Why English? | Using CAIR to Explain Clinical AI Research Through Questions, Data, Models, Validation, Results and Limitations

ORIENTATION

What this guide owns

Current fact check

CAIR was published in 2021 as a Clinical AI Research checklist proposal. Its primary paper presents recommendations across the usual report sections and is intended to improve reporting of medical AI research to clinicians; EQUATOR lists it as guidance applying to the whole report.

Canonical boundary

This article owns clinician-facing explanation of clinical AI research with CAIR. REFORMS owns field-agnostic ML-based science, CLIX-M owns evaluation of explanations, and specialised trial, imaging, prediction-model and implementation guides retain their own purposes.

Worked example

Every chapter returns to a fictional emergency-department model that predicts which adults with a suspected wrist injury need urgent specialist review. It is fictional, low-stakes and educational; it does not report a real school, clinical decision or guaranteed outcome.

Warm up with eduKate’s guides to explaining scientific evidence and asking better questions in class.

CHAPTER 1 OF 16

1. Name AI and Its Intended Clinical Use

Back to contents

Why this move matters

Use the title or abstract to say that AI or machine learning is used and state the targeted condition and purpose.

Begin with the clinical problem

Use the title or abstract to say that AI or machine learning is used and state the targeted condition and purpose. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The report says the model supports urgent wrist-review prioritisation. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

AI is not a clinical purpose by itself. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Name AI and Its Intended Clinical Use”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 2 OF 16

2. Lead With the Clinical Problem

Back to contents

Why this move matters

Describe the unmet need, current pathway, intended user and place where the tool might add information or change action.

Begin with the clinical problem

Describe the unmet need, current pathway, intended user and place where the tool might add information or change action. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The introduction explains delays and the existing triage process before discussing models. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A novel algorithm can answer a low-priority problem. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Lead With the Clinical Problem”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 3 OF 16

3. Write the Research Question as a Decision

Back to contents

Why this move matters

State population, input, output, comparator, timing and intended decision in one testable question.

Begin with the clinical problem

State population, input, output, comparator, timing and intended decision in one testable question. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The question asks whether the score improves urgent-review prioritisation at arrival. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A broad aim such as improve diagnosis hides the decision point. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Write the Research Question as a Decision”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 4 OF 16

4. Separate Participant and Input Eligibility

Back to contents

Why this move matters

Report patient inclusion and exclusion separately from image, signal, text or record requirements, with reasons.

Begin with the clinical problem

Report patient inclusion and exclusion separately from image, signal, text or record requirements, with reasons. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Adults are eligible by presentation; records are separately excluded for missing required views. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Data-quality exclusions can silently change the patient population. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Separate Participant and Input Eligibility”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 5 OF 16

5. Describe Data Acquisition, Selection and Timing

Back to contents

Why this move matters

Give sites, dates, devices, sampling, clinical context, repeated cases and the moment each input became available.

Begin with the clinical problem

Give sites, dates, devices, sampling, clinical context, repeated cases and the moment each input became available. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Records come from two hospitals and only information available at triage enters the model. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Later information can leak the answer into an earlier prediction task. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Describe Data Acquisition, Selection and Timing”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 6 OF 16

6. Protect Training, Validation and Test Independence

Back to contents

Why this move matters

Explain partition units, repeated patients, duplicates, preprocessing and differences between training and testing.

Begin with the clinical problem

Explain partition units, repeated patients, duplicates, preprocessing and differences between training and testing. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Every visit from one patient stays in one partition and the external hospital remains untouched. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Random record splitting can create patient leakage. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Protect Training, Validation and Test Independence”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 7 OF 16

7. Report Human Handling and Ground Truth

Back to contents

Why this move matters

Describe annotators, expertise, instructions, training, blinding, disagreement, consensus and noise in labels.

Begin with the clinical problem

Describe annotators, expertise, instructions, training, blinding, disagreement, consensus and noise in labels. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Two specialists review outcomes and disagreements are resolved before model testing. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A confident label can still be uncertain or subjective. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Report Human Handling and Ground Truth”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 8 OF 16

8. Explain Missing, Poor-Quality and Rare Data

Back to contents

Why this move matters

State how missingness, outliers, low-quality inputs and rare conditions were handled and why.

Begin with the clinical problem

State how missingness, outliers, low-quality inputs and rare conditions were handled and why. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The report shows missingness by outcome and keeps a rare-pathology audit. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Deleting difficult cases can make the model look cleaner than practice. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Explain Missing, Poor-Quality and Rare Data”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 9 OF 16

9. Specify Model Architecture and Training

Back to contents

Why this move matters

Name model type, inputs, preprocessing, pretraining, overlap with current data, loss, regularisation, optimisation, epochs and other material settings.

Begin with the clinical problem

Name model type, inputs, preprocessing, pretraining, overlap with current data, loss, regularisation, optimisation, epochs and other material settings. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The architecture and training schedule are reported with the reasons for key choices. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A model name alone is not a reproducible method. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Specify Model Architecture and Training”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 10 OF 16

10. Record the Exact Model Version

Back to contents

Why this move matters

Give code, weights, software and model versions, access date and which version produced every reported result.

Begin with the clinical problem

Give code, weights, software and model versions, access date and which version produced every reported result. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The locked evaluation version is distinguished from later development builds. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Performance cannot be assigned safely to a family name. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Record the Exact Model Version”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 11 OF 16

11. Define the Output and Decision Threshold

Back to contents

Why this move matters

Say whether the output is a probability, class, region, text or other object and how thresholds or post-processing were chosen.

Begin with the clinical problem

Say whether the output is a probability, class, region, text or other object and how thresholds or post-processing were chosen. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: A probability becomes an urgent flag at a prespecified threshold. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A threshold chosen after seeing test results overstates performance. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Define the Output and Decision Threshold”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 12 OF 16

12. Choose Familiar and Appropriate Performance Measures

Back to contents

Why this move matters

Justify primary and additional metrics, exact variants, prevalence, confidence intervals and comparison methods.

Begin with the clinical problem

Justify primary and additional metrics, exact variants, prevalence, confidence intervals and comparison methods. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Sensitivity and workload at the triage threshold accompany discrimination and calibration. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Accuracy can mislead when urgent cases are uncommon. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Choose Familiar and Appropriate Performance Measures”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 13 OF 16

13. Report Errors and Failure Analysis

Back to contents

Why this move matters

Describe false positives, false negatives, difficult subgroups, input failures and plausible harm, or explain why analysis was not possible.

Begin with the clinical problem

Describe false positives, false negatives, difficult subgroups, input failures and plausible harm, or explain why analysis was not possible. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: Missed urgent cases are reviewed for shared data and workflow patterns. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Average performance can hide a concentrated danger. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Report Errors and Failure Analysis”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 14 OF 16

14. Explain Access, Licensing and Reuse

Back to contents

Why this move matters

State how code, data, model and documentation can be accessed, with licences, restrictions and reasons when sharing is impossible.

Begin with the clinical problem

State how code, data, model and documentation can be accessed, with licences, restrictions and reasons when sharing is impossible. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: A controlled test set and versioned code are available under documented conditions. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Open code cannot reproduce unavailable clinical data or labels. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Explain Access, Licensing and Reuse”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 15 OF 16

15. Report Ethics, Bias and Generalisability

Back to contents

Why this move matters

Discuss consent or waiver, privacy, representation, subgroup performance, bias, intended settings and conditions where use is unsafe.

Begin with the clinical problem

Discuss consent or waiver, privacy, representation, subgroup performance, bias, intended settings and conditions where use is unsafe. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The report limits transfer beyond the hospitals and age groups studied. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

A diverse table does not by itself establish fairness. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “Report Ethics, Bias and Generalisability”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

CHAPTER 16 OF 16

16. End With a Bounded Clinical Claim

Back to contents

Why this move matters

Connect results to the original question, comparator and study design; state evidence needed before implementation.

Begin with the clinical problem

Connect results to the original question, comparator and study design; state evidence needed before implementation. State the condition, intended use, position in the pathway and people affected before describing the algorithm.

Translate the pipeline for a clinician

Worked example: The conclusion supports prospective workflow evaluation rather than immediate autonomous use. Trace patient selection, input data, labels, preprocessing, model, output, threshold, action and error in ordinary clinical sequence.

Keep technical detail without technical theatre

Name architecture, parameters, versions and measures only where they let readers reproduce, compare or understand the claim. Explain specialist terms at first use.

Keep the evidence boundary

Clear language should reduce, not inflate, certainty. Clear reporting helps clinicians judge a study; it does not make a model clinically effective, unbiased, safe or ready for deployment.

Student, parent and career transfer

Make one evidence card for “End With a Bounded Clinical Claim”: source location, plain-language meaning, limit and next question. A parent or teacher can ask for one everyday explanation and one specialist explanation; both must preserve the same conditions. This habit supports medicine, nursing, allied health, clinical data science, biomedical engineering, journal editing, product evaluation and science communication without promising a grade, admission, funding decision or job.

PRACTICE LAB

Build a cair decision trail

Student route

1. Choose one original report inside this guide’s scope.
2. Copy its objective and underline actor, action, context, evidence and intended decision.
3. Draw the specific reasoning chain used in this article.
4. Mark each step reported, inferred, contested or unknown.
5. Rewrite the conclusion so it preserves the narrowest important limitation.

Parent or teacher route

Ask the learner to explain explain clinical ai so clinicians can follow the problem, data, model, output, evidence and limitations without guessing twice: once to a classmate and once to a method specialist. Praise a precise correction and a well-stated unknown; both show disciplined attention to evidence.

Early-career route

Create a one-page cair audit with source link, date checked, scope, method logic, decisive wording, contrary evidence, limit and next action. A colleague should be able to continue without guessing.

Continue with eduKate’s guides to reading course and career information and giving useful feedback in group projects.

USEFUL NEXT READING

Connect this narrow owner to its neighbour

Read the CHEERS-AI health-economic evaluation guide for a related but different reader job. Return here when the question is specifically how to explain clinical ai so clinicians can follow the problem, data, model, output, evidence and limitations without guessing. This boundary keeps one canonical owner per intent.

For English practice across research, school and career communication, visit the English at work and in public life guide.

FREQUENTLY ASKED QUESTIONS

Questions students and parents often ask

What is CAIR?

A Clinical AI Research checklist proposal for reporting medical AI research to clinicians.

Is CAIR a regulatory standard?

No. It supports reporting and appraisal; it does not approve a product.

Why separate participant and input eligibility?

Data-quality exclusions can change which patients the results represent.

Why state the model version?

Models, code and interfaces change, so results need an identifiable object.

Does a high score prove clinical benefit?

No. Workflow, decisions, harms and patient outcomes require further evaluation.

Why report errors?

Errors reveal where the model may fail and which consequences matter clinically.