68/100

Moderate evidence

Cite
More
EmailLinkedInX

AI models matched endoscopists at spotting stomach lesions likely to turn worse, study of 329 patients finds

Study title: Retrieval-augmented generation models for predicting pathological upgrading in endoscopically resected gastric low-grade intraepithelial neoplasia: A dual-center evaluation

Evidence type
Controlled or validated comparison
Sample
n = 372
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Retrieval-augmented generation models for predicting pathological upgrading in endoscopically resected gastric low-grade intraepithelial neoplasia: A dual-center evaluation

A retrospective study at two tertiary hospitals tested whether AI language models could predict which gastric low-grade intraepithelial neoplasia lesions would turn out to be high-grade neoplasia or early gastric cancer after endoscopic removal. It covered 329 patients with 372 lesions, and 148 lesions (40%) were upgraded.

Read the study abstract

BACKGROUND Gastric low-grade intraepithelial neoplasia (LGIN) exhibits highly variable clinical outcomes. The discordance between histological findings from endoscopic forceps biopsy and endoscopic submucosal dissection (ESD) complicates the formulation of appropriate treatment strategies. AIM To evaluate a prediction framework using large language models (LLMs) enhanced with retrieval-augmented generation (RAG) for the pathological upgrading of gastric LGIN. METHODS We retrospectively enrolled 329 patients with…

The gain

In a retrospective study of 329 patients with 372 gastric lesions, AI language models given a literature knowledge base predicted which low-grade lesions were upgraded after removal with an AUC of 0.873, comparable to two experienced endoscopists.

In brief

The retrieval-augmented DeepSeek-V3 scored an AUC of 0.873 and Qwen3-235b 0.862. Neither differed significantly from two experienced endoscopists or a generalized linear mixed model baseline. Biopsy and removal results often disagree, which complicates treatment choices, so a tool that flags likely upgrades could inform decisions about removal.

The work is retrospective and the authors say the models need prospective validation before they can support clinical decisions.

Main points

  1. The study enrolled 329 patients with 372 gastric lesions from two tertiary hospitals; all had a biopsy diagnosis of low-grade neoplasia before endoscopic submucosal dissection.
  2. 148 lesions (40%) were upgraded to high-grade intraepithelial neoplasia or early gastric cancer.
  3. RAG-enhanced DeepSeek-V3 reached an AUC of 0.873 (95% CI 0.828-0.901); Qwen3-235b reached 0.862 (95% CI 0.818-0.893).
  4. DeepSeek-V3 had 85.1% sensitivity, 80.4% specificity, 74.1% positive and 89.1% negative predictive value at a 0.357 threshold.
  5. Qwen3-235b had 84.5% sensitivity, 77.2% specificity, 71.0% positive and 88.3% negative predictive value at a 0.453 threshold.
  6. Decision curve analysis showed clinical utility across threshold probabilities of 15%-80%, and accuracy held up in lesions measuring 1-2 cm.

The full story

The researchers built a knowledge base from 28 peer-reviewed articles on how low-grade lesions get upgraded. Two language models, DeepSeek-V3 and Qwen3-235b, used it in a zero-shot setting to produce an upgrade probability and a reasoning report for each lesion.

Performance was judged by AUC, calibration plots and decision curve analysis, and compared with a generalized linear mixed model baseline and two experienced endoscopists. DeLong tests found no significant differences from the endoscopists or the baseline (P > 0.05).

At thresholds picked by the maximum Youden index, both models caught about 85% of upgraded lesions. Their negative predictive values were 89.1% and 88.3%. The authors conclude the models may support decisions about endoscopic removal, pending prospective validation.

Sources

1 on file
  1. Peer-reviewed2027-01-07

Talk it over

Three questions to start on, two more if you have time
In the study of 329 patients, 148 lesions (40%) were upgraded after removal. How would you weigh a model with 89.1% negative predictive value against the chance that it misses an upgrade?

Start with a question

AI models matched endoscopists at spotting stomach lesions likely to turn worse, study of 329 patients findshttps://truvace.com/item/retrieval-augmented-generation-models-for-predicting-pathological-upgrading-in-e

  1. In the study of 329 patients, 148 lesions (40%) were upgraded after removal. How would you weigh a model with 89.1% negative predictive value against the chance that it misses an upgrade?

    Answer in the thread
  2. The models were tested on past cases from two tertiary hospitals, with two experienced endoscopists as the comparison. What can a retrospective design like this tell us about how the models would do on new patients?

    Answer in the thread
  3. If you were an endoscopist deciding whether to remove a lesion with a low-grade biopsy result, what would you want to see from a prospective study before letting DeepSeek-V3 or Qwen3-235b shape your choice?

    Answer in the thread
If you have more time
  • The knowledge base came from 28 peer-reviewed articles. What would you want to look up about how those articles were chosen, and how might that change the models' predictions?

    Answer in the thread
  • How might a hospital you know use a probability and reasoning report for each lesion alongside a biopsy result, and who would be responsible for the final decision?

    Answer in the thread

What readers say