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

The DOI matches its Crossref record; 12 of the 12 references it cites were found in a registry or an index; 10 of the 25 sources its claims cite were checked for support.

PED-X-Bench: A Benchmark of Adult-to-Pediatric Extrapolation Decisions in FDA Drug Labels

Open access

Abstract

Abstract Pediatric trials are ethically and logistically difficult, so the U.S. FDA often extrapolates adult data to children when justified. Yet no public resource systematically documents these decisions. We present PED-X-Bench , the first dataset and benchmark that encodes FDA pediatric-extrapolation outcomes as a four-way classification task ( Full, Partial, None, Unlabeled ). PED-X-Bench contains 737 FDA drug-label sections (≈ 1 M words of source text) for approvals issued 2007–2024 across all therapeutic areas. A two-stage o3-mini prompting pipeline mined full FDA label text; nine domain reviewers then adjudicated a stratified sample of 135 labels yielding an accuracy F1 of 0.74 and 0.63 respectively (inter-annotator κ = 0.678) and spot-checking the remainder. For every drug we release the ground-truth label, concise efficacy and pharmacokinetic/safety summaries, and harmonized study metadata. To showcase utility we release two baseline models: (i) a logistic-regression classifier that uses structured metadata from FDA’s pediatric-labeling dataset, and (ii) a fine-tuned BigBird BERT that ingests full label text. Both base-lines perform modestly, leaving ample headroom for future work. PED-X-Bench enables research on pediatric drug development, clinical NLP and drug safety; dataset card and code are made available here: github.com/tatonetti-lab/PedXBench huggingface.co/datasets/apoorvasrinivasan/Ped-X-Bench

Overview

CoChat Summary

TL;DR

PED-X-Bench is the first open dataset encoding FDA adult-to-pediatric extrapolation decisions as a four-way classification task (Full, Partial, None, Unlabeled) over 737 FDA drug-label sections from 2007–2024. Two baseline models (metadata logistic regression and a fine-tuned BigBird transformer) both perform modestly, leaving headroom for future work.

Key findings

  • The corpus contains 737 FDA drug-label sections (≈1M words) for approvals issued 2007–2024; a two-stage o3-mini pipeline labelled them, with nine domain reviewers adjudicating a stratified sample of 135 labels (accuracy 0.74, macro-F1 0.63; inter-annotator κ = 0.678 in the abstract, κ = 0.72 after consensus discussion in Methods).
  • Outcome distribution is skewed: None dominates at 60% (445/737), Partial 31%, Full borrowing is rare at 3%, and 6% are Unlabeled; only 22 fully extrapolated cases exist.
  • Ablations: two-stage o3-mini reached accuracy 0.740 / macro-F1 0.633; swapping to gpt-4o-mini dropped to 0.534 / 0.294; single-stage reached 0.725 / 0.637; single-stage plus verifier reached 0.804 accuracy but macro-F1 fell to 0.538.
  • On the blind test split (34 labels), logistic regression on metadata achieved accuracy 0.647 / macro-F1 0.359, matching BigBird's accuracy (0.647) while doubling its macro-F1 (0.240, loss 0.810).

Methods

  • Data source: FDA's pediatric labeling-change spreadsheet (BPCA, PREA, legacy Pediatric Rule since 1998); after removing veterinary/device entries and collapsing reformulations, 737 human drugs with English PDF labels were retained, converted with pdftotext.
  • Annotation: a two-stage chain-of-thought o3-mini pipeline (evidence extraction, then four-way classification with rationale and ≤150-word efficacy/safety summaries); 135 randomly sampled labels were independently reviewed by nine annotators (eight biomedical data scientists, one clinician) to form a gold test/validation set, with the remaining 602 machine-labelled records as silver training data.
  • Baselines: multinomial logistic regression on a 253-dimensional sparse feature vector (z-scored numeric features, one-hot categoricals, class-balanced weights) and fine-tuned google/bigbird-roberta-base on 4096-token label text, evaluated on a fixed 687/85/34 train/dev/test split using accuracy, macro-F1, and cross-entropy loss.

Limitations

  • Stated: the dataset inherits biases of public labels (negative trials, interim analyses, and unpublished regulatory correspondence are absent); severe class imbalance with only 22 fully extrapolated cases; annotations capture a single 2007–2024 snapshot requiring periodic refresh; and labels address extrapolation at the whole-product level rather than per indication or dosage form.
  • Evident from the text: the gold-standard test set is small (34 labels in the blind test split; 135 adjudicated overall), and the reported inter-annotator agreement differs between the abstract (κ = 0.678) and Methods (κ = 0.72), which the text does not reconcile.

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

Integrity signals

Registry

Crossref · 10.1101/2025.05.22.25328187

Cites

12 of 12 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Two-stage LLM pipeline used to build PED-X-Bench . The model first extracts pediatric evidence from the full FDA label, then classifies the extrapolation outcome while generating concise efficacy and safety summaries. Expert review (green) provides the gold test set; study metadata (blue) are merged to form the final corpus.
Figure 1 Two-stage LLM pipeline used to build PED-X-Bench . The model first extracts pediatric evidence from the full FDA label, then classifies the extrapolation outcome while generating concise efficacy and safety summaries. Expert review (green) provides the gold test set; study metadata (blue) are merged to form the final corpus.From the PDF
Figure 3 Corpus statistics. A) Label distribution. B) Extrapolation outcome by legislative pathway. C) Age-range coverage for each class (violins show min-max bounds). D) Number of pediatric studies cited per label.
Figure 3 Corpus statistics. A) Label distribution. B) Extrapolation outcome by legislative pathway. C) Age-range coverage for each class (violins show min-max bounds). D) Number of pediatric studies cited per label.From the PDF
Figure 4 Study-design and outcome flags by extrapolation class. Bars denote the percentage of applications with each PLC flag.
Figure 4 Study-design and outcome flags by extrapolation class. Bars denote the percentage of applications with each PLC flag.From the PDF

Verification

Claims the paper makes

4 Valid · 1 Not valid · 9 Not checked · 45 Own results and interpretations

21 more claims

Create a free account to see all 59 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 25
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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.

3 of 3

Citations

Cites

Show all 11 references

Metadata

Authors
Srinivasan, A., Berkowitz, J., Friedrich, N. A., Tsang, K., Kuchi, A., Acitores, J., Zietz, M., Czarny, R. S., Liu, H., & Tatonetti, N. P.
Affiliations
Cedars-Sinai Medical Center (US)
Corresponding author
Srinivasan, Apoorva
Venue
medRxiv
Publisher
openRxiv
Published
2025-05-23
OpenAlex
W4410746186
Cited by
0
References
11
Type
posted content
Language
en
Open access
Locations
medRxiv (acceptedVersion, cc-by)
Topics
  • Health Sciences › Medicine › Pediatrics, Perinatology and Child Health › Pharmaceutical studies and practices
  • Health Sciences › Medicine › Health Informatics › Biomedical Text Mining and Ontologies
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Toxicology › Pharmacovigilance and Adverse Drug Reactions
Keywords
FDA drug labels, pediatric drug development, drug safety, efficacy claims, pharmacokinetics, annotated dataset, benchmark, clinical natural language processing, drug approval

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (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: PED-X-Bench: A Benchmark of Adult-to-Pediatric Extrapolation Decisions in FDA Drug Labels. https://app.cochat.ai/papers/10.1101/2025.05.22.25328187
The paper
Srinivasan, A., Berkowitz, J., Friedrich, N. A., Tsang, K., Kuchi, A., Acitores, J., Zietz, M., Czarny, R. S., Liu, H., & Tatonetti, N. P. (2025). PED-X-Bench: A Benchmark of Adult-to-Pediatric Extrapolation Decisions in FDA Drug Labels. medRxiv. https://doi.org/10.1101/2025.05.22.25328187

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.