiv | **一定要標示「尚未經同儕評審」**;發稿前需要編輯審閱 | | **大學新聞稿** | 拿不到實際論文時 | 視為 Tier 2.5;對照期刊預印本/摘要查證 | | **單一研究者的說法** | 沒有已發表的證據 | Tier 4;只能當作「專家意見」並明確加註保留 | | **製藥公司** | 臨床試驗的贊助者 | Tier 3–4;**一定要揭露資金來源**;盡可能和獨立數據交叉驗證 | | **病患軼事** | 訪談、見證、Facebook 貼文 | Tier 7;只能當作說明用的敘事,絕對不能當證據 | ### Step 3:健康線專屬的風險檢查 在 med-news-reporter 的通用倫理檢核之外,再加上: 1. **個人資料保護法(PDPA)+ 醫療法 §72(病患隱私)**: - 病患個案報導必須去識別化:年齡用區間(不寫確切歲數)、不寫具名的機構/醫院、不寫會讓人重新辨識身分的獨特病況。 - 例:❌ "63-year-old Mr. Chen A treated at NTU Hospital on March 15 for liver cancer with rare genetic mutation"(可重新辨識) - 例:✅ "A 60–65-year-old male with common cancer type"(已匿名) 2. **WHO 自殺報導守則**(核心,沒得商量): - **不要寫出**死亡的方式、地點或日期。 - **不要刊登**遺書或詳細的過程敘述。 - **一定要附上**求助專線(1925、安心專線、國際專線)。 - **要用的措辭**:"died by suicide" 或 "suicide"(不要用 "committed suicide"、"successful attempt")。 - 違反這條會提高模仿風險(WHO 記錄的 Werther 效應),這是專業上的責任風險。 3. **醫療廣告法**: - 如果消息來源跟產品有財務利益(藥廠、持股的研究者、有專屬療法的醫院),要揭露。 - 不要放大未經證實的宣稱(例如沒有 RCT 證據的保健食品「治癒」說法)。 4. **藥物名稱使用合理性**: - 優先用學名,不用商品名(用 "ibuprofen",不用 "Advil"),除非商品名對這則新聞很關鍵。 - 新藥:第一次提到時學名和商品名都寫,之後用學名。 5. **傳染病防治法**: - 疫情初期的數字常隨資料累積而修正。要明講:「截至 [日期],[來源] 通報 X 例」。 - 避免暗示因果(「接種後」≠「因接種而」),時間用語要精確。 ### Step 4:證據等級與風險呈現稽核(健康線專屬新增) 輸出前要做: 1. **證據強度稽核**:每項醫學宣稱都要確認有標證據等級。單一個案報告不能寫成「研究顯示」。 2. **相對 → 絕對換算**:每一個「增加/降低 X%」的宣稱,都要搭配絕對基礎值(見上面的鐵律)。 3. **NNT/ARR/基礎發生率**:至少放進其中一項指標,讓臨床意義有個落腳點。 4. **重複驗證狀態**:如果只是單一研究,要講出來(「首度證據」/「需要進一步確認」/「證實了先前的發現」)。 5. **信賴區間/不確定性**:要寫範圍,不能只寫點估計。寫「30% 到 40%」,不要只寫「35%」。 6. **資金揭露**:只要有任何消息來源在結果上有財務利益,就要早早揭露(「由 Pharma Corp X 資助」)。 --- ## 輸出格式 用 med-news-reporter 的基本格式,再在結尾的 meta footer 加上健康線專屬的項目: ```markdown [Headline / sub-headline / body paragraphs per med-news-reporter] --- **稿件類型**: 醫學研究報導 / 公衛警訊 / 藥品核准 / 健康政策 / 患者故事 **字數**: approx. XXX **消息來源層級**: 政府公衛機構 N / 同儕評審期刊 N / 預印本 N / 企業新聞稿 N / 專家意見 N / 患者訪談 N **醫學證據稽核**: - 每項醫學宣稱之證據等級: ✅ / ⚠️ (列出未標的) - 相對風險 + 絕對風險配對: ✅ / ⚠️ (列出缺項: RR 未伴絕對值、NNT、基礎風險) - 單一研究 vs 系統性評論: ✅ / N/A / ⚠️ - 95% CI / 不確定性表述: ✅ / ⚠️ (列出未含的宣稱) **患者隱私檢核**: - 去識別化: ✅ / ⚠️ (列出仍可追蹤身份的資訊) - 同意書揭露: ✅ / N/A / ⚠️ **WHO 自殺守則**: - 適用: N/A / ✅ (已遵守) / ❌ (違反項目) **利益衝突揭露**: - 資金來源: ✅ / N/A / ⚠️ (列出未揭露的利益相關) **待查證事項**: ... **倫理 / 識讀檢核摘要**: 〔交給 med-news-reporter 的 Step 4-5 footer〕 ``` --- ## 範例 請看 `examples/` 目錄: - `sample_input.md`:貼近真實的健康新聞來源素材(臨床研究新聞稿 + 衛生主管機關聲明 + 醫學會回應 + 病患軼事) - `sample_output.md`:產出的稿件 + meta footer + skill 的運作軌跡說明 --- ## 注意事項 - **WHO 自殺報導守則不可選擇遵守**:不報導方法、地點、遺書;必附求助專線(1925);違反會導致模仿效應(Werther Effect)的文件風險。這是法律 + 倫理 + 公衛的三重義務,不是新聞美學選擇。 - **絕對風險不伴相對風險才是報導失敗**:「某藥物將 X 病死亡風險降低 50%」若不說明基礎風險(例如 1000 人中 4 人 → 2 人),讀者會高估臨床意義。必須兩項並列,或都不列。 - **預印本(preprint)是非同儕評審版本**:medRxiv / bioRxiv 的論文尚未經 peer review,在描述時必須明確標註「未經同儕評審」。發稿前應檢驗是否已正式發表在期刊。 - **單一研究 ≠ 醫學共識**:即使 Lancet 刊登,一篇論文不足以宣稱「科學證明」。需要 meta-analysis/多中心驗證或醫學會聲明。LLM 傾向誇大單一研究的通用性。 - **NNT 的臨床意義直覺優於百分比**:「需治療 500 人才有 1 人受益」 vs 「療效提升 0.2%」——同一結果,直觀度差異極大。優先用 NNT;若無法計算則註明。 - **製藥/醫療機構資助須在引述時點明**:研究由廠商贊助、醫師兼任產業顧問、醫院銷售新技術——這些利益衝突不揭露等同隱瞞。放在第一次引述該來源時,不要藏在尾註。 - **患者故事去識別化不等於「模糊化」**:「一名 60 歲男性患者 A」≠ 「患者 X 是成功病例」;前者遵守個資法 §72,後者若細節足夠仍可追蹤(就醫時間+地點+罕見疾病組合)。審核時逐項檢驗。 - **疫情數據更新快,引用需註明日期**:「確診人數 X」不標明「截至 4 月 29 日」,會在重新整理後過時或誤導。時間戳必要。 - **「治癒率」/「完治」/「成功率」各有定義**:不加定義就引用會混淆:是 5 年存活率?完全緩解?部分緩解?必須明確定義或引用原始文件。 --- ## 參考資料 | 檔案 | 用途 | 何時讀 | |------|---------|--------------| | `references/sources_and_beats.md` | 台灣衛生醫療消息來源、機構、官方資料庫 | Step 2 來源審查 | | `references/glossary.md` | 醫學統計、流行病學、臨床試驗術語對照 | 遇到不熟的醫學術語時 | | `references/ethics_and_law.md` | PDPA / 醫療法 §72 / 醫療廣告法 / 自殺守則 | Step 3 風險檢查 | | `references/medical_evidence_reading.md` | 證據等級金字塔、相對風險誤導、P-hacking | Step 1/4 證據等級 | | `references/risk_communication.md` | 風險溝通原則、絕對 vs 相對、不確定性表述 | Step 4 風險呈現 | 相關 skill: - `med-news-reporter`:通用新聞工作流程(這個 skill 是它的特化版) - `med-political`:健康政策與法規新聞 - `stat-hypothesis-testing`:針對 RCT 和 meta-analysis 的進階統計識讀 - `stat-causal-inference`:處理觀察性研究中的因果宣稱 - `hum-source-criticism`:來源審查框架 --- ## 限制 - **不會對照即時的臨床試驗資料庫或醫學期刊做查證。** 這個 skill 只會標出需要查證的宣稱,不會去查 ClinicalTrials.gov、PubMed 或 Cochrane。編輯/使用者必須自己查證或提供資料。 - **適用的司法管轄範圍**:法規引用(PDPA、醫療法、醫療廣告法)反映的是台灣法律。用在美國(HIPAA)、歐盟(GDPR)、英國、香港或中國大陸時,原則仍適用,但條文需要替換。 - **WHO 自殺報導守則適用全球**,但各地的執法與後果不同。不管在哪個國家發稿,一律要遵守這份守則。 - **統計識讀**:民調品質、p-hacking 風險、效果量解讀,在 `references/medical_evidence_reading.md` 只*部分*涵蓋,要深入做方法學批判請用 `stat-hypothesis-testing` 或 `grad-survey-design`。 - **不能取代法律審查。** 涉及具名病患、藥品/治療批評或法規爭議的健康新聞,發稿前應該先給律師看過。 - **不會產生原創報導。** 如果提供的素材不夠或互相矛盾,這個 skill 會標出缺口,但不會自己去找更多訪談或更多研究。
英文原文
---
name: "med-health"
description: "Use when writing a medical or health news story — clinical research breakthroughs, public health alerts, drug approvals, epidemiology, health policy, patient stories, risk communication — from research papers, press releases, health authority statements, or interviews. Specializes the med-news-reporter workflow for health-beat discipline: relative risk framing, absolute baseline inclusion, evidence-hierarchy verification, deidentification protocol, and WHO suicide-reporting compliance. Triggers on phrases like '寫一篇醫學新研究', 'draft a health news piece', '整理流行病新聞', '幫我把這份臨床試驗結果寫成新聞', '健康新聞報導', 'write up this drug approval', 'health story from this study'. Do NOT use for medical advice (→ consult healthcare provider), pharmaceutical marketing (→ mkt-pharma), hospital PR/press release in house voice (→ pr-press-release)."
metadata:
category: "WP-50 大眾傳播"
tags: ["news", "journalism", "health-news", "medical-journalism", "public-health", "epidemiology", "media-ethics"]
---
# Medical & Health News Reporting
> **This skill specializes med-news-reporter for the medical/health beat.** Read med-news-reporter first for the general 6-step workflow (type selection, material audit, fact-check, balance, ethics, literacy). This file adds **health-specific discipline** on top.
## Overview
Distilled from health-journalism curricula at Stanford Medicine+Muse, Johns Hopkins SFDH, AHCJ (Association of Health Care Journalists), Columbia Mailman, NTU Public Health, and Taiwan health-media ethics standards. Covers five sub-types: **research breakthroughs / public-health alerts / drug approval / health policy / patient stories**. Core challenge: translating statistical evidence for public understanding without misrepresenting risk or false certainty.
```
IRON LAW: Relative Risk Without Absolute Risk Is Misleading
Every medical claim in the form "X% increase/decrease in risk" MUST cite
absolute baseline numbers: baseline incidence, NNT (Number Needed to Treat),
absolute risk reduction, or absolute risk change. "50% reduction in risk of
heart attack" is meaningless without "from 4 in 1000 to 2 in 1000 per year".
LLM default: lead with the relative risk (sounds dramatic), omit baseline.
Readers then overestimate the clinical significance. Override that default
by naming the denominator first, then the percentage.
```
Why this is non-obvious: "50% reduction" *sounds* much more impactful than "2 fewer heart attacks per 1000 per year", yet both describe the same result. Research-to-media translation routinely inverts this — the press release says "50% reduction", the outlet runs that number, and readers assume a larger clinical effect than evidence supports. This is the single most common source of health-news overclaim.
**Rationalization Table — these justifications DO NOT override the Iron Law:**
| Claude might think... | Why it's still a violation |
|---|---|
| "'50% reduction' is the research result, I'll just quote it" | Quoting a relative-risk figure *without* the absolute baseline is relaying an incomplete fact. The journal paper has the baseline; the press release usually does not. Cite both or cite neither + flag. |
| "The baseline is in the methods section, readers can look it up" | Readers will not. The article is the only context they read. Omitting it is misleading by omission, not just incomplete. |
| "Adding the absolute number makes the story less dramatic" | That is the *point*. Accuracy is not a bug. If the absolute effect is small, the reader deserves to know. |
| "NNT is too technical for general audiences" | True, and it's also the clearest way to show clinical significance. Use NNT in a side sentence ('meaning doctors would need to treat about 500 people to prevent one case'). Not optional. |
| "The researcher said 'statistically significant'—that's the main story" | Statistically significant ≠ clinically significant. A study of 100,000 people can show a 0.5% effect as "significant" if it's real. Report both p-value and effect size. |
---
## When to Use
**Trigger conditions:**
- User supplies health/medical material — journal abstracts, clinical trial results, drug approval announcements, public health advisories, epidemiological data, health-policy statements, patient interviews — and asks for a news piece.
- User asks for "醫學新聞", "健康新聞", "health story", "research reporting", "drug news", "epidemic coverage", "health policy piece", "clinical breakthrough".
- User paraphrases: "寫一篇醫學新研究", "整理流行病新聞", "幫我把這份臨床試驗結果寫成新聞", "turn this NIH press release into a story", "draft a piece on this WHO alert".
**Input signals:**
- Named disease, drug, treatment, researcher, institution, study name, clinical trial identifier, or epidemiological data.
- Direct quotes from clinicians, researchers, health authorities (CDC, 衛福部, WHO, etc.).
- Statistical claims (relative risk, incidence, prevalence, mortality, efficacy, confidence intervals).
- Regulatory status (FDA approval, 食藥署 listing, Phase III trial completion).
**When NOT to use:**
- "What do I have?" / personal medical symptom advice → direct user to healthcare provider, not journalism skill.
- Pharmaceutical company press release in the company's own voice → use `pr-press-release`.
- Hospital marketing / institutional PR ("Our Advanced Surgery Center Achieves...") → use `pr-*`.
- Promotion of unproven remedy or supplement as scientific fact → refuse; suggest user consult source integrity first.
---
## Methodology
### Step 0: Defer general workflow to med-news-reporter
Read or have already loaded `med-news-reporter` for: material audit, fact-checking, source-strength tagging, balance principle, media-ethics check, media-literacy self-check. **Do not re-implement those steps here.** This file specializes Steps 1, 2, 3, and adds health-specific Step 7 (Evidence Hierarchy & Risk Framing Audit).
### Step 1: Classify the health-story sub-type
| Sub-type | Signals | Sub-template focus |
|----------|---------|--------------------|
| **Research breakthrough** | Journal paper, pre-print, press release from university/NIH | Evidence level check; RR + AR framing; replication status |
| **Public health alert** | CDC alert, 衛福部 advisory, WHO statement, disease outbreak | Absolute numbers (cases, deaths); transmission risk; at-risk population; response guidance |
| **Drug approval** | FDA/食藥署 approval, Phase III completion, clinical trial results | Trial design rigor; efficacy + side-effect rate; NNT; cost/access; alternative treatments |
| **Health policy** | Coverage decision, vaccine recommendation, screening guideline, regulation | Policy rationale; affected population; evidence basis; expert consensus; dissenting opinion |
| **Patient story** | Interview, testimonial, case narrative | De-identification protocol; generalizability limits; attribution; expert context |
If material spans sub-types (e.g. a policy change triggered by a study), classify by the *primary news driver*.
### Step 2: Source vetting & evidence-hierarchy tagging
Every health claim must carry **evidence-level tag** at first mention:
```
Evidence Hierarchy (strongest → weakest):
1. Meta-analysis / systematic review of RCTs
2. Large RCT (n > 500)
3. Small RCT (n < 500)
4. Cohort study / case-control study
5. Case series / case report
6. Expert opinion / editorials
7. Anecdote / single patient story
```
**Bad tagging:** 「新研究表示...」(which study? what strength?)
**Good tagging:** 「今年發表在 Lancet 的一項 1,200 人隨機對照試驗表示...」or 「基於個案報告(證據等級 5)...但尚未進行人體試驗」
**Source tier (extends med-news-reporter):**
| Tier | Examples | Treatment |
|------|----------|-----------|
| **Government health authority** | CDC, 衛福部、疾管署、食藥署、WHO | Direct citation; highest credibility tier |
| **Peer-reviewed journal** | Lancet, JAMA, BMJ, Nature Medicine, 台灣醫學會期刊 | Always cite journal name + DOI; include publication date |
| **Preprint / not yet peer-reviewed** | medRxiv, bioRxiv | **Must flag as "not yet peer-reviewed"**; requires editor review before publication |
| **University press release** | Without access to actual paper | Treat as Tier 2.5; verify against journal preprint / abstract |
| **Single researcher quote** | Without published evidence | Tier 4; acceptable only as "expert opinion" with explicit caveat |
| **Pharmaceutical company** | Clinical trial sponsor | Tier 3–4; **always disclose funding source**; cross-verify against independent data when possible |
| **Patient anecdote** | Interview, testimonial, Facebook post | Tier 7; only acceptable as illustrative narrative, never as evidence |
### Step 3: Health-specific risk check
Beyond med-news-reporter's general ethics check, add:
1. **個人資料保護法 (PDPA) + 醫療法 §72 (Patient Privacy)**:
- Patient case reports must be de-identified: age range (not exact), no named institution/hospital, no unique medical conditions that allow re-identification.
- Example: ❌ "63-year-old Mr. Chen A treated at NTU Hospital on March 15 for liver cancer with rare genetic mutation" (re-identifiable)
- Example: ✅ "A 60–65-year-old male with common cancer type" (anonymized)
2. **WHO Suicide Reporting Guidelines** (essential, non-negotiable):
- **Do NOT name the method, location, or date** of death.
- **Do NOT publish a suicide note or detailed narrative.**
- **ALWAYS include helpline number(s)** (1925, 安心专线, international).
- **Do use framing**: "died by suicide" or "suicide" (not "committed suicide", "successful attempt").
- Breaking this rule increases copycat risk (Werther Effect documented by WHO); this is a professional liability.
3. **醫療廣告法 (Medical Advertising Law)**:
- If a source has financial interest in the product (pharma company, researcher with stock, hospital with proprietary treatment), disclose it.
- Do not amplify unproven claims (e.g., supplement "cure" claims without RCT evidence).
4. **藥物名稱使用合理性**:
- Generic name preferred over brand name ("ibuprofen" not "Advil") unless brand is essential to the story.
- New drugs: include both generic + brand on first mention; thereafter use generic.
5. **傳染病防治法 (Communicable Disease Control Act)**:
- Early epidemic numbers are often revised as data accumulates. State explicitly: "as of [date], [source] reports X cases".
- Avoid implied causation ("after the vaccine" ≠ "caused by the vaccine"); use temporal language precisely.
### Step 4: Evidence Hierarchy & Risk Framing Audit (health-specific addition)
Before output, apply:
1. **Evidence Strength Audit**: for each medical claim, verify it cites the evidence level. Single case reports must not be presented as "research shows".
2. **Relative → Absolute Conversion**: every claim of "X% increase/decrease" must be paired with absolute baseline (see Iron Law above).
3. **NNT / ARR / Baseline Incidence**: include at least one of these metrics to ground clinical significance.
4. **Replication Status**: if this is a single study, state so ("first evidence" / "needs confirmation" / "confirms earlier findings").
5. **Confidence Interval / Uncertainty**: include the range, not just the point estimate. "30% to 40%" not just "35%".
6. **Funding Disclosure**: if any source has financial stake in the result, disclose it early ("funded by Pharma Corp X").
---
## Output Format
Use the med-news-reporter base format, with health-specific additions to the meta footer:
```markdown
[Headline / sub-headline / body paragraphs per med-news-reporter]
---
**稿件類型**: 醫學研究報導 / 公衛警訊 / 藥品核准 / 健康政策 / 患者故事
**字數**: approx. XXX
**消息來源層級**: 政府公衛機構 N / 同儕評審期刊 N / 預印本 N / 企業新聞稿 N / 專家意見 N / 患者訪談 N
**醫學證據稽核**:
- 每項醫學宣稱之證據等級: ✅ / ⚠️ (列出未標的)
- 相對風險 + 絕對風險配對: ✅ / ⚠️ (列出缺項: RR 未伴絕對值、NNT、基礎風險)
- 單一研究 vs 系統性評論: ✅ / N/A / ⚠️
- 95% CI / 不確定性表述: ✅ / ⚠️ (列出未含的宣稱)
**患者隱私檢核**:
- 去識別化: ✅ / ⚠️ (列出仍可追蹤身份的資訊)
- 同意書揭露: ✅ / N/A / ⚠️
**WHO 自殺守則**:
- 適用: N/A / ✅ (已遵守) / ❌ (違反項目)
**利益衝突揭露**:
- 資金來源: ✅ / N/A / ⚠️ (列出未揭露的利益相關)
**待查證事項**: ...
**倫理 / 識讀檢核摘要**: 〔交給 med-news-reporter 的 Step 4-5 footer〕
```
---
## Examples
See `examples/` directory for:
- `sample_input.md` — realistic health-news source material (clinical study press release + health authority statement + medical society response + patient anecdote)
- `sample_output.md` — produced piece + meta footer + skill-trace explanation
---
## Gotchas
- **WHO 自殺報導守則不可選擇遵守**: 不報導方法、地點、遺書;必附求助專線(1925);違反導致 copycat 效應(Werther Effect)之文件風險。這是法律 + 倫理 + 公衛的三重義務,不是新聞美學選擇。
- **絕對風險不伴相對風險才是報導失敗**: 「某藥物將 X 病死亡風險降低 50%」若不說明基礎風險(例如 1000 人中 4 人 → 2 人),讀者高估臨床意義。必須兩項並列,或都不列。
- **預印本 (preprint) 是非同儕評審版本**: medRxiv / bioRxiv 的論文尚未經 peer review,在描述時必須明確標註「未經同儕評審」。發稿前應檢驗是否已正式發表在期刊。
- **單一研究 ≠ 醫學共識**: 即使 Lancet 刊登,一篇論文不足以宣稱「科學證明」。需meta-analysis / 多中心驗證或醫學會聲明。LLM 傾向誇大單一研究的通用性。
- **NNT 的臨床意義直覺優於百分比**: 「需治療 500 人才有 1 人受益」 vs 「療效提升 0.2%」——同一結果,直觀度差異極大。優先用 NNT;若無法計算則注明。
- **製藥/醫療機構資助須在引述時點明**: 研究由廠商贊助、醫師兼任產業顧問、醫院銷售新技術——這些利益衝突不揭露等同隱瞞。放在第一次引述該來源時,不要藏在尾註。
- **患者故事去識別化不等於「模糊化」**: 「一名 60 歲男性患者 A」≠ 「患者 X 是成功病例」;前者遵守個資法 §72,後者若細節足夠仍可追蹤(就醫時間+地點+罕見疾病組合)。審核時逐項檢驗。
- **疫情数据更新快,引用需註明日期**: 「確診人數 X」不標明「截至 4 月 29 日」會在重新整理後過時或誤導。時間戳必要。
- **「治癒率」/ 「完治」/ 「成功率」各有定義**: 無定冠詞引用會混淆:是 5 年存活率?完全緩解?部分緩解?必須明確定義或引用原始文件。
---
## References
| File | Purpose | When to read |
|------|---------|--------------|
| `references/sources_and_beats.md` | 台灣衛生醫療消息來源、機構、官方資料庫 | Step 2 source vetting |
| `references/glossary.md` | 醫學統計、流行病學、臨床試驗術語對照 | When unfamiliar medical terminology appears |
| `references/ethics_and_law.md` | PDPA / 醫療法 §72 / 醫療廣告法 / 自殺守則 | Step 3 risk check |
| `references/medical_evidence_reading.md` | 證據等級金字塔、相對風險誤導、P-hacking | Step 1/4 evidence hierarchy |
| `references/risk_communication.md` | 風險溝通原則、絕對 vs 相對、不確定性表述 | Step 4 risk framing |
Related skills:
- `med-news-reporter` — general news workflow (this skill specializes it)
- `med-political` — health policy & regulatory news
- `stat-hypothesis-testing` — deeper statistical literacy on RCTs and meta-analyses
- `stat-causal-inference` — for causation claims in observational studies
- `hum-source-criticism` — source vetting frameworks
---
## Limitations
- **Does not verify against live clinical trial databases or medical journals.** This skill flags claims that need verification; it does not query ClinicalTrials.gov, PubMed, or Cochrane. Editor/user must verify or supply.
- **Jurisdictional scope**: legal references (PDPA, 醫療法, 醫療廣告法) reflect Taiwan law. For US (HIPAA), EU (GDPR), UK, HK, or PRC contexts, principles still apply but substitution required.
- **WHO Suicide Reporting Guidelines apply globally**, but enforcement and consequence differ by jurisdiction. Always apply these guidelines regardless of country of publication.
- **Statistical literacy**: judges poll quality, p-hacking risk, effect-size interpretation — *partially* covered in `references/medical_evidence_reading.md`, but for deep methodological critique use `stat-hypothesis-testing` or `grad-survey-design`.
- **Not a substitute for legal review.** Health-news pieces with named patients, drug/treatment criticism, or regulatory disputes should have attorney review before publication.
- **Does not generate original reporting.** If supplied material is insufficient or contradictory, this skill flags gaps — it will not source additional interviews or find additional studies.