Which Q1 checklist for medical AI studies should you use?

Direct answer: Choose your Q1 checklist for medical AI studies by what the paper builds. A model applied to images follows CLAIM, a model that predicts a clinical outcome from patient data follows TRIPOD+AI, and a project aimed at hospital deployment adds the FUTURE-AI principles.

Two guidelines carry most of the load. The Checklist for Artificial Intelligence in Medical Imaging, known as CLAIM, tells authors what to report when a study builds or tests an AI model on medical images. Tejani et al. (2024) published the 2024 update of the checklist, and it is registered with the EQUATOR Network, which promotes reporting guidelines in health research. TRIPOD+AI tells authors what to report when a study develops or evaluates a prediction model, whether it uses regression or machine learning. Collins et al. (2024) present it as a 27-item checklist with a 13-item checklist for abstracts and state that it replaces the 37-item TRIPOD checklist of 2015.

The two overlap when an imaging model predicts a diagnosis or prognosis. Read both, follow the one that fits the design better, and say in the methods which one you used. Collins et al. (2024) place CLAIM beside TRIPOD+AI in a table of reporting guidelines for machine-learning health research, with CLAIM covering "medical imaging studies using artificial intelligence".

Source: the three guideline papers in References, read on 2026-10-27.

Direct answer: Often, and at the same places each time. Of 215 deep-learning papers excluded from a COVID-19 imaging review, 110 (51%) failed at least three mandatory CLAIM criteria, and in a survey of 516 diagnostic AI studies only 31 (6%) used external validation.

Roberts et al. (2021) searched for machine learning models that detect or predict COVID-19 from chest radiographs and CT scans. They identified 2,212 studies, screened 415, assessed 320 in full text and kept 62 after quality review. Of the 215 excluded deep-learning papers, 110 (51%) failed at least three mandatory CLAIM criteria, and the most common gaps were insufficient documentation of final model selection (61%, 132 papers), of preprocessing (58%, 125 papers) and of training details (49%, 105 papers). The authors concluded that none of the models they reviewed was of potential clinical use. The review covers one disease in a period of rushed publication, so the rates are not a general estimate for all medical AI.

Kim et al. (2019) looked at a broader field. From 2,748 collected articles they found 516 eligible studies of diagnostic AI on medical images. Only 31 (6%) used external validation, 4 (0.8%) were prospective, 15 (2.9%) were multi-institutional and 5 (1.0%) used a diagnostic cohort, and no study met all three design features the authors asked for. This is a 2019 sample, so newer papers may do better, but the same three items still appear in current checklists.

You can use these audits as a quick self-test. Treat this as a Q1 checklist for medical AI studies and count the mandatory items you cannot document from your own methods section. If the count reaches three, your paper sits where more than half of the excluded papers in Roberts et al. (2021) sat, and the gap is cheaper to close before submission than after a desk rejection.

Direct answer: Reviewers open the items on data partitioning, external testing, statistical uncertainty and error analysis, because those decide whether the reported score can be trusted. Fill the lines that cover those topics first, using the numbering in the published CLAIM 2024 checklist.

The partitioning items carry the most weight. They ask authors to state the level at which training, tuning and test sets are separated, and for images the safe level is the patient, as the next section shows. The items on external testing ask for evaluation on data from another source, or an explicit and justified statement of why it was not done. Reviewers also look for confidence intervals when metrics are compared, and for examples of the cases the model got wrong. A paper that supplies all four lets a reader judge the headline score without guessing.

The table below maps manuscript sections to the topics of the CLAIM 2024 update. It is an editing tool built by MAAS from the topics the checklist covers, not part of the guideline, so check each line against the published item numbers before you submit.

Source: topics covered by the CLAIM 2024 update (Tejani et al., 2024); mapping by MAAS, 2026-10-27.

Direct answer: When one patient contributes several images, a random split by image places most test images next to a sibling from the same patient in training. The example below shows that this happens to about 99.8% of test images, which is why reviewers ask authors to state the split level.

Koçak et al. (2025) write that "a common and serious source of bias in model validation lies in data leakage", and give the case of repeated scans of one patient appearing in both training and validation sets. A short calculation shows how fast that happens. Suppose a dataset holds 4,000 images from 800 patients, 5 images each, and you assign each image to training or test at random with an 80/20 split. For one test image, each of the 4 other images from the same patient lands in training with probability 0.8, so the chance that none does is 0.2 to the fourth power, or 0.16%. About 99.8% of test images therefore have a same-patient sibling in training, and the model can score well by recognising the patient and not the disease. Splitting by patient before any image is touched removes the problem.

The calculation above is an illustration worked out by MAAS, not a result from a study, and the exact figure depends on how many images each patient has. Reviewers who see a random split by image usually ask for a patient-level split, so state the level in the methods and in the checklist.

Sharing is the second concern. Lee et al. (2025) reviewed AI model-development studies in five leading radiology journals in 2022 and found that 107 of 268 studies (39.9%) made their model publicly available. Among 113 deep-learning studies, 13 (11.5%) shared fully, and training code was available in 25 (22.1%). Koçak et al. (2025) add that fewer than about 6% of articles in radiology journals share the data used to build the model. Prepare the patient-level split table and a statement on code and data access before you submit, because both answer questions reviewers would otherwise ask in round one. The fuller workflow is in our guide to reproducible AI health research.

Direct answer: Use TRIPOD+AI when the main product is a prediction model for diagnosis, prognosis or risk, and CLAIM when it is an AI method applied to images. FUTURE-AI is a development and deployment framework, so it can support design claims but does not replace a reporting checklist.

TRIPOD+AI grew out of a 2015 statement for regression models. Collins et al. (2024) explain that machine learning methods had spread and the older checklist needed updating, so the new version gives one guidance for prediction model studies whatever the modelling method. It also defines the term "evaluation data" for data used to estimate performance, which must be separate from the training data so that no participant appears in both.

Lekadir et al. (2025) describe a consortium formed in 2021 of 117 interdisciplinary experts from 50 countries. Over two years they reached consensus on six guiding principles (fairness, universality, traceability, usability, robustness and explainability) and 30 best practices that cover the AI lifecycle from design to monitoring. The paper itself separates its role from reporting guidelines: TRIPOD+AI and CLAIM "do not provide best practices for the actual development and deployment of the AI tools, but promote standardised and complete reporting" (Lekadir et al., 2025). A claim of deployment readiness without evidence invites criticism, so use FUTURE-AI language only for claims your data can support, such as a bias analysis across patient groups.

A paper that does both jobs, for example an imaging model that predicts survival from a scan plus clinical variables, can need TRIPOD+AI logic for the prediction part and CLAIM items for the imaging part. Pick one checklist as the formal submission and bring the missing items of the other into the methods.

Direct answer: Read the scope, page limit and data policy of the target journal first, then match your contribution to them. IEEE Transactions on Medical Imaging, for example, limits an initial submission to ten pages and treats sharing code and test data as best practice.

The author instructions of IEEE Transactions on Medical Imaging (TMI) are explicit. The journal states that an initial submission over ten pages including references is returned without review, and that printed papers beyond eight pages carry a charge of US$250 per page for pages nine and ten (IEEE Transactions on Medical Imaging, 2025). On reproducibility the wording is measured: "Upon publication, TMI considers it best practice to share custom computer code and test data in a way that allows readers to reproduce the published results" (IEEE Transactions on Medical Imaging, 2025).

The same method works for any target. Write down the journal's scope, page limit and data policy and set them beside your paper. If a clinical evaluation of an existing method adds no new method, a clinical imaging journal is the honest fit. Our guides to choosing a Scopus journal and to Q1 versus Q2 journals cover the indexing side.