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

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 372 gastric LGIN lesions from two tertiary hospitals who underwent ESD following a pre-operative diagnosis of gastric LGIN via biopsy. A RAG knowledge base was constructed based on 28 peer-reviewed articles on LGIN upgrading. Two LLMs (DeepSeek-V3 and Qwen3-235b) were evaluated in a zero-shot setting to generate upgrading probability predictions and reasoning reports for each lesion. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA). The performance of RAG-enhanced LLMs was compared with that of a generalized linear mixed model baseline and two experienced endoscopists. Subgroup analyses were also conducted.

RESULTS

Pathological upgrading to high-grade intraepithelial neoplasia or early gastric cancer was confirmed in 148 (40%) lesions. The AUCs for the RAG-enhanced DeepSeek-V3 and Qwen3-235b were 0.873 [95% confidence interval (95%CI): 0.828-0.901] and 0.862 (95%CI: 0.818-0.893), respectively, demonstrating performance comparable to that of experienced endoscopists and baseline model (DeLong test, P > 0.05). Based on the maximum Youden index, DeepSeek-V3 achieved a sensitivity of 85.1%, specificity of 80.4%, positive predictive value of 74.1%, and negative predictive value of 89.1% at the optimal threshold of 0.357; Qwen3-235b achieved 84.5%, 77.2%, 71.0%, and 88.3%, respectively, at the threshold of 0.453. DCA demonstrated superior clinical utility across threshold probabilities of 15%-80%. In the subgroup analysis, RAG-enhanced LLMs maintained higher accuracy in lesions measuring 1-2 cm.

CONCLUSION

This study evaluated a novel RAG-enhanced framework using LLMs for predicting pathological upgrading in gastric LGIN. These models may have the potential to support clinical decision-making for ESD, pending prospective validation.

Key Words: Large language model; Retrieval-augmented generation; Gastric low-grade intraepithelial neoplasia; Pathological upgrading; Predictive modeling; Endoscopic submucosal dissection; Artificial intelligence; Clinical decision support

Core Tip: This study evaluated a novel retrieval-augmented generation (RAG) -enhanced large language model framework using dual-center data to predict pathological upgrading in gastric low-grade intraepithelial neoplasia. The RAG-enhanced DeepSeek-V3 and Qwen3-235b models achieved areas under the curve of 0.873 [95% confidence interval (95%CI): 0.828-0.901] and 0.862 (95%CI: 0.818-0.893), respectively, demonstrating performance comparable to experienced endoscopists and confirming their potential clinical utility.

INTRODUCTION

Gastric cancer remains a prevalent malignancy globally and is one of the leading causes of cancer-related mortality. The 5-year survival rate is heavily dependent on early detection and intervention[1]. Computer-aided detection systems leveraging deep learning have achieved remarkable success in automated lesion detection in the assessment of gastrointestinal neoplasms[2-4]. However, such techniques predominantly focus on the detection and delineation of suspicious lesions. The clinical management of detected lesions remains challenging, particularly for biopsy-diagnosed gastric low-grade intraepithelial neoplasia (LGIN). LGIN is a well-established precancerous lesion within the Correa cascade of gastric carcinogenesis[5]. Clinical management of LGIN is challenging due to the wide variability in its outcomes[6]. The diagnosis of gastric LGIN is primarily based on endoscopic forceps biopsy (EFB). However, EFB has been shown to underestimate the severity of lesions, with 12%-63% of patients initially diagnosed with LGIN subsequently found to have high-grade intraepithelial neoplasia (HGIN) or early gastric cancer (EGC) following endoscopic resection[7-12]. This discordance leads to significant clinical challenges, including the risk of under-treatment or over-surveillance, both of which contribute to substantial economic and psychological burdens[12].

Endoscopic risk factors for pathological upgrading are well-established, including lesion size > 2 cm, depressed morphology, surface erythema or nodularity, ulceration, intestinal metaplasia, and irregular microsurface (MS) and microvascular (MV) patterns observed on magnifying endoscopy[6,12-19]. Several studies have developed traditional statistical models, particularly using multivariate logistic regression, to integrate these risk factors and estimate the likelihood of pathological upgrading in gastric LGIN[6,12,14,20,21]. However, these models have limitations. Multivariate logistic regression models are static, trained on fixed historical datasets, and often assume linear relationships. As a result, these models struggle to capture the complex, non-linear interactions between variables or dynamically integrate new medical literature.

Large language models (LLMs) represent a significant advancement in artificial intelligence (AI), demonstrating exceptional capabilities in medical text comprehension, reasoning, and knowledge synthesis[22,23]. Retrieval-augmented generation (RAG) is an architectural framework that allows the embedding of customized data into LLMs, grounding them in external, verifiable knowledge sources[24]. By retrieving relevant information from a curated database before generating a response, RAG mitigates hallucinations, enhances factual accuracy, and improves interpretability[24]. These advantages make RAG-enhanced LLMs particularly well-suited for clinical prediction tasks, where decisions must be evidence-based, transparent, and capable of dynamically updating in response to evolving literature.

In this study, we constructed a RAG knowledge base based on the studies on LGIN upgrading and employed two open-source LLMs, DeepSeek-V3 and Qwen3-235b. For each lesion, the models dynamically retrieved relevant evidence on endoscopic features, using this information to generate risk predictions for pathological upgrading along with interpretative reports. We hypothesized that the RAG-enhanced LLMs would achieve superior predictive accuracy and would offer promising clinical applicability.

MATERIALS AND METHODS

Study design and patients

This retrospective observational study was approved by the Institutional Review Boards of Shanghai Sixth People's Hospital and Shanghai Tongren Hospital, with a waiver for informed consent. This study was reported in accordance with the transparent reporting of a multivariable model for individual prognosis or diagnosis for LLMs guidelines[25]. The completed checklist is provided in Supplementary material. Patients who underwent endoscopic submucosal dissection (ESD) for gastric LGIN at Shanghai Sixth People's Hospital between July 2020 and March 2025, and at Tongren Hospital between January 2021 and September 2025, were included.

Inclusion criteria: (1) Age ≥ 18 years; (2) Pre-operative EFB diagnosis of gastric LGIN according to the World Health Organization (WHO) 2019 criteria[26]; (3) Availability of high-quality white-light and magnifying endoscopy with narrow-band imaging (ME-NBI) for review; (4) Availability of complete pathological data; and (5) ESD performed within 6 months of the initial biopsy.

Exclusion criteria: A history of gastric or esophageal surgery, coexisting malignant tumors, or poor-quality images (Figure 1).

Figure 1 Flowchart of patient enrollment. EFB: Endoscopic forceps biopsy; LGIN: Low-grade intraepithelial neoplasia; ME-NBI: Magnifying endoscopy with narrow-band imaging; ESD: Endoscopic submucosal dissection.

Data collection and endoscopic feature assessment

Clinical data and endoscopic images were extracted from electronic medical records. Two experienced endoscopists independently reviewed all endoscopic images and recorded a standardized set of features, with blinding to final pathology results. Discrepant assessments were resolved by consensus discussion with a third senior endoscopist. We systematically reviewed and validated well-established risk factors for pathological upgrading, as reported in the literature[6,12-19], evaluating them using both white-light and ME-NBI.

Features assessed using white-light endoscopy included the following: (1) Lesion size (maximum diameter); (2) Location (cardia/fundus, body, angle, antrum/pylorus); (3) Paris classification[27] (0-I, 0-IIa, 0-IIb, 0-IIc, 0-IIa+IIc, 0-IIc+IIa, 0-III); (4) Surface ulceration (defined as a mucosal break with white coating, visible at endoscopy); (5) Surface nodules; (6) Erythema; and (7) Background mucosal atrophy.

Background mucosal atrophy was classified into 3 stages according to the Kimura–Takemoto classification: Mild (absence of atrophy or atrophy limited to the antrum, C0-C1); moderate (atrophy extending along the lesser curvature up to, but not beyond, the cardia, C2-C3); and severe (atrophy extending across the cardia and involving the fundus and/or greater curvature of the body, O1-O3)[28,29].

For magnifying endoscopy, the following features were assessed: (1) MS and MV patterns, which were classified as regular, irregular, or absent according to the vessels plus surface classification system[30,31]; and (2) Light blue crest (LBC) sign[30].

Pathological evaluation and outcome definition

Gastric intestinal metaplasia was defined as the presence of intestinal metaplasia on preoperative biopsy pathology or the detection of the LBC sign by ME-NBI endoscopy. All endoscopic resection specimens were processed according to standard protocols, with pathological evaluation performed by experienced gastrointestinal pathologists. The primary outcome was pathological upgrading, defined as a final diagnosis of HGIN or EGC in the endoscopic resection specimen, as per the WHO 2019 classification[26]. Lesions that remained as LGIN or were downgraded to inflammation were classified as non-upgrading.

Construction of the RAG knowledge base

To ground the LLM predictions in domain-specific evidence, we constructed a comprehensive RAG knowledge base composed of 28 peer-reviewed articles published between 2010 and 2025. Four electronic databases (PubMed, EMBASE, Web of Science, and the Cochrane Library) were searched for studies focusing on the risk factors and prediction of LGIN upgrading. The Newcastle-Ottawa Scale[32] was used to assess the quality of studies and only studies scoring ≥ 6 were included. The search strategy and flowchart of study selection are presented in Supplementary Table 1 and Supplementary Figure 1. The complete list of 28 articles included in the RAG knowledge base is shown in Supplementary Table 2.

We used RAGFlow, an open-source RAG engine (v0.22.1, https://ragflow.io/), to construct the knowledge base. The abstract and results sections of each article were converted into dense vector embeddings using the Qwen/text-embedding-v4 model. These embeddings were indexed in a vector database for efficient similarity searches.

LLM prediction framework

Two open-source LLMs were evaluated: DeepSeek-V3 (deepseek-chat, version 0325) and Qwen3-235b (qwen3-235b-a22b, Alibaba Cloud, May 2025). Technical specifications were rigorously standardized to ensure reproducibility. DeepSeek-V3 and Qwen3-235b were used via an application programming interface with temperature = 0.1, top_p = 0.75, and max_tokens = 2048. The RAG pipeline employed Qwen/text-embedding-v4 with a chunk size of 256 tokens, top-k retrieval of 30 chunks, and a cosine similarity threshold of 0.65. A cross-encoder reranker (gte-reranker, Alibaba NLP) was applied with a threshold of 0.75 to optimize evidence relevance. Both LLMs were evaluated in a zero-shot setting with frozen weights and were not trained or fine-tuned on the study cohort. To prevent data leakage, no lesion-level or patient-level data from either center were used for few-shot example curation, retrieval optimization, hyperparameter tuning, or prompt engineering before inference. Consequently, all 372 lesions served as an independent evaluation cohort.

For each lesion, a structured prompt was created containing the patient's anonymized endoscopic features. First, the system queried the vector database using key terms derived from the case (e.g., size 2 cm, 0-IIc, ulceration, MV irregular). The top 30 most semantically relevant text chunks from the literature were retrieved. Second, the LLM was instructed with a structured prompt that included: (1) System instructions: “You are an expert in digestive endoscopy evaluating a gastric lesion that was biopsied as low-grade intraepithelial neoplasia (LGIN) before resection. Pathological upgrading is defined as post-resection histology showing high-grade intraepithelial neoplasia (HGIN) or early gastric cancer”; (2) Lesion feature summary; (3) Retrieved evidence with citations; and (4) Task: "Based on the endoscopic features below—and your knowledge of the literature—please estimate the probability of pathological upgrade. Please respond in English". For details, the full prompt template, and representative input-output examples are included in Supplementary material.

Probability prediction

The LLM's response was used to extract the predicted probability and the reasoning text. The predicted probabilities are model confidence estimates; in this study, these were treated as proxies for frequentist probabilities for clinical prediction modeling. For each lesion, the LLMs generated a probability interval, with the median of this interval taken as the predicted probability. Each lesion was processed through three independent application programming interface calls. The mean of three runs was used for primary analysis.

Validation by experienced endoscopists

To compare the diagnostic accuracy of the AI model against human experts, two additional experienced endoscopists (each with over 10 years of clinical experience in EGC diagnosis) blindly reviewed all lesions to assess the possibilities of pathological upgrading based on their clinical expertise. Corresponding confidence scores were documented.

Statistical analysis

Continuous variables are expressed as mean ± SD and compared using the t-test. Categorical variables are presented as n (%) and analyzed using Pearson's χ² test. Interobserver agreement of endoscopic features between the two experienced endoscopists was assessed using Cohen's kappa. LLMs generate confidence estimates rather than true frequentist probabilities. Platt scaling[33,34] was applied as a post hoc calibration analysis. Logistic regression was constructed with LLM-derived predictive probabilities as the sole predictor of observed pathological upgrading. A generalized linear mixed (GLM) model with a random intercept for patient identity was constructed using the endoscopic variables as a baseline comparator.

Discriminatory performance was evaluated using the area under the receiver operating characteristic curve (AUC). To account for within-patient clustering (37 patients with 2 lesions and 3 patients with 3 Lesions), the AUCs and their 95% confidence intervals (95%CIs) were calculated using patient-level cluster bootstrap resampling (1000 iterations, bias-corrected and accelerated method). As a robustness check, we performed a sensitivity analysis using only one randomly selected lesion per patient (n = 329). DeLong tests were used to compare the receiver operating characteristic curves of the RAG-enhanced LLMs against the GLM model and two experienced endoscopists. The optimal probability thresholds for each model were determined using Youden’s index. Calibration was assessed using calibration plots generated with 1000 bootstrap resamples.

Clinical utility was quantified by decision curve analysis (DCA), which calculates the net benefit across a range of threshold probabilities. Subgroup analyses were conducted based on lesion size and post-resection histological findings. Statistical analyses were performed using R software version 4.5.2 (R Core Team, Vienna, Austria, http://www.r-project.org/). A P value < 0.05 was considered statistically significant. Post hoc power analysis indicated that the study had > 80% power (α = 0.05) to confirm that both LLMs achieved discriminative performance significantly exceeding an AUC threshold of 0.80, given 148 events and 224 non-events. All statistical methods and analyses were reviewed and validated by a biomedical statistician.

RESULTS

Patient characteristics

A total of 329 patients (mean age 67.0 ± 9.0 years; 56% male) with 372 gastric LGIN lesions were included in the study. Pathological upgrading to HGIN or EGC was observed in 148 (40%) lesions. Compared to the non-upgraded group, the upgraded lesions were significantly larger and more frequently exhibited depressed morphology, background atrophy, surface ulceration/erythema/nodularity, and irregular MV/MS patterns (all P < 0.05; Table 1). Endoscopic features demonstrated substantial interobserver agreement between the two reviewers (Supplementary Table 3). These findings are consistent with prior studies that have identified these features as independent risk factors[6,12,20,35].

Table 1 Baseline characteristics of patients with positive and negative pathologic upgrade, n (%).

Group		Positive (n = 148)	Negative (n = 224)	P value
General characteristics	Age (year)	67.1 ± 9.3	66.9 ± 8.8	0.94
	Sex (male/female)	90/58	118/106	0.12
White light endoscopy characteristics	Lesion size (cm)	1.7 (0.8)	1.1 (0.4)	< 0.001
	Location			0.032
	Antrum	72 (49)	143 (64)
	Angularis	31 (21)	35 (16)
	Gastric body	38 (26)	37 (17)
	Cardia and others	7 (4.7)	9 (4.0)
	Paris classification			< 0.001
	0-IIa	41 (28)	111 (50)
	0-IIa + IIc	23 (16)	24 (11)
	0-IIb	37 (25)	62 (28)
	0-IIc	41 (28)	25 (11)
	0-IIc + IIa	6 (4.1)	2 (0.9)
	Background atrophy			0.003
	Mild	12 (8.1)	48 (21)
	Moderate	112 (76)	148 (66)
	Severe	24 (16)	28 (13)
	Surface nodularity			< 0.001
	Present	52 (35)	19 (8.5)
	Absent	96 (65)	205 (92)
	Surface ulceration			< 0.001
	Present	34 (23)	11 (4.9)
	Absent	114 (77)	213 (95)
	Erythema			< 0.001
	Present	96 (65)	60 (27)
	Absent	52 (35)	164 (73)
ME characteristics	MS pattern			< 0.001
	Irregular/absent	130 (88)	120 (54)
	Regular	18 (12)	104 (46)
	MV pattern			< 0.001
	Irregular/absent	119 (80)	87 (39)
	Regular	29 (20)	137 (61)
Biopsy pathology	Intestinal metaplasia			0.75
	Present	133 (90)	199 (89)
	Absent	15 (10)	25 (11)

ME: Magnifying endoscopy; MV: Microvascular; MS: Microsurface.

Discriminative performance

Based on the maximum Youden index, the optimal threshold for DeepSeek-V3 was 0.357, yielding a sensitivity of 85.1%, specificity of 80.4%, positive predictive value (PPV) of 74.1%, and negative predictive value (NPV) of 89.1%. For Qwen3-235b, the optimal threshold was 0.453, with a sensitivity of 84.5%, specificity of 77.2%, PPV of 71.0%, and NPV of 88.3%. The RAG-enhanced LLMs demonstrated discriminative performance with AUCs of 0.873 (95%CI: 0.828-0.901) for DeepSeek-V3 and 0.862 (95%CI: 0.818-0.893) for Qwen3-235b (Figure 2). The AUCs for the two experienced endoscopists were 0.856 (95%CI: 0.810-0.888) and 0.851 (95%CI: 0.809-0.884), respectively. The sensitivity analysis using only one lesion per patient (n = 329) showed minimal AUC changes (Δ < 0.01; DeepSeek-V3: 0.873 vs 0.880; Qwen3-235b: 0.862 vs 0.863). The GLM model achieved an AUC of 0.879 (95%CI: 0.840-0.917). Neither DeepSeek-V3 (DeLong test, P = 0.60) nor Qwen3-235b (DeLong test, P = 0.20) differed significantly from the GLM model. In addition, no significant difference was found between DeepSeek-V3 and endoscopist 1 (P = 0.11) or endoscopist 2 (P = 0.18). Qwen3-235b showed no significant difference compared with the two endoscopists (P = 0.60 and P = 0.55, respectively).

Figure 2 Model discriminative performance. Receiver operating characteristic curves for the retrieval-augmented generation-enhanced large language models and experienced endoscopists in predicting pathological upgrading in gastric low-grade intraepithelial neoplasia. AUC: Area under the receiver operating characteristic curve; 95%CI: 95% confidence interval; GLM: Generalized linear mixed.

Calibration and clinical utility

Calibration plots revealed good predictive accuracy between the actual and predicted probabilities (Figure 3). DCA demonstrated superior clinical utility of the RAG-enhanced LLMs across threshold probabilities of 15%-80%, with higher net benefit compared to treat-all/none strategies (Figure 4).

Figure 3 Model calibration curves. The dotted line (apparent) represents the performance of the retrieval-augmented generation-enhanced large language models, whereas the solid line (bias-corrected) accounts for and adjusts potential biases. The diagonal reference line (ideal) indicates perfect calibration, where the predictions match the observed outcomes exactly. GLM: Generalized linear mixed.

Figure 4 Decision curve analysis on the prediction models. The net benefit of using the retrieval-augmented generation-enhanced large language models across various threshold probabilities is compared to the strategies of treating all patients or following up all patients. GLM: Generalized linear mixed.

Subgroup analysis

In lesions measuring 1-2 cm (n = 229), where clinical decision-making is most ambiguous, both DeepSeek-V3 and Qwen3-235b maintained high accuracy. However, performance declined in the subsets of lesions ≥ 2 cm and < 1 cm (Table 2). In the subgroup analysis based on post-resection histology, both DeepSeek and Qwen maintained high discriminative accuracy, with peak performance observed for EGC (Table 2).

Table 2 Subgroup analysis for predicting pathological upgrading.

Stratification factors	Subgroups	Upgraded/non-upgraded (n)	AUC
			DeepSeek-V3	Qwen3-235b
Lesion size (cm)	< 1	19/94	0.790 (0.589-0.896)	0.785 (0.617-0.889)
	1-2	103/126	0.852 (0.793-0.892)	0.837 (0.776-0.878)
	≥ 2	26/4	0.594 (0.473-0.721)	0.605 (0.305-0.765)
Postoperative pathology	HGIN	105/224	0.851 (0.795-0.886)	0.846 (0.789-0.885)
	EGC	43/224	0.924 (0.875-0.952)	0.900 (0.849-0.934)

AUC: Area under the receiver operating characteristic curve; HGIN: High-grade intraepithelial neoplasia; EGC: Early gastric cancer.

Interpretability and reasoning analysis

For each lesion, the RAG-enhanced LLMs provided the principal risk and protective factors for pathological upgrading and synthesized an evidence-driven probability estimate. For clinical convenience, the models have been publicly deployed at http://xiyoucloud.pro:12599/RAG-Enhanced_LLMs_for_Gastric_LGIN_Risk_Prediction (Supplementary Figure 2).

DISCUSSION

This study evaluated a novel, evidence-based predictive framework for forecasting pathological upgrading in gastric LGIN by integrating LLMs with a RAG architecture. Our results demonstrate that the RAG-enhanced LLMs (both DeepSeek-V3 and Qwen3-235b) exhibit clinically reliable predictive accuracy, with performance comparable to that of experienced endoscopists and the baseline model (DeLong test, P > 0.05 for all comparisons). The proposed models offer several advantages: They synergistically integrate structured endoscopic data with unstructured medical knowledge, enhance interpretability, and demonstrate good clinical utility, as evidenced by DCA.

In recent years, numerous studies have sought to improve the diagnostic accuracy of LGIN, with the ultimate goal of refining the prognostic prediction for gastric LGIN[6,12,14,18,20,21,36]. However, the findings from these studies have yet to yield a unified consensus, and a relatively objective and quantifiable predictive model remains elusive. Consequently, the accurate prediction of clinical outcomes for gastric LGIN remains a significant clinical challenge. Given the rapid advancements in AI, we hypothesize that LLMs could be a powerful tool to address this gap. In our study, the RAG-enhanced LLMs achieved robust discriminative performance, and the high net benefit demonstrated by DCA suggests that such a tool holds promising clinical applicability. In the subgroup analysis, the RAG-enhanced LLMs maintained higher accuracy for lesions sized 1-2 cm, where clinical decision-making is most ambiguous. In contrast, for lesions ≥ 2 cm, predictive performance was suboptimal. This underperformance is likely attributable to the limited sample size of lesions ≥ 2 cm. Accordingly, our RAG-enhanced LLMs are not recommended for clinical application in this specific subgroup. In the subgroup of post-resection EGC, discrimination was notably higher, likely due to the more conspicuous endoscopic dysplastic characteristics of EGC lesions.

The superiority of LLMs lies in their capacity to model complex, non-linear interactions between variables, which may be overlooked by linear models such as logistic regression[37]. The LLMs were able to capture intricate interactions between risk factors for pathological upgrading, such as the compounded risk of ulceration in lesions with background intestinal metaplasia, a relationship that is difficult to capture using linear models. In the present study, while RAG-enhanced LLMs did not significantly outperform the conventional GLM model in this retrospective cohort, the latter was fitted and internally validated on the same dataset, representing its theoretical performance ceiling. Conventional statistical models are prone to performance decay upon external validation[38]. In contrast, the RAG-enhanced LLMs were evaluated under a zero-shot setting with no within-cohort training or fine-tuning, suggesting their performance may generalize more stably to independent cohorts. Thus, despite similar AUC values observed in this retrospective evaluation, the LLM framework may be a more robust solution for real-world clinical implementation. The RAG framework also enables the model to be grounded in the latest clinical evidence for each prediction, a feature absent in static statistical models that are trained on a fixed dataset from the past[6,12,14,20]. For instance, probe-based confocal laser endomicroscopy (pCLE)[39], a high-resolution imaging technique, may become a critical predictor of pathological upgrading in gastric LGIN. By incorporating the latest evidence from pCLE studies into the RAG knowledge base, RAG-enhanced LLMs can recalibrate upgrade probabilities in real-time, based on the microscopic mucosal features observed. Additionally, the generation of natural-language reasoning represents a significant advancement in clinical AI, shifting from opaque outputs to transparent, clinician-friendly explanations. This capability facilitates clinicians' understanding of why a specific risk level is assigned.

There are several limitations to our study. First, the RAG knowledge base may be subject to publication bias, particularly toward positive findings. Incorporating negative studies and pre-registered trial data could help mitigate this issue. Second, our retrospective study was restricted to patients who had already undergone ESD, thereby introducing spectrum bias. Accordingly, our findings predominantly reflect the ESD-eligible subgroup instead of all biopsy-diagnosed LGIN lesions. Future prospective studies should include patients under surveillance to verify the generalizability of our model. Third, the LLMs generate confidence estimates rather than true frequentist probabilities, requiring cautious interpretation. Fourth, the ≥ 2 cm subgroup was small, yielding wide confidence intervals and suboptimal model performance. Prospective multicenter validation is required before clinical deployment.

CONCLUSION

This study evaluated a novel RAG-enhanced framework using LLMs for predicting pathological upgrading in gastric LGIN. These models provide probabilistic predictions with interpretable, evidence-based reasoning that may have the potential to support clinical decision-making for ESD, pending prospective real-world validation.

ACKNOWLEDGEMENTS

We thank Dr. Li Shen (Clinical Research Center, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine) for statistical assistance.