Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 61 of the 61 references it cites were found in a registry or an index; 33 of the 56 sources its claims cite were checked for support.

The SAFE procedure: a practical stopping heuristic for active learning-based screening in systematic reviews and meta-analyses

Open access
Boetje, Josien, van de Schoot, Rens · 2024 · Systematic Reviews

Published version of a preprint

Abstract

Active learning has become an increasingly popular method for screening large amounts of data in systematic reviews and meta-analyses. The active learning process continually improves its predictions on the remaining unlabeled records, with the goal of identifying all relevant records as early as possible. However, determining the optimal point at which to stop the active learning process is a challenge. The cost of additional labeling of records by the reviewer must be balanced against the cost of erroneous exclusions. This paper introduces the SAFE procedure, a practical and conservative set of stopping heuristics that offers a clear guideline for determining when to end the active learning process in screening software like ASReview. The eclectic mix of stopping heuristics helps to minimize the risk of missing relevant papers in the screening process. The proposed stopping heuristic balances the costs of continued screening with the risk of missing relevant records, providing a practical solution for reviewers to make informed decisions on when to stop screening. Although active learning can significantly enhance the quality and efficiency of screening, this method may be more applicable to certain types of datasets and problems. Ultimately, the decision to stop the active learning process depends on careful consideration of the trade-off between the costs of additional record labeling against the potential errors of the current model for the specific dataset and context.

Overview

CoChat Summary

TL;DR

The paper introduces the SAFE procedure, a conservative four-phase stopping heuristic for active learning-based screening in systematic reviews and meta-analyses. It combines multiple stopping checks to balance the cost of continued labeling against the risk of missing relevant records.

Key findings

  • SAFE consists of four phases: screen a random training set, apply active learning, find more relevant records with a different model, and evaluate quality. In Phase 2, screening stops only when all four conditions are met: all key papers are marked relevant; at least twice the estimated total relevant records (RR_T) have been screened; at least 10% of the total dataset has been screened; and no relevant records have been identified in the last, for example, 50 records.
  • An expert meeting with 26 information specialists, data scientists, and active-learning users from the Netherlands and Germany on 12-01-2023 supported a minimum percentage of records screened combined with a threshold of 50 consecutive irrelevant records as “safe and reasonable.” Visual inspection of the recall plot was viewed positively but not precise enough as a stopping rule on its own.
  • Experts agreed key papers are useful for validating screening results but may introduce bias if used as prior knowledge in active learning. The paper also notes that humans typically miss around 10% of relevant records.
  • Time estimates include Phase 4 taking maybe only 1–2 h, while neural network training in Phase 3 may take up to 6 h on a high-performance cluster; deep learning may be less practical for datasets over 50,000 records due to training times.

Methods

  • The stopping heuristic was initially developed in December 2022, inspired by the procedure used by Brouwer et al., and was peer-reviewed on 12-01-2023 by 26 experts. Feedback was collected digitally via Wooclap, discussed by the authors, and used to adapt the heuristic.
  • The paper presents the resulting SAFE procedure, its assumptions, four phases, example stopping values, and practical implementation guidance; it does not report an empirical dataset analysis or a direct comparison with existing stopping heuristics.
  • The expert feedback focused on use of a minimum percentage to screen, a conservative number of consecutive irrelevant records, visual inspection of the recall plot, and use of key papers as prior knowledge.

Limitations

  • The paper states that a direct comparison with existing stopping heuristics has not been presented, and future comparative analysis is needed to validate effectiveness and efficiency in different contexts.
  • The paper states that the cut-off values are examples and may not suit small datasets; for a dataset of only 500 records, screening 1% means only 5 records, and the observed fraction may not represent the complete set, potentially leading to over-screening.
  • The paper states that the proposed batch size depends on the research question, domain, and desired recall, and that the procedure may not apply to all active learning scenarios or types of data and models; it also notes diminishing returns and training-time concerns in the deep learning phase.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1186/s13643-024-02502-7

Cites

61 of 61 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Graphical overview of the SAFE procedure for applying a practical stopping heuristic for active learning-aided systematic reviewing [4] Note: Disclaimer: The numbers provided in this figure are arbitrary and should not be considered universally applicable. Researchers are responsible for choosing appropriate values based on their specific situation and requirements. The SAFE procedure is a flexible framework, and its effectiveness depends on the careful selection of parameters tailored to the context of each systematic review
Fig. 1 Graphical overview of the SAFE procedure for applying a practical stopping heuristic for active learning-aided systematic reviewing [4] Note: Disclaimer: The numbers provided in this figure are arbitrary and should not be considered universally applicable. Researchers are responsible for choosing appropriate values based on their specific situation and requirements. The SAFE procedure is a flexible framework, and its effectiveness depends on the careful selection of parameters tailored to the context of each systematic reviewCC BY · PubMed Central
Fig. 2 Example recall plot comparing the number of identified relevant records against the number of viewed records for active learning (grey line) versus random screening (blue line) [5]
Fig. 2 Example recall plot comparing the number of identified relevant records against the number of viewed records for active learning (grey line) versus random screening (blue line) [5]CC BY · PubMed Central
Fig. 3 Screening speed over time compared between active learning using the SAFE procedure and random screening [6]
Fig. 3 Screening speed over time compared between active learning using the SAFE procedure and random screening [6]CC BY · PubMed Central

Verification

Claims the paper makes

17 Valid · 4 Not valid · 4 Not checked · 48 Own results and interpretations

2 more claims

Create a free account to see all 73 claims and to verify the 4 references still open.

See all claims — free account
  • It may state two separate claims as one. 28
  • Probably not a substantial claim. 59

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 53 references

Research works printed but not linked: 1

  1. Qin X, Liu J, Wang Y, Deng K, Ma Y, Zou K, Li L, Sun X. Application of nature language processing in systematic reviews. Chin J Evid Based Med. 2021;21(6):715–20. https://doi.org/10.7507/1672-2531.202012150. · doi:10.7507/1672-2531.202012150

Cited by

See all 146

Metadata

Authors
Boetje, J., & van de Schoot, R.
Affiliations
University of Applied Sciences Utrecht (NL)
Corresponding author
Boetje, Josien
Venue
Systematic Reviews
Published
2024-03-01
Volume
13
Issue
1
Pages
81-81
ISSN
2046-4053 (ISSN-L)
OpenAlex
W4392356867
Cited by
146
References
53
Citations by year
2026: 80 · 2025: 50 · 2024: 15
FWCI
18.22
Type
journal article
Language
en
Open access
Licence
Locations
  • Systematic Reviews (publishedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DSpace (Open University in the Netherlands) (submittedVersion, cc-by)
Funders
Nederlandse Organisatie voor Wetenschappelijk Onderzoek (406.22.GO.048, 406.22)
Topics
  • Social Sciences › Decision Sciences › Statistics, Probability and Uncertainty › Meta-analysis and systematic reviews
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Algorithms
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
Keywords
active learning (machine learning), systematic review screening, ASReview, cost-benefit trade-off
MeSH terms
Heuristics, Systematic Reviews as Topic, Humans, Software, Problem-Based Learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: The SAFE procedure: a practical stopping heuristic for active learning-based screening in systematic reviews and meta-analyses. https://app.cochat.ai/papers/10.1186/s13643-024-02502-7
The paper
Boetje, J., & van de Schoot, R. (2024). The SAFE procedure: a practical stopping heuristic for active learning-based screening in systematic reviews and meta-analyses. Systematic Reviews. https://doi.org/10.1186/s13643-024-02502-7

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.