為什麼文獻回顧是論文的根基？
多數初學者把文獻回顧當成「交代有人做過什麼」——這是錯的。文獻回顧的真正目的是建構研究 gap、證明你的研究是 novel 的、並展現方法學上的判斷力。一個好的文獻回顧 = 一個有說服力的 Introduction。
從審稿人的角度看，文獻回顧揭露三件事：(1) 你是否真的讀懂該領域？(2) 你能否批判而非僅羅列？(3) 你的研究問題在邏輯上是否真的接續既有文獻？這三題若答不出來，再漂亮的數據都救不了論文。
Beginners often treat the literature review as "listing who did what." That's wrong. The real purpose is to construct the research gap, demonstrate that your work is novel, and show methodological judgment. A strong literature review equals a persuasive Introduction.
From a reviewer's perspective, a literature review answers three questions: (1) Do you actually understand the field? (2) Can you critique rather than merely list? (3) Does your research question logically follow from prior work? If these are unanswered, no amount of pretty data will save the paper.
一、三種文獻回顧類型
並非所有「回顧」都長一樣。選錯類型，意味方法學標準、報告規範、可投稿期刊的範圍都不同。
Not all "reviews" are the same. Choosing the wrong type means different methodological standards, reporting guidelines, and target journals.
Narrative Review
傳統敘事型，由領域專家依個人視角綜合既有研究。範圍廣、無正式 protocol、不要求窮盡式檢索。
適用：論文 Chapter 2、短篇 commentary、教科書章節。
限制：選擇偏誤高、不可複製。
The traditional form: domain experts synthesize prior work from their perspective. Broad scope, no formal protocol, no exhaustive search required.
Use for: Thesis Chapter 2, short commentary, textbook chapters.
Limits: High selection bias; not replicable.
Systematic Review
遵循 PRISMA 2020 protocol，預先註冊（PROSPERO）、窮盡式檢索、可複製。常與 meta-analysis 搭配做量化合成。
適用：臨床證據合成、藥物療效評估、Cochrane review。
限制：耗時 6–18 個月、需多人協作。
Follows the PRISMA 2020 protocol — pre-registered (PROSPERO), exhaustive search, fully replicable. Often paired with meta-analysis for quantitative synthesis.
Use for: Clinical evidence synthesis, drug efficacy, Cochrane reviews.
Limits: 6–18 months; requires a team.
Scoping Review
遵循 JBI / Arksey-O'Malley 方法學，目的是繪製領域地圖——回答「目前做了什麼？」而非「X 有沒有效？」。
適用：新興領域、概念尚未統一、需識別研究 gap。
限制：不做品質評估、不做量化合成。
Follows JBI / Arksey-O'Malley methodology to map the landscape — answers "what's been done?" not "does X work?"
Use for: Emerging fields, unsettled concepts, identifying gaps.
Limits: No quality appraisal; no quantitative synthesis.
二、PubMed / MeSH 檢索策略
檢索能力決定你能找到什麼證據。關鍵是：MeSH terms（受控詞彙）+ 自由文字（free text, [tiab]）的組合，並用 Boolean 運算子 (AND / OR / NOT) 串接。
主要資料庫：PubMed（生醫主力）、Embase（歐洲藥物文獻較全）、Web of Science（跨領域、引用追蹤）、Scopus（與 WoS 互補）、Cochrane（臨床試驗）、Google Scholar（補遺、找灰文獻）。Systematic review 通常要求至少 3 個資料庫。
Search skill determines what evidence you can find. The key: combine MeSH terms (controlled vocabulary) with free text ([tiab]), joined by Boolean operators (AND / OR / NOT).
Main databases: PubMed (biomedical mainstay), Embase (European drug literature), Web of Science (cross-discipline + citation tracing), Scopus (complements WoS), Cochrane (clinical trials), Google Scholar (gap-filling + grey literature). Systematic reviews typically require at least 3 databases.
# Topic: omega-3 supplementation for mild cognitive impairment ("Fatty Acids, Omega-3"[Mesh] OR "omega-3"[tiab] OR "omega 3"[tiab] OR "n-3 PUFA"[tiab] OR "DHA"[tiab] OR "EPA"[tiab]) AND ("Cognitive Dysfunction"[Mesh] OR "mild cognitive impairment"[tiab] OR "MCI"[tiab] OR "cognit*"[tiab]) AND ("Randomized Controlled Trial"[ptyp] OR "clinical trial"[tiab]) NOT ("animals"[Mesh] NOT "humans"[Mesh]) # Filters applied via UI: 2015-2025, English, Humans, Adult 19+ # * = wildcard (truncation); [tiab] = title/abstract; [Mesh] = MeSH heading # [ptyp] = publication type
# Embase uses Emtree (its own controlled vocabulary), not MeSH ('omega 3 fatty acid'/exp OR 'docosahexaenoic acid'/exp OR 'eicosapentaenoic acid'/exp OR omega NEXT/1 3:ti,ab) AND ('mild cognitive impairment'/exp OR mci:ti,ab OR cognit*:ti,ab) AND ('randomized controlled trial'/exp OR random*:ti,ab) AND [humans]/lim AND [2015-2025]/py AND [english]/lim # /exp = explode (include narrower terms); :ti,ab = title/abstract # NEXT/1 = adjacent within 1 word; /lim = limit; /py = publication year
# Google Scholar — limited Boolean, no MeSH; use for gap-filling ("omega-3" OR "DHA" OR "EPA") AND ("mild cognitive impairment" OR "MCI") AND ("randomized" OR "RCT") -animal -mouse -rat # UI filters: custom date range 2015–2025; sort by relevance # Tip: use "Cited by" and "Related articles" for citation chasing # Limit: only first 1000 results returned; not exhaustive
| 運算子 | 作用 | 例 |
|---|---|---|
| AND | 兩個概念都需出現（縮小結果）Both concepts required (narrows results) | cancer AND immunotherapy |
| OR | 任一同義詞出現即可（擴大結果）Any synonym works (broadens results) | "NSCLC" OR "non-small cell lung cancer" |
| NOT | 排除某詞（慎用，易過濾掉相關文獻）Excludes a term (use cautiously — drops relevant hits) | cognition NOT review[ptyp] |
| * (wildcard) | 截詞，匹配多種詞尾Truncation — matches multiple endings | cognit* → cognitive, cognition… |
| " " | 片語精確匹配Exact phrase match | "randomized controlled trial" |
| [Mesh] / [tiab] | PubMed 欄位標籤：受控詞 / 標題與摘要PubMed field tags: controlled term / title & abstract | "Diabetes Mellitus"[Mesh] |
三、PRISMA 2020 流程
PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 是 systematic review / meta-analysis 的國際報告規範。2020 版取代 2009 版，加入了報告偏誤、自動化工具、註冊資訊等要求。
PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) is the international reporting standard. The 2020 version replaces the 2009 version, adding requirements for reporting bias, automation tools, and registration.
Identification
從資料庫(PubMed, Embase, Cochrane…) 與其他來源(citation chasing、會議論文、灰文獻、領域專家推薦) 撈出所有可能相關的紀錄。
需報告：每個資料庫命中數、去除重複後數量、自動化工具篩除的數量。
Pull all potentially relevant records from databases (PubMed, Embase, Cochrane…) and other sources (citation chasing, conference papers, grey literature, expert referrals).
Report: hits per database, number after de-duplication, records removed by automation tools.
Screening
標題 + 摘要篩選。兩位獨立審查者依預先定義的 inclusion / exclusion criteria 進行；不一致時由第三人仲裁或討論共識。
需報告：篩除數量與主要排除原因（如：非人類研究、非 RCT、不符合 PICO）。
Title + abstract screening. Two independent reviewers apply pre-specified inclusion / exclusion criteria; disagreements are resolved by a third reviewer or consensus.
Report: number excluded and main reasons (e.g., non-human, non-RCT, off-PICO).
Eligibility (Full-text)
對通過 screening 的論文取得全文閱讀，依 criteria 判斷是否納入。
需報告：每篇被排除的論文 ID + 排除理由（這是 PRISMA 的硬性要求，常被忽略）。
Retrieve full text for screened-in papers and judge eligibility against the criteria.
Report: ID + reason for every excluded paper at this stage (a hard PRISMA requirement, often skipped).
Included
最終納入合成的研究數量。可進一步分為「納入質性合成」與「納入量化合成 (meta-analysis)」兩個子數字。
需報告：總納入數、研究特徵摘要表 (Table 1)、品質評估 (RoB 2 / ROBINS-I / Newcastle-Ottawa)。
Final number of studies included for synthesis. Often split into "included in qualitative synthesis" and "included in quantitative synthesis (meta-analysis)."
Report: total included, study characteristics table (Table 1), quality appraisal (RoB 2 / ROBINS-I / Newcastle-Ottawa).
| PRISMA 2009 | PRISMA 2020 | |
|---|---|---|
| 檢核項 | 27 | 27 (內容更詳盡expanded content) |
| 流程圖階段 | 4 階段、單一來源4 stages, single source | 4 階段、區分「資料庫」與「其他來源」雙欄位4 stages, separate columns for "databases" vs "other sources" |
| 自動化工具 | 未提及Not mentioned | 需報告使用 AI / ML / Rayyan 等工具Must report use of AI / ML / Rayyan etc. |
| 報告偏誤 | 建議Suggested | 必填（funnel plot、Egger's test、缺失資料分析）Required (funnel plot, Egger's test, missing-data analysis) |
| 預先註冊 | 建議Encouraged | 需報告 PROSPERO / OSF 註冊碼或說明未註冊原因Must report PROSPERO / OSF registration ID or explain why not |
| 證據確定性 | 未強制Not required | 建議使用 GRADEGRADE recommended |
四、引用追蹤與 citation tools
關鍵字檢索常會漏掉領域核心論文（搜尋詞不一致、新概念尚未進入 MeSH）。引用追蹤 (citation chasing) 是 PRISMA 認可的補充方法。
Keyword search often misses seminal papers (vocabulary mismatch, new concepts not yet in MeSH). Citation chasing is a PRISMA-recognized complement.
🌳 找到該領域所有核心論文的流程
五、好壞文獻回顧句型對照
❌ 空洞泛論
「許多研究顯示，腸道菌與肥胖有關。」
(「許多研究」=多少？指誰？讀者要憑空相信。)
"Many studies have shown that gut microbiota is related to obesity."
("Many" = how many? Whom? The reader is asked to take it on faith.)
✅ 具體有出處
「自 Turnbaugh 等人 (2006, Nature) 首次報告肥胖小鼠的 Firmicutes/Bacteroidetes 比例升高以來，至少 14 項人體研究 (n > 50) 重複了此關聯，但 Sze & Schloss (2016) 的 meta-analysis 指出此關聯在跨族群分析中並不穩健 [3,7,12]。」
"Since Turnbaugh et al. (2006, Nature) first reported elevated Firmicutes/Bacteroidetes ratios in obese mice, at least 14 human studies (n > 50) have replicated this association; however, the meta-analysis by Sze & Schloss (2016) showed it does not hold across populations [3,7,12]."
❌ 流水帳列舉
「Smith (2018) 做了 X。Lee (2019) 做了 Y。Wang (2020) 做了 Z。」
(三位作者並列，沒整合、沒批判、沒指出 gap。)
"Smith (2018) did X. Lee (2019) did Y. Wang (2020) did Z."
(Three names listed in parallel — no synthesis, no critique, no gap identified.)
✅ 整合 + 指出 gap
「儘管 Smith (2018) 與 Lee (2019) 在體外系統證實 X 蛋白質參與細胞遷移，Wang (2020) 在小鼠模型中卻未能重現此現象——三項研究皆未在人類原代細胞檢驗，本研究即填補此空缺。」
"Although Smith (2018) and Lee (2019) showed in vitro that protein X drives cell migration, Wang (2020) failed to reproduce this in mouse models. None of the three tested human primary cells — a gap this study addresses."
❌ 過度依賴單一來源
「根據 Wikipedia / 一篇 2005 年的綜述…」
(Wikipedia 不是學術引用；10 年前的綜述在快速領域已過時。)
"According to Wikipedia / a 2005 review…"
(Wikipedia is not a citable source; a 10-year-old review is outdated in fast-moving fields.)
✅ 近期同儕審查文獻
「最新的綜述 (Garcia 等人, 2024, Nat Rev Genet) 整合了 2018–2023 年間 87 項 scRNA-seq 研究，指出 cell-type annotation 一致性仍低於 70%。」
"The most recent review (Garcia et al., 2024, Nat Rev Genet) synthesized 87 scRNA-seq studies (2018–2023) and reported cell-type annotation concordance below 70%."
📝 自我檢測
1. PRISMA 2020 流程的四個階段，正確順序為何？
1. What is the correct order of the four PRISMA 2020 stages?
2. 關於 systematic review 與 scoping review 的差異，下列何者最正確？
2. Which best describes the difference between a systematic review and a scoping review?
3. 以下哪個 PubMed 檢索字串最正確、能找到「以 RCT 評估免疫治療於 NSCLC 的研究」？
3. Which PubMed query best retrieves "RCTs evaluating immunotherapy in NSCLC"?