CoChat

Extracting the Sample Size From Randomized Controlled Trials in Explainable Fashion Using Natural Language Processing

Open access

The DOI matches its Crossref record; 13 of the 13 references it cites were found in a registry or an index; none of the 2 sources its claims cite is checked for support yet.

Abstract

Background: Extracting the sample size from randomized controlled trials (RCTs) remains a challenge to developing better search functionalities or automating systematic reviews. Most current approaches rely on the sample size being explicitly mentioned in the abstract. Methods: 847 RCTs from high-impact medical journals were tagged with six different entities that could indicate the sample size. A named entity recognition (NER) model was trained to extract the entities and then deployed on a test set of 150 RCTs. The entities’ performance in predicting the actual number of trial participants who were randomized was assessed and possible combinations of the entities were evaluated to create predictive models. Results: The most accurate model could make predictions for 64.7% of trials in the test set, and the resulting predictions were within 10% of the ground truth in 96.9% of cases. A less strict model could make a prediction for 96.0% of trials, and its predictions were within 10% of the ground truth in 88.2% of cases. Conclusion: Training a named entity recognition model to predict the sample size from randomized controlled trials is feasible, not only if the sample size is explicitly mentioned but also if the sample size can be calculated, e.g., by adding up the number of patients in each arm.

Overview

CoChat Summary

TL;DR

This preprint trains a named entity recognition model to extract sample-size-related entities from RCT abstracts and combine them into predictions of the number randomized. The most accurate model predicted for 64.7% of test trials, with 96.9% of those predictions within 10% of ground truth.

Key findings

  • The conditional model made predictions for 64.7% of the 150-trial test set; 92.8% were within 1% and 96.9% within 10% of ground truth, with a mean absolute percentage error of 1.82%.
  • The accurate entities model predicted for 96.0% of trials; 80.6% were within 1% and 88.2% within 10%, with a mean absolute percentage error of 7.9%.
  • The ordered model predicted for 98.0% of trials; 78.9% were within 1% and 86.4% within 10%, with a mean absolute percentage error of 13.7%.
  • Among individual entities, randomized total was extracted in 42.0% of trials and had a mean absolute percentage error of 2.4%; arms was extracted in 74.0% with a mean absolute percentage error of 11.1%; general total was extracted in 65.3% with a mean absolute percentage error of 19.0%.

Methods

  • A random sample of 996 RCTs from seven major journals (British Medical Journal, JAMA, JAMA Oncology, Journal of Clinical Oncology, Lancet, Lancet Oncology, New England Journal of Medicine) published between 2010 and 2022 was labeled; abstracts were retrieved from PubMed and parsed with regular expressions.
  • Six entities were tagged in each abstract: randomized total, analysis total, completion total, general total, arm, and arm event. Ground truth was the number of people randomized, retrieved from the abstract and, if uncertain, the full publication.
  • 150 annotated examples were randomly assigned to an unseen test set; the remaining 846 were used to train and validate a RoBERTa-base named entity recognition model with Adam, using an 85:15 training/validation split. Entities were cleaned with regular expressions and conditional statements, then combined into ordered, accurate entities, and conditional prediction models.

Limitations

  • The paper states that only trials from seven journals were used for training and testing, so the model’s ability to generalize to trials from other journals, especially those with unstructured abstracts, cannot be assessed.
  • Evident from the text: the approach depends on abstracts, and predictions are made only when relevant entities are found or meet model conditions, so trials without extractable sample-size information may receive no prediction or less accurate estimates.
  • Evident from the text: the abstract reports 847 RCTs tagged, while the Methods report 996 labeled RCTs and 846 used for training/validation after a 150-trial test set, an inconsistency in the reported dataset size.

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

Integrity signals

Registry

Crossref · 10.1101/2024.07.09.24310155

Cites

13 of 13 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 0 Not valid · 1 Not checked · 18 Own results and interpretations

5 more claims

Create a free account to see all 19 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 7
  • Probably not a substantial claim. 2

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 13 references

Research works printed but not linked: 1

  1. Reidsma D , Hofs DHW , Jovanovic N. Designing Focused and Efficient Annotation Tools. In: Measuring Behaviour, 5th International Conference on Methods and Techniques in Behavioral Research. Noldus Information Technology; 2005:149–152.

Cited by

Metadata

Authors
Windisch, P., Dennstädt, F., Koechli, C., Förster, R., Schröder, C., Aebersold, D. M., & Zwahlen, D. R.
Affiliations
  • Kantonsspital Winterthur (CH)
  • University of Bern (CH)
Corresponding author
Windisch, Paul
Venue
medRxiv
Publisher
openRxiv
Published
2024-07-10
OpenAlex
W4400491757
Cited by
2
References
11
Citations by year
2025: 1 · 2024: 1
Type
posted content
Language
en
Open access
Locations
medRxiv (acceptedVersion)
Topics
  • Health Sciences › Medicine › Health Informatics › Biomedical Text Mining and Ontologies
  • Social Sciences › Decision Sciences › Statistics, Probability and Uncertainty › Meta-analysis and systematic reviews
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
randomized controlled trials, named entity recognition, explainable AI, natural language processing, systematic review automation, entity tagging, prediction model

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.

A language model drew the claims from the paper on 2026-10-04, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-04 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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

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.