Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2004
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-20T10:42:26.657502Z
status: published
lens: p_space
sector: health
headline: Differential Diagnosis Performance of ChatGPT in Oral and Maxillofacial Pathology
dek: Objective: To evaluate ChatGPT (GPT-5.1) for ranked differential diagnoses in oral and maxillofacial pathology, assess image effects on diagnostic performance and examine output stability across repeated runs.
Methods: This retrospective diagnostic accuracy study included 160 biopsy-confirmed cases. Cases were evaluated using a standardised prompt under two conditions: text-only and text plus clinical photographs and/or radiographs. Each case underwent three independent runs. Performance was measured using Top-1…
gain_title: (none)
problem_title: ChatGPT listed the biopsy-confirmed diagnosis first in only about 43 to 44 percent of oral and maxillofacial pathology cases, and adding clinical images did not significantly improve accuracy, so it is not reliable enough for clinical decision support.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: ChatGPT listed the biopsy-confirmed diagnosis first in only about 43 to 44 percent of oral and maxillofacial pathology cases, and adding clinical images did not significantly improve accuracy, so it is not reliable enough for clinical decision support.
problem_evidence: ChatGPT ranked the reference diagnosis first in fewer than half of cases | Images did not significantly improve accuracy
quick_read: Researchers tested ChatGPT (GPT-5.1) on 160 biopsy-confirmed oral and maxillofacial pathology cases, asking it for ranked differential diagnoses. Each case was run three times, once with text only and once with text plus clinical photographs or radiographs.

The results show a gap between suggesting the right answer somewhere in a list and committing to it as the top choice. That limits use in clinical settings, though the authors see room for hypothesis generation and education. How well images could help with more standardised inputs remains unclear.
limitation: The authors note that heterogeneous and non-standardised visual inputs may have limited detection of any benefit from adding images, so the lack of an image effect may not be conclusive.
tag: Evidence-backed problem
key_points: Top-1 accuracy was 44.0% with text only and 43.3% with text plus images. | Top-10 accuracy reached 80.7% for text-only and 82.9% for text plus image inputs. | No significant differences between input conditions were seen at any Top-K threshold. | Top-1 results were stable across three runs, but broader Top-K thresholds varied in the text-only condition.
rundown: The study was a retrospective diagnostic accuracy analysis using 160 cases confirmed by biopsy. A standardised prompt was applied under two input conditions, and accuracy was scored at Top-1, Top-3, Top-5 and Top-10 thresholds.

Accuracy rose as the list widened, with the reference diagnosis appearing in about four of five cases at the Top-10 level. Adding photographs or radiographs did not produce a statistically significant gain at any threshold.

The authors conclude the findings support further investigation for hypothesis generation and educational use, not clinical decision support or replacement of expert judgement and histopathological confirmation.
sources:
- peer_reviewed | Oral Diseases | https://doi.org/10.1111/odi.70501 | 2026-10-19
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

0bc02721a17c8dd0a9872e20632608916c9b26923d84ee1488ee88dd00c29e04

previous

0000000000000000000000000000000000000000000000000000000000000000