Use the right framework for the question

Once the study has been reconstructed, a reporting or risk-of-bias framework can make omissions easier to find. It cannot do the reconstruction for the reader. Nor can a completed checklist prove that a reference standard was valid, that a test set was untouched, that a model is causal, or that using it improves care. Frameworks are lenses layered on the study, not substitutes for evidence.

Completeness is not proof

Three questions are often collapsed into one. Reporting asks whether the paper has disclosed the information needed to understand and reproduce its design. Risk of bias asks how the design or analysis could systematically mislead the result. Computation asks whether a numerical definition can be implemented consistently and checked against a reference. A reporting item can be present while the underlying choice is still biased. A reproducible feature calculation can still be a poor measurement of the intended biology. A low-risk-of-bias prediction can still have no decision utility.

Different tools answer different questions

The study type determines which lenses are relevant. CLAIM is used as a reporting lens for medical-imaging AI model studies such as classification, segmentation, and reconstruction. TRIPOD+AI is a reporting lens for clinical prediction models, including models built with machine learning. PROBAST+AI is a risk-of-bias lens for prediction models. CLEAR is a reporting lens for clinical radiomics. IBSI addresses the definitions and computation of radiomic features and filters so that an implementation can be compared with reference definitions and values.

These roles can overlap without becoming interchangeable. A deep-learning classifier on medical images that is also presented as a diagnostic prediction model may need CLAIM and TRIPOD+AI as complementary reporting lenses, with PROBAST+AI for prediction-model risk of bias. A radiomics study should use a radiomics-specific reporting lens rather than treating CLAIM as its primary radiomics checklist. IBSI can make feature computation more inspectable, but it does not establish cohort validity, clinical usefulness, or absence of bias.

Match the study to its lenses

Question about the paper	Useful lens	What the lens helps expose	What it cannot establish
Is this medical-imaging AI model work reported sufficiently?	CLAIM	data, model, evaluation, and implementation omissions	that the model is unbiased or clinically useful
Is this a clinical prediction model reported sufficiently?	TRIPOD+AI	target, predictors, outcome, missing data, validation, calibration, and utility reporting	a causal effect or a successful deployment
Could the prediction result be systematically misleading?	PROBAST+AI	participant, predictor, outcome, analysis, and validation risks	every imaging-specific computation or shortcut
Is a clinical radiomics study described sufficiently?	CLEAR	acquisition, segmentation, feature, analysis, and reporting details	proof that a feature is biological or useful
Can a radiomic feature calculation be compared across implementations?	IBSI	mathematical definitions, filters, settings, and reference computations	validity of the clinical question or patient cohort

The names, editions, item counts, endorsements, and current links of these resources are changeable reference information. They belong in the maintained companion Web and should be checked against the current public source when a paper is appraised. They should not carry a stable manuscript claim merely because a particular edition was available at the time of writing.

Choose evidence by the question

Evidence also has a hierarchy that depends on the question. A normative standard is the right authority for a technical definition or computation contract. The original peer-reviewed full text is needed for claims about a study’s methods, cohort construction, split, preprocessing, results, and limitations. Reporting and risk-of-bias guidance answers what should be disclosed or assessed. A high-quality review can map a field and its primary literature. Abstracts and metadata support paper identity and broad orientation only. Consequential interpretation of Methods or Results requires inspection of the relevant full text; when that text cannot be inspected, the claim must be narrowed or withheld.

This prevents citation proximity from becoming evidence. A framework cited in a paper does not show that the paper followed it. A paper’s reference list does not show that the cited source supports every sentence nearby. For a consequential claim, record the source and the level inspected, and narrow the claim if the relevant passage, table, or method cannot be checked.

Read omissions without turning the checklist into a verdict

Use a framework after writing the reconstruction. The checklist can tell you that a radiomics paper does not report its image processing or feature settings, or that a prediction paper does not describe calibration. It cannot tell you whether the hidden processing was appropriate or whether the unreported model would have transported. Those are separate scientific judgements.

The correct consequence of an omission is conditional. If the omitted detail is necessary to identify the estimand or independence boundary, the corresponding claim remains unresolved. If it is useful for reproducibility but not required for the narrow result, the result may survive with a reporting limitation. If a source-grounded review finds that the design actually breached a boundary, the claim must be remediated for that reason. A checklist helps locate the decision; it does not make the decision automatically.

The final case applies all these lenses to one paper. It is deliberately read as a reconstruction first, a metric report second, and a claim about clinical use only if the design provides evidence for that additional step.