Compare healthcare AI evidence and its limits
By ChatGPT Health. Educational synthesis prepared with AI assistance. Sources checked September 28, 2026. Named clinical review: pending. This page is not a clinical protocol.
Compare study designs, absolute results, applicability and unanswered questions across five clinical areas. Keep the conclusion at the level of the outcome that was measured.
How to use this comparison
These are selected examples across different clinical tasks. They are not a systematic review or a ranking of products. Outcomes, populations and comparators differ, so the numbers cannot be pooled or used to declare one specialty’s AI more effective than another.
Medication alerts
Design and comparison: Prospective observation in one ward; existing decision support remained in use.
Measured result: Alerts affected 0.4% of orders; 43% prompted subsequent order changes.
Interpretation: Clinician rated usefulness and prescribing behavior are not a randomized patient harm endpoint.
Before local use: Test missed errors, pharmacist workload and adverse events against local practice.
Mental health chatbot
Design and comparison: 210 adults; specialized Therabot versus waitlist; four week intervention, eight week follow up.
Measured result: At four weeks, mean depression symptom score change was −6.13 versus −2.63.
Interpretation: This is a symptom scale result, not a remission rate or a comparison with human therapy. Children and general chatbots are outside the tested claim.
Before local use: Examine deterioration, escalation, retention and longer follow up; ask for an active comparator.
Pediatric diabetes eye screening
Design and comparison: 164 participants aged 8 to 21; AI exam at a diabetes visit versus referral plus education.
Measured result: Six month eye exam completion: 100% versus 22%.
Interpretation: Completion is a care process outcome. Vision benefit and broad pediatric diagnostic validity were not established by this result.
Before local use: Check the current intended use, local access and whether abnormal findings reach appropriate follow up.
Oncology screening
Design and comparison: MASAI randomized trial in Sweden; 105,915 participants analyzed; AI supported versus standard double reading.
Measured result: Interval cancers: 1.55 versus 1.76 per 1,000. Ratio 0.88, 95% CI 0.65 to 1.18. Sensitivity 80.5% versus 73.8%.
Interpretation: Noninferiority was established with a prespecified 20% margin. The confidence interval includes no difference; superiority in interval cancers was not established.
Before local use: Match reading workflow, equipment and population. Track recalls, harms and downstream outcomes.
Stroke rehabilitation devices
Design and comparison: Cochrane synthesis of 101 trials, 4,224 adults; devices plus physiotherapy versus physiotherapy or usual care.
Measured result: Independent walking at treatment completion probably improved; estimated number needed to treat was nine.
Interpretation: Speed and distance did not clearly improve, and follow up benefit was uncertain. Robotics evidence does not isolate an AI contribution.
Before local use: Match impairment and recovery stage. Measure sustained function, supervision needs and burden.
Record what you cannot verify
Use the source paper and supplementary material to complete funding, commercial interests, exclusions, missing data, adverse events and subgroup results. If a detail is unavailable, mark it unknown. This comparison does not assign formal GRADE ratings; certainty language for rehabilitation follows the Cochrane review.
Sources
- Medication safety study, 2019
- Therabot trial, 2025
- ACCESS trial, 2024
- MASAI analysis, 2026
- Cochrane rehabilitation review, 2025