AI-Powered Systematic Literature Review Workflows for Researchers

# AI Tool Workflow for mHealth/T2D SLR in LMICs

## Stage 1: Search & Discovery

- **Semantic Scholar (free)**: Citation graph exploration across 200M+ papers. Useful for finding seed papers on "mHealth + T2D + LMIC" and tracing influential work. **Strength**: free API, good coverage of global health literature. **Limitation**: not a replacement for structured PubMed/Scopus/Embase Boolean searches required for PRISMA.
- **ResearchRabbit (free)**: Visual citation mapping ("Similar Work," "Earlier/Later Work"). Helps identify grey literature and LMIC-specific studies that keyword searches miss. **Limitation**: search paths are not easily reproducible — use only to *supplement* database searches, and document seed papers in your PRISMA log.

## Stage 2: Screening & Deduplication

- **Rayyan (freemium; free tier sufficient)**: Purpose-built for SLR screening with AI-assisted relevance ranking after ~50 manual decisions. Handles ~3,000 abstracts easily, supports dual-blind screening, conflict resolution, and exports PRISMA flow data. Deduplicates across PubMed/Scopus/Embase RIS exports. **Strength**: PRISMA-aligned, transparent audit trail. **Limitation**: AI suggestions must not auto-exclude records — keep "human decides" mode on for compliance.
- **Covidence (paid, ~$200+/yr — likely over budget)**: Better extraction features than Rayyan but exceeds $30/month. Skip unless institution provides access.

## Stage 3: Full-Text Analysis & Extraction

- **Elicit ($12/month Plus plan)**: Extracts PICO elements, outcomes, sample sizes, and intervention details into structured tables — ideal for comparing heterogeneous mHealth interventions (SMS, apps, IVR) across LMIC settings. **Strength**: saves hours on ~50–100 included full texts. **Limitation**: can hallucinate numeric values; **every extracted data point must be human-verified against the PDF** or it breaks PRISMA data-integrity requirements.
- **SciSpace (freemium; ~$12–20/month)**: Alternative with a strong "Chat with PDF" feature for parsing methods sections and risk-of-bias details. Useful when papers use varied outcome measures (HbA1c, fasting glucose, adherence).

## Stage 4: Synthesis & Writing

- **Claude (free tier or Pro $20/month)**: Long-context window (200K tokens) handles synthesizing 20–30 full papers simultaneously — useful for drafting narrative synthesis across LMIC subgroups (Sub-Saharan Africa vs. South Asia). **Strength**: strong at nuanced comparison and identifying heterogeneity. **Limitation**: cannot access live databases; do not let it generate citations (fabrication risk).
- **Consensus (freemium; Premium $9/month)**: Quickly checks consensus on specific claims (e.g., "Does SMS reduce HbA1c in T2D?"). Good for identifying counter-evidence during discussion section. **Limitation**: results reflect indexed literature only — not a substitute for your own included studies.

## Stage 5: Citation & Reporting

- **Scite ($20/month — choose this OR Elicit, not both, to stay under budget)**: Shows whether cited studies have been supported or contradicted by later work. Critical for a 2015–2024 review where early mHealth RCTs may have been superseded. **Strength**: strengthens discussion/limitations. **Limitation**: classifications are automated and occasionally mislabeled — spot-check key citations.
- **Zotero + ChatGPT free tier**: Manage references in Zotero; use ChatGPT only for formatting/language polishing, never citation generation.

## Recommended 5-Step Workflow (~$24–32/month)

1. **Scope & search**: Build Boolean strings in PubMed/Scopus/Embase; use ResearchRabbit + Semantic Scholar to validate completeness and find LMIC grey literature.
2. **Screen**: Import all records to **Rayyan**, deduplicate, title/abstract screen with AI ranking assist (dual reviewer, human-final).
3. **Extract**: Use **Elicit** to pre-populate extraction tables (PICO, HbA1c outcomes, intervention type, country income tier); verify every cell manually.
4. **Synthesize**: Upload included PDFs to **Claude Pro** for thematic synthesis and heterogeneity analysis by region/intervention type; run specific claims through **Consensus**.
5. **Report & cite**: Manage references in Zotero; use **Scite** to audit citation reliability; generate PRISMA 2020 flow diagram from Rayyan exports.

**Budget**: Rayyan (free) + Elicit ($12) + Claude Pro ($20) = $32, *or* swap Claude for ChatGPT free to hit $12.

## PRISMA Compliance Note

AI-assisted screening and extraction are permitted under PRISMA 2020 **only if** (a) decisions are human-verified, (b) the tool, version, date, and prompts are documented in your methods, and (c) AI is not used to exclude records autonomously. Log all AI touchpoints in your protocol (register on PROSPERO before screening begins).

# AI Tool Workflow for mHealth Type 2 Diabetes SLR in LMICs

## 1. Search & Discovery

**Research Rabbit** (Freemium)
- Visualizes citation networks and paper relationships; automatically generates similar paper suggestions from your seed papers
- Ideal for discovering mHealth diabetes papers you might miss via traditional database filters
- Strengths: Maps knowledge gaps visually; identifies clusters of related work without manual screening
- Limitations: Relies on indexed papers (may miss non-English publications common in LMICs); cannot export PRISMA-compliant search strategies directly
- Pricing: Free tier sufficient for initial mapping

**Semantic Scholar** (Free)
- AI-powered paper indexing across PubMed, Scopus, and preprints; extracts key claims and methodology automatically
- Strengths: Superior filtering by study type (RCTs, observational) and relevance scoring; excellent for narrowing 3,000 abstracts pre-screening
- Limitations: Citation counts sometimes inconsistent with Scopus; lacks formal search syntax documentation for reproducibility
- Pricing: Free

**Recommendation**: Use Research Rabbit to map citation networks from 5-10 seed mHealth diabetes papers, then use Semantic Scholar's advanced filters (study type, year, country income classification) to refine your 3,000-paper corpus before screening.

---

## 2. Screening & Deduplication

**Covidence with AI-Assisted Screening** (Paid)
- Native deduplication and abstract screening with machine learning that learns your inclusion/exclusion criteria after ~50 manual screens
- Ideal for your 3,000 abstract volume; supports PRISMA flow diagrams automatically
- Strengths: Audit trail for reproducibility; inter-rater reliability metrics built-in; exports PRISMA-compliant documentation
- Limitations: Steep learning curve; AI recommendations require calibration; cost ~$25-40/month per user
- Pricing: $30-40/month (exceeds budget slightly but justifiable for 3,000 abstracts)

**Rayyan (by QCR International)** (Free)
- Lightweight screening platform with machine learning prioritization (highlights likely-relevant abstracts first); handles deduplication
- Strengths: No cost; QCRI-backed (reproducibility standard); suitable for 3,000 abstracts; transparent methodology
- Limitations: ML model less sophisticated than Covidence; slower interface; no built-in PRISMA output generation
- Pricing: Free

**Recommendation**: Start with **Rayyan** to stay within budget. After screening ~200 abstracts manually, enable ML prioritization to rank remaining 2,800 abstracts. This reduces screening burden by ~30-40% while maintaining transparency.

---

## 3. Full-Text Analysis & Extraction

**SciSpace (Copilot)** (Freemium)
- AI assistant that reads PDFs, extracts study characteristics (design, sample size, interventions, outcomes) via chat interface
- Ideal for quickly extracting data from 100-200 included full texts; supports custom extraction templates
- Strengths: Reduces manual data extraction time by 50%; handles non-English abstracts (useful for LMIC studies); free tier covers 10 papers/month
- Limitations: Cannot batch process; accuracy decreases for poorly scanned PDFs or non-standard formats; requires quality control checking
- Pricing: Free for limited use; $10/month for 50 papers

**Elicit** (Freemium)
- Automates extraction of quantitative findings, population characteristics, and methodology from full texts
- Strengths: Outputs structured tables directly; excellent for meta-analysis preparation; supports custom search within papers
- Limitations: Better for RCTs than mixed-method studies; sometimes misses context-dependent outcomes relevant to LMIC settings
- Pricing: Free tier (5 papers/month); $15/month for unlimited

**Recommendation**: Use **SciSpace** for initial extraction ($10/month), with manual spot-checking of 20% of papers to validate accuracy. Reserve Elicit for outcome-specific extraction only if conducting meta-analysis.

---

## 4. Synthesis & Writing

**Claude 3.5 Sonnet** (via Claude.ai or API, ~$10-15/month)
- Generates narrative synthesis by summarizing study characteristics, grouping findings by intervention type or population
- Strengths: Produces draft tables synthesizing effectiveness across LMIC contexts; identifies intervention heterogeneity
- Limitations: Cannot independently assess quality; prone to hallucinating effect sizes if not given extracted data as input
- Pricing: $20/month for Claude Pro (or pay-per-token on API, ~$10-15 for SLR volume)

**Recommendation**: Feed Claude your extracted data table (from SciSpace/Elicit) with explicit instructions to synthesize by: (1) intervention type (SMS, app, wearable), (2) country income level, (3) HbA1c/glycemic control outcomes. Use output as draft for your narrative synthesis section.

---

## 5. Citation & Reporting

**Scite** (Freemium)
- Maps citation context (supporting vs. contradicting evidence); flags papers with citation problems
- Strengths: Ensures your citations are accurately used; identifies weak evidence in your included studies
- Limitations: Limited free tier (20 citations/month)
- Pricing: Free tier; $15/month for full access

**Zotero** (Free) + **PRISMA 2020 Checklist** (Free)
- Reference management and PRISMA compliance tracking
- Pricing: Free

**Recommendation**: Zotero for reference management throughout; use Scite's free tier spot-check on 20 high-impact papers in your synthesis.

---

## Integrated 5-Step Workflow

1. **Search phase** (Week 1): Semantic Scholar + Research Rabbit → curated 3,000-paper corpus with documented search strategy
2. **Deduplication + Screening** (Weeks 2-4): Rayyan with ML prioritization → ~100-150 included full texts
3. **Data Extraction** (Weeks 5-6): SciSpace + 20% manual QC → structured extraction table
4. **Synthesis** (Week 7): Claude 3.5 + your extraction table → draft narrative synthesis grouped by intervention type and LMIC context
5. **Reporting** (Week 8): PRISMA 2020 checklist + Zotero → finalized SLR with reproducible search strategy documented

---

## PRISMA Compliance Note

**Critical**: Rayyan and SciSpace lack native PRISMA output generation. Mitigate by: (1) documenting search strategies manually in a supplementary table, (2) using Covidence's PRISMA export *if* budget allows mid-review, (3) manually creating PRISMA flow diagram in Lucidchart (free tier). Ensure inter-rater reliability data and screening criteria are logged for transparency.

**Total estimated cost: $20-30/month** (Rayyan free + SciSpace $10 + Claude $15), fitting your budget.

## 1. Search & Discovery

- **Elicit** *(freemium)*
  - **What it does:** Uses semantic search and structured paper summaries to help refine search terms, identify key papers, and surface related studies.
  - **Why it fits this project:** Useful early on for expanding terms around *mHealth*, *type 2 diabetes*, and LMIC synonyms before running formal database searches in PubMed/Scopus/Embase.
  - **Strengths:** Good for query refinement; can help identify intervention/outcome terminology you may miss; supports quick relevance checking.
  - **Limitations:** Not a replacement for database searching; coverage is not identical to PubMed/Scopus/Embase, so results are **not reproducible enough to serve as the primary PRISMA search source**.
  
- **Research Rabbit** *(free)*
  - **What it does:** Citation-network discovery tool that maps related papers, authors, and clusters.
  - **Why it fits this project:** Helpful for backward/forward citation chasing after your core set of diabetes mHealth studies is identified.
  - **Strengths:** Strong for finding adjacent studies and influential papers from LMIC/public health networks.
  - **Limitations:** Best for supplementary searching only; discovery logic is not fully transparent, so **document it as citation searching, not main search strategy**.

## 2. Screening & Deduplication

- **Rayyan** *(free / paid tiers)*
  - **What it does:** AI-assisted title/abstract screening with tagging, conflict resolution, and duplicate handling.
  - **Why it fits this project:** Screening ~3,000 abstracts is exactly Rayyan’s sweet spot, especially on a limited budget.
  - **Strengths:** Widely used in SLRs; easy inclusion/exclusion labels; supports blinded dual screening; duplicate detection is practical.
  - **Limitations:** AI ranking is assistive, not fully auditable; deduplication is decent but not perfect across PubMed/Scopus/Embase exports—**manual spot-checking still needed for PRISMA counts**.
  
- **Covidence with AI** *(paid; usually above your budget unless institution provides access)*
  - **What it does:** End-to-end review platform with deduplication, screening, extraction, and PRISMA flow support.
  - **Why it fits this project:** If your university already subscribes, it can reduce workflow friction.
  - **Strengths:** Strong review management; more structured than Rayyan.
  - **Limitations:** Likely too expensive for $30/month; AI features are convenient but not essential. If self-paying, **Rayyan is the better value**.

## 3. Full-Text Analysis & Extraction

- **SciSpace** *(freemium)*
  - **What it does:** AI-assisted PDF reading, question answering, and table extraction from full texts.
  - **Why it fits this project:** Useful for extracting intervention components, outcomes (e.g., HbA1c), study setting, and LMIC context from full texts.
  - **Strengths:** Good for rapid orientation across many PDFs; easier than manually hunting through methods/results sections.
  - **Limitations:** Extraction can hallucinate or omit details; **do not treat outputs as final data extraction without human verification**. Reproducibility is limited because answers depend on prompts.

- **ChatGPT** *(paid optional; free version available with limits)*
  - **What it does:** Helps turn PDFs/notes into structured extraction tables, compare studies, and draft evidence matrices.
  - **Why it fits this project:** Good for standardizing extraction fields: country, population, intervention type, comparator, duration, outcomes, risk-of-bias notes.
  - **Strengths:** Flexible for creating custom extraction templates tailored to mHealth and public health interventions.
  - **Limitations:** Not a source database; can infer unsupported claims if prompted loosely. **Always verify against the article PDF and keep manual extraction records**.

## 4. Synthesis & Writing

- **Claude** *(free / paid)*
  - **What it does:** Strong long-context summarization and thematic synthesis from extracted notes.
  - **Why it fits this project:** Particularly useful for synthesizing heterogeneous intervention studies from multiple LMIC settings.
  - **Strengths:** Handles long documents well; good at producing comparative narrative summaries.
  - **Limitations:** Can sound authoritative even when wrong; **not suitable for unsupervised evidence synthesis**. Keep a traceable chain from extraction sheet to narrative claim.
  
- **Elicit** *(freemium)*
  - **What it does:** Can help organize findings by question and compare study characteristics.
  - **Why it fits this project:** Useful for grouping evidence by intervention type (SMS, app-based coaching, telemonitoring) and outcomes.
  - **Strengths:** Faster than manual clustering for broad themes.
  - **Limitations:** Again, not fully transparent enough for sole analytic basis.

## 5. Citation & Reporting

- **Scite** *(paid/freemium depending plan)*
  - **What it does:** Shows citation context—whether papers are supporting, contrasting, or mentioning prior work.
  - **Why it fits this project:** Helpful when interpreting mixed evidence on mHealth effectiveness.
  - **Strengths:** Adds nuance to discussion section and helps identify contested findings.
  - **Limitations:** Citation classifications are imperfect; **not a substitute for critical appraisal**.

- **Zotero + Zotero AI add-ons if desired** *(free)*
  - **What it does:** Reference management, PDF storage, tagging, and citation insertion.
  - **Why it fits this project:** Best free option for a PhD student; integrates well with Word and supports transparent recordkeeping.
  - **Strengths:** Free, stable, exportable library, good for reproducibility.
  - **Limitations:** Native AI is limited; you may need manual organization.

## Recommended 5-Step Workflow for Your Review

1. **Develop reproducible database searches manually** in PubMed, Scopus, and Embase; use **Elicit** only to refine keywords and synonyms for mHealth, T2D, and LMICs.
2. **Export all records to Zotero**, then import into **Rayyan** for deduplication and title/abstract screening; use dual screening if possible.
3. **Retrieve full texts** and use **SciSpace** to quickly locate intervention details and outcomes; record all final extraction in a manual spreadsheet.
4. **Use ChatGPT or Claude** to convert verified extraction notes into structured comparative summaries by intervention type, country, and outcome.
5. **Use Scite and Zotero** during writing to contextualize findings, manage references, and produce transparent reporting materials.

## PRISMA-Compliance Note

For PRISMA compliance, **AI tools should assist, not replace, documented human decisions**. Your formal search must come from named databases with saved strategies; screening/extraction decisions should be reproducible and ideally dual-checked; and any AI-assisted summaries must be verified against the original articles. On your budget, the most defensible stack is **Rayyan + Zotero + free/freemium Elicit/SciSpace**, with optional paid LLM use only during synthesis.

## 1. Search & Discovery
- **Semantic Scholar (free)**: AI-powered academic search engine; scans PubMed/Scopus-like sources for mHealth/T2D papers. Fits by surfacing 2015-2024 LMIC-focused studies via semantic queries like "mHealth type 2 diabetes low-income countries." Strengths: Free, fast relevance ranking over 200M papers. Limitations: No direct Embase integration; reproducibility via query logs only. 
- **Research Rabbit (free)**: Visual literature maps from seed papers. Fits for exploring mHealth intervention networks in public health. Strengths: Free collections/clusters for ~3K results. Limitations: Relies on user seeds; less precise for date/LMIC filters.

## 2. Screening & Deduplication
- **Rayyan (free/freemium)**: AI-assisted abstract screening/deduplication for SLRs. Fits 3K abstracts from PubMed/Scopus/Embase via RIS/CSV import; highlights inclusion criteria (e.g., mHealth effectiveness in LMICs). Strengths: Free AI prioritization (95% accuracy), blind mode, export for PRISMA. Limitations: AI decisions not fully auditable, reducing reproducibility; freemium limits collaborators.
- **Elicit (freemium, $12/mo Plus)**: Bulk abstract screening with AI extraction. Fits by ranking abstracts on PICO (effectiveness, T2D, LMICs). Strengths: Extracts key data fast. Limitations: Free tier caps 200 papers/mo; opaque AI may challenge PRISMA reporting.

## 3. Full-Text Analysis & Extraction
- **SciSpace (freemium, $12/mo Copilot)**: AI PDF analyzer/extractor. Fits for pulling outcomes (e.g., HbA1c changes) from mHealth trials. Strengths: Table extraction, Q&A on full texts. Limitations: Free tier limits uploads; AI hallucinations risk data errors, impacting reproducibility.
- **Elicit ($12/mo)**: Automates data extraction from full texts. Strengths: Custom columns for SLR (e.g., intervention type, country). Limitations: Column limits on free; non-transparent models hinder PRISMA audits.

## 4. Synthesis & Writing
- **Consensus (freemium)**: AI evidence synthesizer for health topics. Fits by meta-answering "mHealth effectiveness for T2D in LMICs?" from screened papers. Strengths: Yes/no/extraction summaries. Limitations: Free caps queries; biased toward high-citation papers.
- **Claude (free/Pro $20/mo)**: LLM for thematic synthesis. Strengths: Handles pasted excerpts for narrative review. Limitations: No citation tracking; reproducibility via prompts only.

## 5. Citation & Reporting
- **Scite (freemium, $20/mo Smart Citations)**: Citation analyzer/contextualizer. Fits PRISMA flow/export. Strengths: Supporting/contradicting labels for mHealth claims. Limitations: Free limits searches; partial for non-Scopus citations.
- **ChatGPT (free/Plus $20/mo)**: Generates PRISMA diagrams/tables. Strengths: Cheap formatting. Limitations: Hallucinations; log prompts for reproducibility.

**Recommended 5-Step Workflow**: 1) Use Semantic Scholar/Research Rabbit for initial PubMed/Scopus/Embase searches (date/LMIC filters). 2) Import to Rayyan/Elicit for dedupe/screen 3K abstracts. 3) SciSpace/Elicit for full-text extraction on ~200 includes. 4) Consensus/Claude for theme/effect synthesis. 5) Scite/ChatGPT for citations/PRISMA flow.

**PRISMA-Compliance Note**: Document all AI use/decisions in methods; tools' opacity limits full reproducibility—manually verify 20% samples. Total cost: <$25/mo (Elicit + SciSpace). (428 words)