A study from Seoul Asan Medical Center demonstrates that artificial intelligence significantly bridges the experience gap in otoscopic evaluations.

SEOUL — Artificial intelligence is proving to be a powerful ally for early-career physicians. A newly published study shows that integrating a deep learning-based diagnostic tool into ear examinations substantially sharpens accuracy, particularly among medical residents who are still building their clinical expertise.

The research initiative was spearheaded by Professor Ahn Jung-ho from the Department of Otolaryngology and Professor Kwon Ji-hoon from the Department of Biomedical Informatics at Seoul Asan Medical Center. The team trained a specialized convolutional network using a dataset of 7,335 clinical otoscopy images gathered between 2018 and 2020.

Middle ear complications like chronic otitis media and otitis media with effusion can present subtle visual overlaps during routine otoscopy, often making reliable diagnosis heavily reliant on a clinician's years on the job. To address this hurdle, the hospital's AI model was engineered to classify these distinct conditions, alongside recognizing secondary anomalies such as myringitis, attic retractions, and fungal infections. The system functions as a collaborative aid: it highlights predicted pathology based on visual cues, leaving the ultimate diagnostic call to the human physician.

In internal cross-validation tests, the AI system itself achieved an impressive 94.8% accuracy rate across primary diagnostic categories. When tested against a blind validation set of 100 separate images, the algorithm maintained a 95% accuracy score.

To measure real-world clinical impact, the researchers enlisted a diverse cohort of 13 medical professionals—comprising three expert otolaryngology professors, five senior residents, and five junior residents. Participants initially evaluated a batch of 100 test images unaided. After a washout period of three months, the same clinicians reviewed the exact same image set while incorporating recommendations from the AI tool.

The comparative data revealed a statistically significant performance boost across all experience tiers. Junior residents experienced the most dramatic turnaround: their average diagnostic accuracy leaped by 32.2 percentage points, soaring from 54.4% to 86.6%. Senior residents and senior faculty members also saw noticeable gains of 28.8 and 11 percentage points, respectively. Notably, the accuracy gap between veteran professors and junior trainees narrowed from a wide 24.3 percentage points down to a mere 3.1 percentage points when the AI was employed.

Further error-analysis indicated that out of all independent reviews, 384 previous misdiagnoses (29.5%) were successfully corrected to accurate conclusions with the AI's guidance. Conversely, a small minority of correct initial opinions—roughly 46 instances, or 3.5%—were incorrectly overridden after consulting the software.

"Our findings demonstrate that AI can effectively elevate the diagnostic capabilities of less-seasoned medical personnel," noted Professor Ahn Jung-ho. "With ongoing validation trials ahead, we anticipate this technology will become a valuable asset for both clinical workflows in middle ear disorder management and resident education."

The findings of this research have been formally released in the peer-reviewed international journal JMIR Medical Informatics.

[ⓒ Global Economic Times. 무단전재-재배포 금지]