literature-review

name

literature-review

description

Literature-review METHODOLOGY knowledge: how to design search strategies, frame RQs (PICO/PEO/SPIDER), choose databases, set inclusion criteria, screen in two stages, snowball, and synthesize. Use for HOW-TO/conceptual questions about reviewing. To actually RUN a review use the /lit-review or /slr commands. Triggers on: 'literature review', 'systematic review', 'research question', 'PICO', 'inclusion criteria', 'screening protocol', 'literature search', 'search strategy', 'bibliographic database', 'synthesis', 'what is SCIE/ESCI/Scopus', 'quartile meaning', 'predatory journal signs'.

Literature Review — Full-Cycle Workflow

This skill guides you through the literature review pipeline: from research question to final synthesis. The approach is systematic by default but scales to rapid scoping reviews. Bilingual (EN/TR) search is built in; duplicate detection handles multilingual result sets.

The Workflow (9 Steps)

1. Define Research Questions (RQs)

Start with 1–4 clear, specific research questions. Each should be:

• Answerable: your literature search can address it.
• Specific enough: not "What is machine learning?" but "What ML architectures are most effective for time-series anomaly detection in IoT systems?"
• Framed for downstream use: each RQ will tag a subset of evidence cards; drafts will retrieve cards by RQ tag.

Example RQs:

• RQ1: What deep learning architectures have been applied to bearing fault diagnosis?
• RQ2: What datasets exist for bearing fault diagnosis benchmarking?
• RQ3: How does CNN performance compare to traditional signal-processing baselines?

2. Select Framing Methodology

Choose one, based on your question type and paper type:

Framing	Best for	Output
PICO (Population, Intervention, Comparison, Outcome)	Clinical/health studies, RCTs, observational studies	Inclusion/exclusion criteria for patient populations, interventions, and measured outcomes
PEO (Population, Exposure, Outcome)	Observational studies, epidemiology, social sciences	Criteria for who is studied, what exposure/exposure variation, what outcomes
SPIDER (Sample, Phenomenon, Design, Evaluation, Research type)	Qualitative studies, exploratory research, theory development	More flexible framing for diverse methodologies
Concept-based (concept 1 × concept 2 × context)	Systematic reviews of theoretical concepts, scoping reviews	3-dimensional keyword matrix: concept A (terms), concept B (terms), context/population (terms)

For the bearing fault example (mixed methods): Concept-based with three dimensions:

• Concept 1: fault detection methods (CNN, autoencoders, signal processing, …)
• Concept 2: bearing types / fault types (rolling-element, journal, …)
• Context: industrial IoT, condition monitoring, vibration analysis

3. Build a Dual-Language Keyword Matrix

See bilingual-search.md for the full protocol. In brief:

Concept 1 (Methods):
  EN: CNN, deep learning, neural network, autoencoder, signal processing, …
  TR: derin öğrenme, yapay sinir ağı, sinyal işleme, …

Concept 2 (Fault types):
  EN: bearing fault, anomaly detection, failure mode, …
  TR: rulman arızası, anormallik tespiti, arıza modu, …

Context:
  EN: condition monitoring, IoT, predictive maintenance, …
  TR: durumsal izleme, IoT, öngörülü bakım, …

Decision on search languages: always produce EN+TR matrices. The bilingual-search skill rule: if a topic is Turkey-local (Turkish education policy, Turkish health legislation, Turkish dialect research), search TR sources mandatory. If topic is international (deep learning, COVID-19 vaccines), still search TR for completeness but expect EN to dominate.

4. Select Databases & Apply Search Strategy

Standard academic databases:

• Multidisciplinary: Google Scholar, Semantic Scholar, Crossref, OpenAlex
• Computer science/engineering: arXiv, DBLP, IEEE Xplore, ACM Digital Library
• Life sciences: PubMed, PubMed Central, EMBASE
• General: JSTOR, ProQuest, EBSCO

Turkish-specific sources (see turkish-databases.md):

• TR Dizin (trdizin.gov.tr) — indexes Turkish academic journals
• DergiPark (dergipark.org.tr) — open-access Turkish journals
• YÖK Thesis Center (tez.yok.gov.tr) — all Turkish theses (BSc, MSc, PhD)

Use the scout agent (literature-scout) to run multi-database searches in parallel. The scout will:

• Log every search query, database, date, and hit count
• De-duplicate across sources (DOI-first, then title-normalized matching)
• Enrich results with venue index status (SCI-E, Scopus, TR Dizin, etc.)
• Return a ranked candidate list (by relevance + venue quality)

5. Two-Stage Screening

Stage 1 (Title/Abstract): You review candidate list (typically 50–300 hits) and mark:

• ✓ Include: title/abstract clearly relevant; advance to full text.
• ✗ Exclude: clearly out of scope; record reason (e.g., "Not peer-reviewed", "Different population", "Predatory journal flagged").
• ? Unclear: advance to full-text review for decision.

Stage 2 (Full-Text): For each included paper, the source-analyst agent reads the PDF and:

• Confirms relevance against inclusion/exclusion criteria
• Extracts structured evidence cards (findings, methods, datasets, metrics, limitations)
• Scores methodological rigor (1–5 strength rating)
• Flags conflicting findings

6. Forward & Backward Snowballing (Optional)

Backward snowballing: Look at references in your included papers. Are there seminal papers or highly-cited sources you missed?

Forward snowballing: Use Google Scholar or Semantic Scholar to find papers citing your key included papers. Are recent papers building on them?

Snowballing is most valuable in rapidly-evolving fields (ML, COVID-19 literature) or when your database searches yield surprisingly few hits.

7. Saturation & Stopping Criterion

Continue searching and screening until:

• Knowledge saturation: new papers contribute no new concepts, methods, or findings; you're seeing the same themes repeated.
• Study saturation: the evidence cards show consistent findings; conflicting findings have been identified and characterized.
• Practical limit: you've reviewed 50–200 papers (typical for narrative reviews) or 100+ papers (systematic reviews). For large datasets (500+ papers), use staged analysis: screen a batch, synthesize, decide if more searching is needed.

For systematic reviews, use the PRISMA flow diagram (see prisma.md): track identified / de-duplicated / screened / excluded (with reason) / included. The diagram communicates transparency.

8. Synthesis (The Hard Part)

See synthesis.md for deep guidance. In brief:

Synthesis ≠ annotated bibliography. An annotated bibliography lists papers and summarizes each. Synthesis integrates findings across papers: it identifies patterns, themes, contradictions, and gaps.

Patterns to look for:

• Thematic: group papers by concept (e.g., "CNN architectures", "transfer learning strategies", "explainability methods").
• Chronological: evolution of the field (early classical methods → neural networks → recent transformer-based approaches).
• Methodological: what methods dominate? What methodological gaps exist?
• Empirical: do all studies agree on dataset difficulty? Metric hierarchy? Baseline performance?
• Conflict: what findings contradict? Why? (different populations, datasets, evaluation protocols, publication bias?)

Rule: ≥2 sources per paragraph. If you cite only one source in a paragraph, you're summarizing, not synthesizing. The minute you cite a second source, you create a synthesis: comparison, confirmation, or conflict.

9. Write Up & Document

Outcomes:

• sources/registry.jsonl — all screened sources, with status (included/excluded/duplicate) and DOI/venue metadata.
• evidence/cards.jsonl — structured findings, methods, datasets from included papers, tagged by RQ.
• search-log.md — (if systematic) every query run, by database, hit count, date.
• PRISMA flow diagram — (if systematic) counts at each stage.
• literature-synthesis.md — thematic / chronological / methodological narrative tying papers together.

---

PICO/PEO/SPIDER Templates (Quick Reference)

PICO (for clinical/health studies)

Population:       Demographics, disease, health condition. E.g., "adults 18–65 with Type 2 diabetes"
Intervention:     The treatment/exposure being studied. E.g., "metformin therapy"
Comparison:       The alternative or control. E.g., "placebo" or "standard care"
Outcome:          Measured result. E.g., "HbA1c reduction", "quality of life"
Study design:     RCT, cohort, case-control. (Often implicit in RQ.)
Timeframe:        Duration of intervention and follow-up.

Inclusion criteria:
  - Population: age, comorbidities, disease stage
  - Intervention: drug, dosage, duration
  - Comparison: required or optional?
  - Outcomes: primary + secondary
  - Study design: RCT only, or observational too?
  - Language: EN only, or others?
  - Date range: e.g., 2015–present

Exclusion criteria:
  - Opinion pieces, editorials, non-peer-reviewed
  - Populations with confounding conditions
  - Outcomes not measuring clinical efficacy (e.g., only lab markers)
  - Conferences proceedings (depends on field; keep separate)

PEO (for observational/exposure studies)

Population:       Who is being studied?
Exposure:         What is the risk factor / exposure of interest?
Outcome:          What is the measured result?

Typically used in epidemiology, environmental health, social sciences.

SPIDER (for qualitative / exploratory)

Sample:           Who participated? How many? Sampling strategy?
Phenomenon:       What is being investigated? Central concept?
Design:           Study methodology (grounded theory, phenomenology, ethnography)?
Evaluation:       How was quality assessed? (rigor, credibility, transferability?)
Research type:    Qualitative, mixed-methods, theoretical?

More flexible than PICO; allows exploratory, iterative research.

---

Handling Bilingual Results

See bilingual-search.md for full protocol. Key rules:

1. Always build dual matrices: even if you expect EN results to dominate, building a TR matrix is the search discipline. It surfaces Turkish-language gaps or hidden TR expertise.

2. Deduplication (critical): when you search both EN and TR, you'll find the same paper twice—once in English, once translated to Turkish. De-duplicate via:

   • DOI-first match (if both have DOI, same DOI = duplicate)
   • Title-normalized matching (normalize whitespace + diacritics + case, check ≥95% similarity)

3. Track language in records: every registry entry and card logs lang: "en" or lang: "tr". In synthesis, note when findings come from EN-only vs TR-only sources (a finding known only in Turkish is noteworthy).

---

Citations & Venue Quality

Before including a source:

• Run /journal-check to classify venue (SCI-E, Scopus, TR Dizin, predatory-flagged).
• Evidence from high-quality venues (SCI-E Q1, peer-reviewed) is weighted higher than pre-prints or low-quality outlets.
• See venue-quality.md for index meanings and red flags.

---

Common Mistakes (Avoid These)

Mistake	Fix
Searching only one database	Use ≥3 (e.g., Semantic Scholar + Google Scholar + IEEE Xplore) for engineering topics
Using only English keywords	Build dual matrix; many non-English papers use EN abstracts + TR titles, easily missed
Including non-peer-reviewed or predatory sources	Check /journal-check; exclude predatory flags automatically
Stopping at Title/Abstract screening	Many false negatives; papers with vague abstracts need full-text review
Synthesis = list of papers	Every paragraph must synthesize ≥2 sources; if listing, it's annotation, not synthesis
Ignoring conflicting findings	Conflicts are the interesting part; acknowledge & explain in Discussion
No search log (for systematic reviews)	Search-log.md is mandatory for reproducibility; list queries, dates, hit counts

---

See Also

• bilingual-search.md — EN/TR decision matrix, transliteration, dedup rules
• turkish-databases.md — TR Dizin, DergiPark, YÖK Thesis Center: coverage, query patterns, metadata quirks
• venue-quality.md — SCI-E/SSCI/AHCI/ESCI/Scopus/TR Dizin explained; quartiles; predatory red flags; /journal-check usage
• prisma.md — PRISMA 2020 flow discipline; systematic review workflows
• synthesis.md — thematic / chronological / methodological synthesis patterns; "synthesis ≠ annotated bibliography" rule
• literature-scout agent — automates multi-database search + venue enrichment + dedup
• evidence_db.py query --tags rq1 — retrieve cards for drafting, filtered by RQ