WHY ENGLISH?
Use four reporting pillars to make a medical AI claim understandable
MINIMAR stands for MINimum Information for Medical AI Reporting. Published in 2020, it proposes a compact way to describe medical AI so readers can understand intended predictions, target populations, hidden biases and the ability to generalise.
Its central reporting pillars are study population, model architecture, model evaluation and transparency. That four-part frame remains useful for students and early-career readers because it turns a dense paper into a set of precise English questions.
MINIMAR is a proposal for minimum reporting, not a guarantee of quality, fairness, safety or clinical benefit. Newer, design-specific guidance may be more appropriate for a current manuscript, but MINIMAR remains a clear lesson in how language connects people, models and evidence.
Article contents
Open all 19 chapters
- 1. Start With the Intended Prediction
- 2. Define the Target Population and Setting
- 3. Report Recruitment and Selection
- 4. Describe Demographics and Clinical Context
- 5. Name Data Sources and Measurement Processes
- 6. Report Missing Data and Data Quality
- 7. Prevent Patient and Time Leakage
- 8. Identify the Model Architecture and Version
- 9. Explain Feature Engineering and Preprocessing
- 10. Report Training, Tuning and Hyperparameter Search
- 11. Describe Outputs, Thresholds and Calibration
- 12. Use an Independent Evaluation Set
- 13. Choose Evaluation Metrics and Uncertainty
- 14. Perform Error and Subgroup Analysis
- 15. Compare With Clinical Baselines Fairly
- 16. Make Code, Parameters and Artefacts Available Responsibly
- 17. Discuss Generalisability and Hidden Bias
- 18. Separate Transparent Reporting From Clinical Readiness
- 19. Use MINIMAR as a Transferable Reading Habit
This article uses a fictional model that predicts unplanned return to hospital after discharge. The example shows how each reporting pillar changes what a reader can responsibly conclude.
The guide also distinguishes minimum information from sufficient evidence. A paper can report every requested detail and still require independent validation, risk-of-bias appraisal and clinical evaluation.
CHAPTER 1 OF 19 · Study population
1. Start With the Intended Prediction
Read the mechanism
State exactly what the model predicts, for whom, from which information and at what time.
A prediction target needs an event definition, horizon and use context. Vague phrases such as predicts risk hide when the prediction is made and what decision it may support.
Did You Know? MINIMAR was proposed to make intended predictions and target populations understandable before readers inspect technical performance.
Follow the evidence trail
The model estimates the probability of an unplanned hospital return within 30 days, using information available at discharge for adult medical patients.
Practical check
- Is the event defined?
- Is the prediction time fixed?
- Is the intended decision or user named?
Limits and responsible interpretation
A defined prediction does not show that acting on it improves care.
Rewrite the objective as input time, target event, horizon and intended user.
Study and career transfer: “Start With the Intended Prediction” trains the learner to convert a broad claim into named evidence and a repeatable action. In medical AI appraisal, that habit supports careful handovers, reviews and project decisions. At home, a parent can ask for one everyday explanation and one technical explanation; the two versions should differ in vocabulary, not in truth.
CHAPTER 2 OF 19 · Study population
2. Define the Target Population and Setting
Put the wording to work
Describe who the model is meant to serve and where it may be used.
Eligibility, setting, geography and care pathway shape prevalence, measurements and outcomes. The intended population should be distinguished from the available sample.
Did You Know? A model can be evaluated on a convenient sample that only partly represents its stated users.
Test the claim in context
The intended population is adults discharged from general medicine, but the sample comes from one tertiary hospital and excludes non-English records.
Practical check
- Are intended and observed populations separate?
- Are exclusions and setting stated?
- Are dates and locations visible?
Limits and responsible interpretation
A broad intended-use statement cannot overcome a narrow evaluation sample.
Create side-by-side columns for intended users and study participants.
For further study, this chapter turns reading into professional communication. A learner should be able to state the decision, point to the supporting record and repeat this practical move: Create side-by-side columns for intended users and study participants. That pattern is useful in prediction-model research, where confident language must still preserve uncertainty, scope and responsibility.
CHAPTER 3 OF 19 · Study population
3. Report Recruitment and Selection
Separate the parts
Explain how records entered the study and which steps removed them.
A flow from eligible population to analysed sample reveals selection mechanisms, missingness and potential bias.
Did You Know? Large datasets can still be selectively assembled through documentation, insurance or data-linkage requirements.
Rebuild the claim carefully
The paper reports all eligible discharges, exclusions for missing follow-up linkage and the characteristics of excluded patients.
Practical check
- Can every exclusion be counted?
- Are data-availability exclusions explained?
- Are included and excluded groups compared?
Limits and responsible interpretation
A flow diagram describes selection but does not eliminate selection bias.
Follow one hypothetical patient through every inclusion gate.
Revision transfer: turn “Report Recruitment and Selection” into a question, answer it from the source and underline the phrase that supports the answer. Then write one assumption beside it and test whether the conclusion changes. This compact routine builds evidence discipline for clinical data science while keeping English active rather than ornamental.
CHAPTER 4 OF 19 · Study population
4. Describe Demographics and Clinical Context
Make the record auditable
Report characteristics relevant to performance, fairness and transportability without reducing people to a checklist.
Counts and distributions should be chosen because they relate to data generation, outcomes or care access. Intersectional and contextual differences may matter.
Did You Know? Hidden bias can remain invisible when only an overall average population is described.
Check what another reader could verify
The study reports age, sex, language, comorbidity, discharge destination and hospital site, with missingness and privacy-aware grouping.
Practical check
- Are characteristics relevant to the use case?
- Is missingness shown?
- Are small groups protected from disclosure?
Limits and responsible interpretation
Demographic balance does not guarantee equitable outcomes or representative experience.
Explain why each reported characteristic could affect data, prediction or action.
Team transfer: assign an author, reviewer and future user to the problem in “Describe Demographics and Clinical Context.” The author explains the choice, the reviewer locates missing evidence and the future user asks whether the record is sufficient to act. In evidence review, this dialogue helps teams find ambiguity before it becomes an operational mistake.
CHAPTER 5 OF 19 · Study population
5. Name Data Sources and Measurement Processes
Read the mechanism
Clinical variables come from tests, notes, devices and administrative systems. Describe how each source was produced.
Measurement process affects meaning. A diagnosis code, laboratory value and clinician narrative have different error structures and timelines.
Did You Know? Medical AI often learns documentation patterns alongside clinical patterns.
Follow the evidence trail
The model uses laboratory results and coded diagnoses available by discharge. Notes added later are excluded to prevent future information leakage.
Practical check
- What system generated each input?
- When did it become available?
- Could documentation practice differ across groups?
Limits and responsible interpretation
A recorded variable may be a proxy for care processes rather than biology.
Add source, timestamp and plausible error to every important feature.
Study and career transfer: “Name Data Sources and Measurement Processes” trains the learner to convert a broad claim into named evidence and a repeatable action. In medical AI appraisal, that habit supports careful handovers, reviews and project decisions. At home, a parent can ask for one everyday explanation and one technical explanation; the two versions should differ in vocabulary, not in truth.
CHAPTER 6 OF 19 · Model architecture
6. Report Missing Data and Data Quality
Put the wording to work
State what is missing, why it may be missing and how the model handles it.
Missingness can carry clinical meaning or reflect unequal access. Imputation, indicators and complete-case analysis need clear justification.
Did You Know? A model may treat absence of a test as information about clinician decisions rather than patient state.
Test the claim in context
The study reports missing laboratory values by site and language group, uses a prespecified imputation method and tests sensitivity.
Practical check
- Is missingness quantified before modelling?
- Are handling rules reproducible?
- Are group differences examined?
Limits and responsible interpretation
No imputation method recovers information that was never observed without assumptions.
Write the assumption behind every missing-data decision.
For further study, this chapter turns reading into professional communication. A learner should be able to state the decision, point to the supporting record and repeat this practical move: Write the assumption behind every missing-data decision. That pattern is useful in prediction-model research, where confident language must still preserve uncertainty, scope and responsibility.
CHAPTER 7 OF 19 · Model architecture
7. Prevent Patient and Time Leakage
Separate the parts
Separate training and evaluation at the level where dependence occurs.
All records for one patient should remain within a partition when repeated visits exist. Time-aware splits may be needed when practice changes.
Did You Know? Random row splits can put information from the same person in both development and test data.
Rebuild the claim carefully
The model trains on earlier patients and evaluates on later discharges, grouping repeated admissions by person.
Practical check
- What is the partition unit?
- Can future information enter inputs?
- Does the test reflect intended time use?
Limits and responsible interpretation
A temporal split still may not represent another hospital or later policy era.
Describe the split as a clinical timeline, not a software command.
Revision transfer: turn “Prevent Patient and Time Leakage” into a question, answer it from the source and underline the phrase that supports the answer. Then write one assumption beside it and test whether the conclusion changes. This compact routine builds evidence discipline for clinical data science while keeping English active rather than ornamental.
CHAPTER 8 OF 19 · Model architecture
8. Identify the Model Architecture and Version
Make the record auditable
Name the algorithm, implementation, version and major components.
Architecture reporting should be detailed enough to understand the modelling family and reproduce the configured system, including ensembles or pretrained components.
Did You Know? Two studies can use the same model label while applying different software versions and preprocessing.
Check what another reader could verify
The paper identifies a gradient-boosted tree implementation, package version, hyperparameter set and calibration layer.
Practical check
- Is the exact implementation named?
- Are all components and versions listed?
- Are proprietary unknowns disclosed?
Limits and responsible interpretation
Technical detail alone does not establish that the architecture suits the clinical question.
Create a versioned model card tied to the paper’s release.
Team transfer: assign an author, reviewer and future user to the problem in “Identify the Model Architecture and Version.” The author explains the choice, the reviewer locates missing evidence and the future user asks whether the record is sufficient to act. In evidence review, this dialogue helps teams find ambiguity before it becomes an operational mistake.
CHAPTER 9 OF 19 · Model architecture
9. Explain Feature Engineering and Preprocessing
Read the mechanism
Describe transformations from raw clinical data to model inputs.
Scaling, encoding, aggregation and text processing can affect performance and leakage. Rules should be fit only on development data when appropriate.
Did You Know? A seemingly simple derived feature can contain future information if its time window crosses the prediction moment.
Follow the evidence trail
Laboratory values are summarised using measurements before discharge; normalisation parameters come only from training data.
Practical check
- Are transformation formulas available?
- Are time windows explicit?
- Were preprocessing parameters isolated from test data?
Limits and responsible interpretation
Reproducible preprocessing can still encode questionable proxies.
For every engineered feature, state clinical meaning and information time.
Study and career transfer: “Explain Feature Engineering and Preprocessing” trains the learner to convert a broad claim into named evidence and a repeatable action. In medical AI appraisal, that habit supports careful handovers, reviews and project decisions. At home, a parent can ask for one everyday explanation and one technical explanation; the two versions should differ in vocabulary, not in truth.
CHAPTER 10 OF 19 · Model architecture
10. Report Training, Tuning and Hyperparameter Search
Put the wording to work
Show how the final model was selected without using the test set.
Separate optimisation data from final evaluation and describe search space, objective, stopping and random seeds.
Did You Know? Repeatedly checking test results turns the test set into part of development even if files remain separate.
Test the claim in context
Hyperparameters are chosen by cross-validation inside the training period; the later test set is opened once after locking decisions.
Practical check
- Which data selected the model?
- Is the search space reported?
- When was the final configuration frozen?
Limits and responsible interpretation
A clean procedure reduces optimistic bias but does not guarantee external validity.
Keep a decision log that records changes before test access.
For further study, this chapter turns reading into professional communication. A learner should be able to state the decision, point to the supporting record and repeat this practical move: Keep a decision log that records changes before test access. That pattern is useful in prediction-model research, where confident language must still preserve uncertainty, scope and responsibility.
CHAPTER 11 OF 19 · Model evaluation
11. Describe Outputs, Thresholds and Calibration
Separate the parts
A probability, score and binary alert are different outputs. Report how one becomes another.
Thresholds should be tied to intended consequences and calibration should show whether predicted risks correspond to observed frequencies.
Did You Know? A model can rank patients well while giving poorly calibrated probabilities.
Rebuild the claim carefully
The study reports risk probabilities, calibration plots and two prespecified thresholds reflecting different service capacities.
Practical check
- Is the raw output defined?
- Who chose the threshold and why?
- Is calibration assessed?
Limits and responsible interpretation
A threshold suitable for one hospital’s resources may be unsuitable elsewhere.
Translate threshold choice into expected alerts, misses and workload.
Revision transfer: turn “Describe Outputs, Thresholds and Calibration” into a question, answer it from the source and underline the phrase that supports the answer. Then write one assumption beside it and test whether the conclusion changes. This compact routine builds evidence discipline for clinical data science while keeping English active rather than ornamental.
CHAPTER 12 OF 19 · Model evaluation
12. Use an Independent Evaluation Set
Make the record auditable
Final performance should come from data not used to train, tune or select the model.
Independence includes people, time, sites and decision processes. Report exactly how the evaluation set differs.
Did You Know? The word validation is ambiguous unless the relationship to development is described.
Check what another reader could verify
Internal testing uses later patients at the development hospital; external testing uses a second health system with independent extraction.
Practical check
- Was the set untouched during development?
- Are site and time relationships explicit?
- Were thresholds fixed beforehand?
Limits and responsible interpretation
One external test does not prove universal generalisability.
Replace validation with a precise description of the data relationship.
Team transfer: assign an author, reviewer and future user to the problem in “Use an Independent Evaluation Set.” The author explains the choice, the reviewer locates missing evidence and the future user asks whether the record is sufficient to act. In evidence review, this dialogue helps teams find ambiguity before it becomes an operational mistake.
CHAPTER 13 OF 19 · Model evaluation
13. Choose Evaluation Metrics and Uncertainty
Read the mechanism
Report metrics appropriate to prevalence, ranking, calibration and decisions, with confidence intervals.
No single number captures the clinical meaning of predictions. Baselines and decision-relevant summaries help readers interpret scale.
Did You Know? Accuracy can look high when the outcome is rare even if the model misses most events.
Follow the evidence trail
The paper reports sensitivity, specificity, predictive values, area measures, calibration and confidence intervals at prespecified thresholds.
Practical check
- Do metrics fit outcome frequency?
- Is uncertainty visible?
- Is current practice a comparator?
Limits and responsible interpretation
Good statistical performance is not the same as useful clinical action.
Explain each metric in one sentence without jargon.
Study and career transfer: “Choose Evaluation Metrics and Uncertainty” trains the learner to convert a broad claim into named evidence and a repeatable action. In medical AI appraisal, that habit supports careful handovers, reviews and project decisions. At home, a parent can ask for one everyday explanation and one technical explanation; the two versions should differ in vocabulary, not in truth.
CHAPTER 14 OF 19 · Model evaluation
14. Perform Error and Subgroup Analysis
Put the wording to work
Inspect who is misclassified, how errors arise and which failures matter most.
Error analysis should connect examples to data quality, model behaviour and consequences. Subgroup analysis needs adequate support and uncertainty.
Did You Know? Overall metrics can conceal severe failures in smaller populations.
Test the claim in context
False negatives cluster among patients discharged to facilities with fragmented follow-up records. The study treats this as a data-system finding, not a biological conclusion.
Practical check
- Are errors sampled systematically?
- Are severe cases described?
- Are subgroup estimates supported?
Limits and responsible interpretation
Post hoc patterns can generate hypotheses but may not support firm causal claims.
Read five consequential errors before celebrating the average.
For further study, this chapter turns reading into professional communication. A learner should be able to state the decision, point to the supporting record and repeat this practical move: Read five consequential errors before celebrating the average. That pattern is useful in prediction-model research, where confident language must still preserve uncertainty, scope and responsibility.
CHAPTER 15 OF 19 · Model evaluation
15. Compare With Clinical Baselines Fairly
Separate the parts
A model should be compared with relevant current practice under equivalent conditions.
Comparators may include simple scores, clinician judgement or existing workflows. Report information access, timing and uncertainty.
Did You Know? A complex model’s apparent advantage can disappear when compared with a strong, properly implemented baseline.
Rebuild the claim carefully
The new model and existing score use the same prediction time and eligible population; clinicians are evaluated separately with stated information.
Practical check
- Is the comparator relevant?
- Are inputs and timing equivalent?
- Are differences clinically meaningful?
Limits and responsible interpretation
Better retrospective performance does not prove better outcomes after deployment.
Ask whether the comparison changes a real decision.
Revision transfer: turn “Compare With Clinical Baselines Fairly” into a question, answer it from the source and underline the phrase that supports the answer. Then write one assumption beside it and test whether the conclusion changes. This compact routine builds evidence discipline for clinical data science while keeping English active rather than ornamental.
CHAPTER 16 OF 19 · Transparency and generalisability
16. Make Code, Parameters and Artefacts Available Responsibly
Make the record auditable
Transparency includes access to code, configuration, data dictionaries and model details, with justified restrictions.
An availability statement should locate each artefact, version it and explain barriers to sharing sensitive data or proprietary components.
Did You Know? MINIMAR includes transparency because readers need to understand and reproduce the model, not only see its score.
Check what another reader could verify
The team shares preprocessing code, parameters, a synthetic test fixture and a data dictionary while clinical records remain governed.
Practical check
- Can artefacts be found?
- Do versions match the paper?
- Are restrictions and access routes clear?
Limits and responsible interpretation
Shared code does not guarantee reproducibility when data or dependencies differ.
Create a reproducibility manifest before submission.
Team transfer: assign an author, reviewer and future user to the problem in “Make Code, Parameters and Artefacts Available Responsibly.” The author explains the choice, the reviewer locates missing evidence and the future user asks whether the record is sufficient to act. In evidence review, this dialogue helps teams find ambiguity before it becomes an operational mistake.
CHAPTER 17 OF 19 · Transparency and generalisability
17. Discuss Generalisability and Hidden Bias
Read the mechanism
Explain where performance may change and which unmeasured processes could create bias.
Generalisability depends on population, prevalence, measurement, workflow, equipment and policy. The discussion should connect each limit to evidence.
Did You Know? MINIMAR explicitly highlights target populations, hidden biases and the ability to generalise.
Follow the evidence trail
The authors limit claims to adult medical discharges in two hospitals and identify language exclusions and follow-up linkage as likely bias sources.
Practical check
- Are claims bounded by observed settings?
- Are likely shifts named?
- Are unsupported populations excluded from conclusions?
Limits and responsible interpretation
A limitation list is useful only when it changes interpretation or next steps.
Turn each limitation into a replication or monitoring plan.
Study and career transfer: “Discuss Generalisability and Hidden Bias” trains the learner to convert a broad claim into named evidence and a repeatable action. In medical AI appraisal, that habit supports careful handovers, reviews and project decisions. At home, a parent can ask for one everyday explanation and one technical explanation; the two versions should differ in vocabulary, not in truth.
CHAPTER 18 OF 19 · Transparency and generalisability
18. Separate Transparent Reporting From Clinical Readiness
Put the wording to work
A complete report can still describe a model that is biased, ineffective or unsafe.
Use reporting guidance to find information, then apply risk-of-bias, validation, implementation and governance tools appropriate to the decision.
Did You Know? EQUATOR notes that reporting guidelines are not design or quality-appraisal tools.
Test the claim in context
The return-risk model is fully documented but lacks prospective workflow evaluation. The team labels it research-only and plans a silent prospective study.
Practical check
- What question does the reporting checklist answer?
- What appraisal remains necessary?
- Is deployment language proportionate?
Limits and responsible interpretation
Transparency is a condition for responsible scrutiny, not certification.
End every review with two columns: reported clearly and supported adequately.
For further study, this chapter turns reading into professional communication. A learner should be able to state the decision, point to the supporting record and repeat this practical move: End every review with two columns: reported clearly and supported adequately. That pattern is useful in prediction-model research, where confident language must still preserve uncertainty, scope and responsibility.
CHAPTER 19 OF 19 · Transparency and generalisability
19. Use MINIMAR as a Transferable Reading Habit
Separate the parts
The four pillars provide a compact method for reading any data-driven claim.
Ask who is represented, what system was built, how it was tested and what can be checked. Then add the domain-specific questions the decision requires.
Did You Know? Minimum information is useful in classrooms because it reduces a complex paper to a navigable evidence map.
Rebuild the claim carefully
A student reads an education-prediction article using population, architecture, evaluation and transparency headings, then identifies where medical-specific assumptions no longer apply.
Practical check
- Can the learner locate each pillar?
- Can they explain missing information?
- Can they avoid transferring conclusions beyond the sample?
Limits and responsible interpretation
A compact framework should lead to deeper inquiry, not replace specialist standards.
Use the four headings for the next research paper you read and write one evidence-backed sentence under each.
Revision transfer: turn “Use MINIMAR as a Transferable Reading Habit” into a question, answer it from the source and underline the phrase that supports the answer. Then write one assumption beside it and test whether the conclusion changes. This compact routine builds evidence discipline for clinical data science while keeping English active rather than ornamental.
PRACTICE LAB · FOUR-PILLAR EVIDENCE MAP
How to Read a Medical AI Abstract Before Opening the Full Paper
Population
Underline every phrase that identifies people, setting, dates and eligibility. Write “intended” beside the population the model is meant to serve and “observed” beside the people actually studied. If the abstract gives only a sample size, record that population information is incomplete and look for the flow diagram and characteristics table in the full paper.
Architecture
Circle the model family, inputs, output and development stage. Note whether the authors developed, tuned or only evaluated the system. An abstract rarely contains the whole implementation, but it should not lead readers to confuse a commercial model evaluation with a new model or a risk score with a clinical recommendation.
Evaluation
Box the test relationship, metrics, comparator and uncertainty. Ask whether the test data were independent at the patient, time and site levels. Translate one headline metric into an ordinary-language consequence. If that translation is impossible, the abstract may not provide enough decision context.
Transparency
Place a star beside statements about code, data, protocols, registration or artefact availability. Then list the information that must be checked in the full text: missing data, preprocessing, thresholds, calibration, subgroup errors, limitations and conflicts. Minimum reporting is a doorway into scrutiny, not a shortcut around it.
Write the two-sentence abstract verdict
Sentence one states the reported finding with population, model and test context. Sentence two states the most important limit or unanswered question. This disciplined pair is useful for literature reviews, presentations and interviews because it shows both comprehension and restraint.
Family practice: A parent can ask, “Who was studied? What was built? How was it tested? What can we check?” A student who can answer those four questions has already moved beyond headline reading.
QUESTIONS PARENTS, STUDENTS AND EARLY-CAREER READERS ASK
Frequently asked questions
Why use MINIMAR when newer AI guidance exists?
MINIMAR offers a compact four-pillar reading frame: population, architecture, evaluation and transparency. For current manuscript preparation, choose the latest design- and domain-specific reporting guidance as well.
What does minimum information mean?
It means a baseline record needed to understand the study. It does not mean the evidence is sufficient for clinical use or that every risk has been addressed.
Why distinguish the intended population from the study sample?
The intended population defines where authors hope the model will be used. The sample shows where evidence actually came from. Generalisation requires examining the gap.
What is the most common evaluation trap for beginners?
Treating one impressive score as the whole result. Beginners should also inspect the test relationship, prevalence, calibration, uncertainty, errors, comparators and decision consequences.
Can the four pillars be used outside medicine?
Yes, as a reading habit. Replace medical assumptions with the relevant domain knowledge and ask who is represented, what was built, how it was tested and what can be checked.
CONTINUE WITH VERIFIED SOURCES
Useful next reading
Read the original proposal and current EQUATOR record, then select newer design-specific guidance when preparing present-day research.
- MINIMAR primary open-access article — original proposal and four-pillar rationale
- EQUATOR MINIMAR record — scope and bibliographic verification
- PubMed MINIMAR record — primary article metadata
- WHO ethics and governance of AI for health — responsible interpretation beyond minimum reporting
Continue across the eduKate English and evidence trail
- Reading MI-CLAIM clinical AI models — extend the minimum frame into a six-part reporting checklist
- Reporting CLAIM 2024 medical-imaging AI — apply specialised guidance to imaging studies
- Why English matters for explaining scientific evidence — practise claim-evidence boundaries
- Why English matters for course and career information — carry evidence literacy into choices
English makes a model claim inspectable when population, architecture, evaluation and transparency stay connected. It is one layer of evidence practice, alongside technical expertise, ethical judgement and multilingual communication.