Skip to content
Imaging Evidence Path

The reporting and risk-of-bias frameworks

Tonight · ~20 min · read · energy: low · setup: none

CLAIM, TRIPOD+AI, PROBAST+AI, CLEAR, IBSI — the alphabet soup is intimidating until you see that each framework answers a different question. This lesson teaches why each exists and when to reach for it, then compresses them into one table. The point is to choose the lens by the study’s purpose, not to memorise checklists.

The anchor: different questions, different frameworks

These frameworks are not competing; they are layered. A reporting framework tells you what a paper must show; a risk-of-bias framework tells you where its design could mislead; a computation standard tells you whether the numbers are reproducible. You choose by the study type and the question you are asking of it.

Why each exists

  • CLAIM 2024 (CLEAR for radiomics — see below) — a reporting guideline for medical-imaging AI model studies (classification, segmentation, reconstruction). It exists because imaging-AI papers routinely omitted the details (data, training, evaluation, deployment) needed to judge them. Note: CLAIM explicitly did not extend its scope to radiomics — use CLEAR for radiomics, not CLAIM.
  • TRIPOD+AI — a reporting guideline for clinical prediction models (regression or ML). It exists because prediction-model papers omitted the calibration, handling of missing data, and clinical-utility analyses that decide whether a model is usable.
  • PROBAST+AI (Moons et al., BMJ 2025) — a risk-of-bias tool for prediction models. It exists to systematically find the leakage paths, circularity, and testing flaws that lessons 2–5 walked through.
  • CLEAR (Kocak et al., Insights Imaging 2023;14:75, PMID 37142815, VERIFIED; 58 items, ESR/EuSoMII-endorsed; with CLEAR-E3 explanations and the METRICS quality score) — the reporting guideline for clinical radiomics studies. Use this instead of CLAIM when the study is radiomics.
  • IBSI 1/2 — computation standards for radiomic features/filters (Ch. 3). They exist so that a feature value is reproducible across software.

The framework table

Choose by study type; combine when scopes overlap.

If the study is… Reporting lens Risk-of-bias lens
Medical-imaging AI model (classification / segmentation / reconstruction) CLAIM 2024 study-type dependent; PROBAST+AI when it is also a clinical prediction model
Clinical prediction model (regression or ML) TRIPOD+AI PROBAST+AI
Clinical radiomics study CLEAR CLEAR / METRICS
Radiomic feature computation / reproducibility IBSI 1/2 (reference values)

CLAIM 2024 is a reporting lens; it is not a formal risk-of-bias tool. For medical-imaging AI, the RoB lens is study-type dependent, and PROBAST+AI applies when the imaging AI is also a clinical prediction model (so CLAIM + TRIPOD+AI can be complementary, with PROBAST+AI supplying the RoB lens).

Critical correction to a common misuse: CLAIM 2024 explicitly did not extend its scope to imaging-biomarker/radiomics research. Do not use CLAIM as the primary radiomics checklist. Use CLEAR. All are linked in resources/standards-and-reference.md.

How to use the table when reading

You do not run the full checklist on every paper. You use the table to know which questions a complete paper of that type should have answered, and notice which are missing. A radiomics paper that reports no binWidth/spacing fails CLEAR before you reach item 10; a prediction-model paper with no calibration fails TRIPOD+AI. The frameworks make the omissions nameable.

Stop and think — then reveal

You are appraising a study that builds a deep-learning ACL-tear classifier on knee MRI and reports AUROC on internal + external test sets, with no calibration or reader-study. It is both a medical-imaging-AI model and a diagnostic prediction model. Which reporting lenses apply, which risk-of-bias lens applies, and which does not apply?

Reporting: CLAIM 2024 (imaging-AI model) and TRIPOD+AI (diagnostic prediction model) — complementary. Risk-of-bias: PROBAST+AI (it is a clinical prediction model). CLEAR does not apply — it is not a radiomics study (no feature battery). So this is exactly the Tran et al. case (lesson 8): CLAIM + TRIPOD+AI for reporting, PROBAST+AI for risk of bias. The missing calibration fails the TRIPOD+AI lens; the reader-study absence is a CLAIM/TRIPOD+AI gap.

What to retain

  1. Frameworks are layered: reporting (what to show) vs risk-of-bias (where it misleads) vs computation (whether it reproduces). Choose by study type.
  2. CLAIM 2024 (imaging-AI model), TRIPOD+AI (prediction model), CLEAR (radiomics), IBSI 1/2 (feature computation); PROBAST+AI supplies the RoB lens for prediction models.
  3. Do not use CLAIM for radiomics — use CLEAR. CLAIM + TRIPOD+AI can be complementary; PROBAST+AI adds the RoB lens.
  4. Use the table to name the questions a complete paper should answer, and notice omissions.

Next: put all seven lessons to work on a real paper — the Tran et al. dissection.