Systematic Mapping Studies

Charting the landscape of research so we can see the forest, not just the trees.

Introduction

In an era where scholarly output doubles every nine years, researchers face a paradox of abundance: the more we know, the harder it becomes to know what we don’t know. A single keyword search on “bee health” in Scopus returns over 12,000 records spanning agriculture, ecology, genomics, and policy. In the field of artificial intelligence, a search for “self‑governing agents” yields nearly 8,000 papers, many of which are scattered across conferences, pre‑print servers, and niche journals.

Systematic Mapping Studies (SMS) are the cartographers of this scholarly terrain. Rather than diving deep into the effectiveness of a single intervention (as a systematic review does), a mapping study surveys the breadth of a domain, classifies its components, and visualises gaps and trends. The result is a high‑level, evidence‑based map that guides funding bodies, policy makers, and fellow scientists toward the most promising—or most neglected—areas. For bee conservation, such a map can reveal whether research is disproportionately focused on honeybees in Europe while wild pollinators in the Global South receive scant attention. For AI governance, a map can expose the imbalance between technical safety work and socio‑legal frameworks.

This article is a definitive, step‑by‑step guide to planning, executing, and reporting a systematic mapping study. It blends the rigor of software‑engineering standards with concrete examples from ecology and AI, and it shows how emerging AI agents can help automate parts of the workflow without sacrificing transparency. By the end, you’ll be equipped to turn a chaotic literature flood into a clear, actionable landscape.

1. What Is a Systematic Mapping Study?

A systematic mapping study (SMS) is a secondary research method that aims to:

1. Identify all primary studies that meet a predefined set of criteria.
2. Classify each study according to a taxonomy (e.g., research type, intervention, geographic region).
3. Visualise the distribution of studies across the taxonomy to highlight concentrations, trends, and gaps.

In contrast to a systematic review, which synthesises findings to answer a narrowly scoped question (e.g., “Does pesticide X reduce colony collapse disorder?”), an SMS answers a broad question such as “What research has been conducted on pesticide impacts on pollinators, and where are the knowledge gaps?”

Key distinguishing features:

The primary purpose of an SMS is strategic: it tells you where to look next. For bee conservation programmes, this might mean spotting a dearth of studies on urban pollinator networks in Southeast Asia, prompting a targeted field campaign. For AI governance, it could highlight that only 12 % of papers on “self‑governing agents” address legal accountability, suggesting a fertile area for interdisciplinary work.

2. Historical Development and Standards

The systematic mapping methodology originated in software engineering in the early 2000s. Katherine Kitchen and Brett Charters (2007) introduced the term “systematic mapping” to differentiate broad field surveys from systematic reviews. Their seminal paper laid out a six‑step process that has since become the de‑facto standard in many disciplines.

Key milestones:

These standards have been adopted beyond software engineering. In ecology, the Ecological Society of America recommends the Kitchen‑Charters workflow for biodiversity literature surveys. In AI ethics, the Partnership on AI cites PRISMA‑SM as the baseline for its annual “State of AI Governance” mapping reports.

3. Methodological Workflow

Below is a canonical workflow that can be adapted to any domain. Each step includes concrete actions, typical artefacts, and pitfalls to avoid.

3.1 Define the Research Scope

• Research Question (RQ): Formulate a broad RQ using the PCC (Population, Concept, Context) framework.
• Example (Bee health): “What research has been conducted on population‑level effects of pesticides on wild pollinators worldwide?”
• Example (AI agents): “What literature addresses self‑governing autonomous agents in real‑world deployments?”

• Inclusion/Exclusion Criteria:
• Temporal: 2000–2024 (captures modern pesticide formulations).
• Language: English, Spanish, Mandarin (to cover major bee‑research regions).
• Document type: Peer‑reviewed articles, conference papers, theses (exclude editorials).

3.2 Develop a Protocol

A protocol is a registered, immutable plan that prevents post‑hoc bias. Register on platforms like Open Science Framework (OSF) or PROSPERO (if the field permits). The protocol should contain:

1. Search strings (see Section 3.3).
2. Databases (Scopus, Web of Science, PubMed, IEEE Xplore, arXiv).
3. Screening process (titles/abstracts → full text).
4. Data extraction template (see Table 1).

3.3 Construct Search Strings

A Boolean search string must balance recall (capturing all relevant papers) and precision (excluding irrelevant hits). Use controlled vocabularies (e.g., MeSH for health, ACM CCS for computing).

Example (Bee health)

(TITLE-ABS-KEY("wild pollinator*" OR "bumblebee*" OR "solitary bee*") 
 AND TITLE-ABS-KEY(pesticid* OR "neonicotinoid*" OR "insecticide*") 
 AND PUBYEAR > 1999) 
 AND (LANGUAGE(English) OR LANGUAGE(Spanish) OR LANGUAGE(Chinese))

Example (AI agents)

(TITLE-ABS-KEY("self‑governing" OR "autonomous" OR "self‑regulating") 
 AND TITLE-ABS-KEY("agent*" OR "system*" OR "robot*") 
 AND TITLE-ABS-KEY("deployment" OR "real‑world")) 
 AND PUBYEAR > 2009

Run the string in each database, export results in BibTeX or CSV format, and merge with a de‑duplication script (e.g., refcleaner in R).

3.4 Screening and Selection

1. Title/Abstract Screening – Two independent reviewers assign include, exclude, or uncertain. Disagreements are resolved by a third reviewer.
2. Full‑Text Screening – Apply the same criteria; record reasons for exclusion (e.g., “No primary data”).

A PRISMA‑SM flow diagram (see Figure 1) visualises the numbers at each stage. In a recent mapping of AI safety literature (2022–2024), the initial search yielded 9,842 records; after de‑duplication 8,117 remained; title/abstract screening left 1,932; full‑text screening produced 527 final studies.

3.5 Data Extraction and Classification

Using the template from Table 1, extract data into a relational database (SQLite) or a spreadsheet. Classification can be manual, but for >1,000 papers, semi‑automated text classification dramatically speeds up the process (see Section 5).

Key classification dimensions (customise per field):

• Research Type: Empirical, Modeling, Review, Position Paper, Tool Development.
• Methodology: Field experiment, Lab assay, Simulation, Survey.
• Geographic Region: Continent, country, climate zone.
• Stakeholder Focus: Farmers, policymakers, beekeepers, AI developers.

3.6 Mapping and Visualisation

The final step is to map the data. Common visual tools:

• Bubble plots (research type vs. year, bubble size = number of papers).
• Heat maps (topic vs. region).
• Network maps (co‑citation or keyword co‑occurrence using VOSviewer).

Interactive dashboards (e.g., Shiny in R or Streamlit in Python) let users filter by year, region, or methodology, turning a static paper into a living resource.

4. Taxonomy and Classification Schemes

A robust taxonomy is the backbone of any mapping study. It must be exhaustive (covers all observed phenomena) and mutually exclusive (no overlap). Below are two exemplar taxonomies, one from ecology and one from AI, that illustrate how to tailor the scheme to your domain.

4.1 Bee‑Health Taxonomy

When applied to a corpus of 2,437 papers on bee stressors (2000–2024), the taxonomy revealed that 78 % of studies focused on honeybees, while only 7 % examined solitary bees. Moreover, 65 % of papers used field monitoring, but only 3 % employed genomic techniques, highlighting a methodological gap.

4.2 AI Self‑Governing Agents Taxonomy

Applying this taxonomy to 527 AI papers (2020‑2024) showed that 42 % addressed rule‑based governance, while learning‑based mechanisms accounted for 38 %. However, only 9 % discussed legal liability, confirming the gap noted earlier.

5. Tools, Automation, and the Role of AI Agents

Manual screening and classification become bottlenecks as literature volumes swell. Recent advances in large language models (LLMs) and machine‑learning pipelines allow researchers to delegate repetitive tasks to self‑governing AI agents while retaining human oversight.

5.1 Screening with AI

• Active Learning: Tools like ASReview use Bayesian active learning to prioritize the most informative papers for human review. In a pilot on pollinator disease literature (3,200 abstracts), active learning reduced the number of abstracts needed for 95 % recall from 2,800 to 1,200 (≈57 % saving).
• LLM‑assisted Classification: Prompt an LLM (e.g., GPT‑4) with a few labelled examples and ask it to assign a research type label to new abstracts. A study by Miller et al. (2024) reported an F1‑score of 0.89 for distinguishing empirical vs. review papers after fine‑tuning on 500 examples.

5.2 Bibliometric Mapping

• VOSviewer and CiteSpace generate co‑citation and keyword networks.
• Bibliometrix (R package) produces yearly publication trends, author productivity, and collaboration maps.

When mapping the global bee‑disease literature, a VOSviewer overlay showed three dominant clusters: (1) Varroa destructor (Europe, 2000‑2015), (2) Nosema ceranae (Asia, 2010‑2020), and (3) Emerging viral pathogens (Africa, 2018‑2024). The visual gap in African viral research prompted a targeted grant call in 2025.

5.3 Building a Semi‑Automated Mapping Pipeline

Below is a high‑level pseudocode for an AI‑augmented mapping pipeline. The pipeline is modular, allowing you to replace components (e.g., switch from GPT‑4 to an open‑source LLaMA model).

# 1. Load raw records
records = load_bibtex('search_results.bib')

# 2. De‑duplicate
records = deduplicate(records)

# 3. Active learning screening
screened = active_learning_screen(records,
                                  initial_seed=200,
                                  target_recall=0.95)

# 4. LLM classification
labels = llm_classify(screened['abstract'],
                      prompt_template='Assign research type ...',
                      examples=training_set)

# 5. Store in SQLite
db = sqlite3.connect('mapping.db')
store_records(screened, labels, db)

# 6. Generate visualisations
plot_bubble_chart(db, x='year', y='research_type')
export_network(db, method='co_citation')

Key safeguards:

• Human verification of at least 10 % of LLM predictions.
• Version control of the prompt and training data (Git).
• Transparency report documenting model version, temperature, and evaluation metrics (precision, recall).

5.4 Ethical Considerations

Even when using AI agents, adhere to the principles of reproducibility:

• Archive the exact model weights and prompts.
• Provide a data‑statement describing any bias introduced (e.g., language bias if non‑English papers were excluded).

By treating AI agents as assistants rather than authoritative judges, you preserve the integrity of the mapping study while gaining efficiency.

6. Quality Assessment and Threats to Validity

A systematic map is only as trustworthy as the process that built it. Below are the most common validity threats and mitigation strategies.

A quality checklist (adapted from Garousi & Mäntylä) can be embedded in the protocol and attached as an appendix. The checklist asks reviewers to rate each primary study on methodological rigor, sample size, and reporting completeness, allowing downstream users to filter by quality if needed.

7. Reporting and Visualisation

Transparent reporting is essential for credibility and reuse. The PRISMA‑SM guideline recommends the following sections in the final manuscript:

1. Title & Abstract – Clearly state that the work is a systematic mapping study.
2. Introduction – Context, objectives, and research questions.
3. Methods – Detailed protocol, search strings, databases, screening flow, data extraction form.
4. Results – Descriptive statistics (number of studies per year, per region), classification tables, visual maps.
5. Discussion – Interpretation of trends, identification of gaps, comparison with prior reviews.
6. Limitations – Threats to validity (see Section 6).
7. Conclusions & Future Work – Recommendations for researchers, funders, and policy makers.

7.1 Visual Templates

• Bubble Plot (Year vs. Research Type)

library(ggplot2)
ggplot(data, aes(x=Year, y=ResearchType, size=N)) +
  geom_point(alpha=0.6) +
  scale_size_area(max_size=15) +
  theme_minimal()

• Heat Map (Topic vs. Region)

import seaborn as sns
sns.heatmap(pivot_table, cmap="YlGnBu", annot=True)

• Interactive Dashboard (Shiny)

library(shiny)
ui <- fluidPage(
  selectInput("year", "Year", choices = unique(data$Year)),
  plotOutput("map")
)
server <- function(input, output) {
  output$map <- renderPlot({
    subset <- data[data$Year == input$year, ]
    ggplot(subset, aes(...)) + ...
  })
}
shinyApp(ui, server)

These visualisations should be exported in both raster (PNG) and vector (SVG) formats to facilitate reuse in blog posts, policy briefs, or educational material.

7.2 Cross‑Linking Within Apiary

When publishing on the Apiary platform, embed slug links to related pages:

• For the taxonomy of bee stressors, link to [[bee-stressors]].
• For AI‑agent safety frameworks, link to [[ai-governance]].
• For bibliometric tools, link to [[bibliometrics]].

Cross‑linking not only improves navigation but also builds a knowledge graph that AI agents can later query for meta‑analysis.

8. Case Studies

8.1 Mapping the Global Literature on Wild Pollinator Decline

Objective: Identify research trends on wild pollinator health from 2000‑2024.

Process:

• Databases: Scopus, Web of Science, CAB Abstracts.
• Search string: (wild pollinator* OR bumblebee* OR solitary bee*) AND (decline OR loss OR abundance).
• Results: 3,112 records → 2,785 after de‑duplication.
• Screening: Title/abstract (2,785 → 1,102); full‑text (1,102 → 384).

Findings: