LLM classification of chronic rhinosinusitis resectability under EPOS 2020

Source article: ChatGPT Versus DeepSeek in Resectability Categorization of Chronic Rhinosinusitis: A Comparison Based on EPOS 2020

Objective To evaluate and compare the performance of ChatGPT-5.5 (2026) and DeepSeek-V3 (2026) in classifying CRS resectability according to the EPOS 2020 guidelines, employing various prompting strategies. Methods In this retrospective study, 150 chronic rhinosinusitis cases were reviewed, and three rhinology experts classified resectability according to the EPOS 2020 criteria. GPT-5.5 and DeepSeek-V3 were evaluated using default-knowledge, in-context, and chain-of-thought (CoT) prompting based on accuracy, sen…

By the Truvace deskOctober 16, 20262 min read1 sourcePeer-reviewedTRV-2026-1612· cite & verify

Trace impact readingContested

P 69

Both sides are scored from claims and sources, not community votes.

G 65

In brief

Researchers retrospectively reviewed 150 chronic rhinosinusitis cases that three rhinology experts had classified for resectability under the EPOS 2020 criteria. They tested ChatGPT-5.5 and DeepSeek-V3 using default-knowledge, in-context and chain-of-thought prompts, and also compared ENT residents working with and without AI-generated reports.

Results matter because prompting strategy appeared to determine whether general-purpose chatbots could approximate expert categorization, and AI assistance improved resident accuracy. What remains uncertain is whether the findings hold outside this retrospective case set, since the authors stress further validation of the models as support tools before widespread clinical use.

Main points

Under default-knowledge prompting, DeepSeek-V3 had higher accuracy than GPT-5.5 (81.3% vs. 72.7%).
With chain-of-thought prompting, both models reached similarly high accuracy in resectability categorization.
Subgroup analyses by nasal polyp status and surgical eligibility showed consistent effects for GPT-5.5 but heterogeneous patterns for DeepSeek-V3.
CoT-assisted ENT residents were significantly more accurate than unassisted peers (GPT-5.5: 97.3% vs. 72.8%).

The gain

With chain-of-thought prompting, both GPT-5.5 and DeepSeek-V3 reached similarly high accuracy in classifying chronic rhinosinusitis resectability, and AI-generated reports raised ENT residents' accuracy to 97.3% in the GPT-5.5 arm versus 72.8% unassisted.

The problem

Models were not shown to be ready for routine use: DeepSeek-V3 gave heterogeneous results across subgroups, default-knowledge accuracy was only 72.7% for GPT-5.5, and the authors call for further validation before widespread clinical implementation.

The rundown

The study used 150 chronic rhinosinusitis cases, with resectability labels from three rhinology experts based on EPOS 2020 criteria. Performance was assessed with accuracy, sensitivity, specificity, PPV, NPV and F1 score.

Three prompting approaches were compared: default knowledge, in-context guidance and chain-of-thought reasoning. Prespecified subgroup analyses examined nasal polyp status and surgical eligibility to check whether prompting effects were consistent.

A separate comparison measured ENT resident performance with and without AI assistance. The authors report that AI-generated reports improved both accuracy and efficiency, but conclude that the models need further validation as support tools.

What this doesn’t fix

The authors state that the models should be further validated as support tools before widespread clinical implementation, and DeepSeek-V3 showed heterogeneous subgroup responses.