Ear-disease AI narrowed the trainee–expert gap—but also changed 46 correct answers to wrong ones
Thirteen clinicians re-read 100 disease-enriched ear images with AI advice. Accuracy rose most for junior residents, but the single-centre retrospective study cannot show how the tool performs in routine care.
The short version
- What happened
- A South Korean model trained on 7,335 otoendoscopy images was tested with three professors, five senior residents and five junior residents on an independent 100-image set.
- Who it affects
- A less experienced clinician may gain a useful second opinion from image support, but a confident model answer can also pull a correct judgement off course. Patients need prospective evidence that the combined clinician–AI workflow improves real diagnoses and care.
- What remains unproven
- Single-centre retrospective study using one South Korean image database and a 100-image assessment set.
Keep in mind
Single-centre retrospective study using one South Korean image database and a 100-image assessment set.
Disease-enriched set excluded fully normal images and cases with discordant annotations, limiting routine-care generalisability.
The same clinicians read the images twice, three months apart; intrareader variability may explain some changes.
No prospective workflow, referral, treatment or patient outcome was measured.
Funded by the Korean Ministry of Food and Drug Safety; the authors did not disclose competing interests in the publisher record.
Source analysis · 1 linked source. Primary analysed source date: 30 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Trainees improved most on the same 100 images
AI advice substantially raised image-classification accuracy in a peer-reviewed South Korean reader study. Three otology professors, five senior residents and five junior residents first classified 100 otoendoscopy images without assistance, then repeated the task with model predictions three months later. For the main categories—chronic otitis media, otitis media with effusion or neither—accuracy rose by 11.0 percentage points for professors, 28.8 points for senior residents and 32.2 points for junior residents.
The model had been trained on 7,335 images from Asan Medical Center. On the 100-image assessment set, its stand-alone primary-class accuracy was 95.0%. The gap between professors and junior residents narrowed from 24.3 to 3.1 percentage points. That is useful evidence about decision support under a one-minute-per-image constraint, not proof of faster or safer patient care.[1]
AI corrected many errors and introduced some
The paired decisions show both sides of assistance. Across 1,300 primary-class assessments, clinicians' first answer differed from the AI recommendation 499 times. They accepted the recommendation in 395 of those cases. Overall, support corrected 384 initial errors, while 46 initially correct decisions became incorrect. Junior residents were especially likely to accept AI advice and were least often right when they rejected it.
The assessment deliberately excluded fully normal images and therefore did not reflect the prevalence seen in ordinary clinics. Difficult cases with discordant annotations were also removed, the readers saw the same images twice, and all training data came from one centre and endoscope setting. Multicentre prospective studies should measure diagnostic errors, referrals, treatment and patient outcomes before broad deployment.[1]
The distinction that matters
What changed when clinicians saw the model's answer
Each of 13 clinicians classified the same 100 images before and after AI advice.
Junior residents
54.4% to 86.6%
Largest primary-class accuracy gain
Senior residents
65.0% to 93.8%
Highest assisted group accuracy
All readers
384 errors fixed; 46 errors introduced
Help and automation risk appeared together