Before You Believe a Health AI Headline: 8 Questions to Ask

What is new

By healthcarewithai.com Editorial Team

This revised guide separates questions that headlines can blur: does an AI predict accurately, does using it improve care, and what regulatory requirements has it met? The CONSORT-AI paper published on 9 September 2020 described predominantly early computer-based validation evidence during its development. That historical observation does not establish how common different study designs are in 2026.

How it works

1. What was actually tested?

Retrospective evaluation analyses existing data; the label does not establish whether care previously involved AI. Prospective evaluation follows a planned assessment as new data accrue. In prospective silent evaluation, outputs are withheld from care decisions. In prospective interventional evaluation, outputs are made available to influence care. Prospective data collection alone therefore does not mean AI affected treatment. TRIPOD+AI addresses data sources and evaluation; DECIDE-AI addresses early clinical evaluation of AI decision support.

A randomised trial assigns patients, clinicians or sites by chance to comparison groups. A well-conducted trial measuring appropriate patient outcomes offers strong evidence about effects, but randomisation does not guarantee benefit, sound execution or applicability elsewhere; CONSORT-AI supports scrutiny of design and reporting.

2. Compared against what?

Ask whether the comparator represents usual care: clinician experience, available information, time constraints and workflow matter. Does the study compare AI alone with clinicians, or clinicians using AI with clinicians without it? These answer different questions. CONSORT-AI's intervention and human-AI interaction requirements help readers assess whether the comparison matches the claimed use.

3. Tested in whom?

External validation evaluates a model using data independent of its development, not necessarily data from another hospital. Geographical validation is one form; evaluation in a later population at the same institution can be another. A random holdout from the development dataset is internal testing, not automatically external validation. TRIPOD+AI specifies that evaluation data must be distinct from training, hyperparameter-tuning and model-selection data, without overlapping participants. Ask whether the population and setting resemble the intended users, and examine subgroup results and uncertainty.

4. Which outcome was measured?

Matching reference labels measures predictive performance, not necessarily clinical benefit. Look for outcomes such as complications, quality of life, unnecessary procedures or survival, alongside harms. Faster diagnosis may matter, but does not alone establish better health. DECIDE-AI distinguishes clinical performance, safety and human factors; CONSORT-AI supports transparent reporting of intervention effects.

5. What do sensitivity and specificity mean?

FDA's diagnostic-test guidance defines sensitivity as the proportion of people with the condition who test positive, and specificity as the proportion without it who test negative. High sensitivity means a low proportion of missed cases among affected people; high specificity means a low proportion of false positives among unaffected people. Neither guarantees small absolute error counts.

For a scored test, changing the decision threshold commonly trades sensitivity against specificity. Positive predictive value (the proportion of positive results that are true positives) is defined in the guidance's glossary as true positives divided by all positive results, and the guidance's prevalence example shows why that ratio can shift with the population: because PPV depends on prevalence as well as sensitivity and specificity, even a fairly specific test can generate many false positives when screening for a rare condition. Look for confidence intervals, not just headline percentages.

6. What does the regulatory label mean?

Marketing authorisation, conformity marking and registration are different statuses. Regulatory pathways address applicable safety and performance or effectiveness requirements; they do not necessarily demonstrate improved patient outcomes compared with usual care.

United States: 510(k) clearance requires substantial equivalence to a legally marketed predicate: the same intended use and either the same technological characteristics, or different characteristics that raise no different safety/effectiveness questions, with supporting evidence demonstrating that the device is as safe and effective as the predicate. De Novo authorisation addresses novel devices without a predicate whose risks can be controlled as class I or II. PMA approval is a distinct pathway generally for class III devices, requiring sufficient valid scientific evidence of safety and effectiveness for intended use.

European Union: Under MDR 2017/745, Article 2(1), software with relevant medical intended purposes can qualify as a medical device. Articles 20, 52 and 61 address CE marking, conformity assessment and clinical evaluation respectively. Laboratory-diagnostic software qualifying as an in vitro diagnostic device instead falls under IVDR 2017/746; Articles 2(2), 48 and 56 address its definition, conformity assessment and performance evaluation.

United Kingdom: MHRA guidance checked for this revision distinguishes Great Britain (England, Scotland and Wales) from Northern Ireland. Great Britain requires device registration with the MHRA, which is not endorsement or a substitute for applicable market-access requirements. Devices can enter Great Britain through the UKCA route with approved-body assessment where required, while eligible CE-marked devices can also be accepted on the Great Britain market under the UK's current arrangements and transition conditions covering EU-approved devices (check the operative rules at publication, as these have changed repeatedly). Northern Ireland follows EU MDR/IVDR requirements, with CE marking and the UKNI indication where applicable. Recheck operative rules and transition conditions at publication.

7. Who funded it, and was it peer reviewed?

Check funding, conflicts and the sponsor's role in design, analysis and reporting, as requested by TRIPOD+AI. Funding alone does not establish validity or invalidity. Independent replication adds another check. medRxiv warns that its preprints have not been certified by peer review and should not guide clinical practice.

8. What happens in real-world use?

Ask about missing inputs, workflow failures, software updates and errors. CONSORT-AI requests reporting of algorithm version, input handling, human interaction and error analysis. A study of one version does not automatically validate later versions.

How strong is the evidence

Use this progression as a guide, not a quality score: retrospective evaluation assesses existing data; independent external evaluation tests transportability; prospective silent evaluation checks performance without influencing care; prospective interventional evaluation studies use in care. Well-conducted randomised comparisons with appropriate outcomes can provide strong causal evidence. Sample size, bias, uncertainty, harms and relevance matter at every stage.

Who could benefit

Patients, clinicians, hospital buyers and journalists can use these questions to match confidence to evidence rather than reject or endorse AI categorically.

Limitations and next steps

This is an educational guide, not a new clinical study or regulatory determination. Reporting guidelines are not quality scores. Read methods, check the exact device and intended use, and verify current jurisdiction-specific rules.

This is not medical advice; talk to your doctor.