Introduction

Gastric cancer (GC) remains one of the leading causes of cancer-related deaths globally, with a particularly high incidence in East Asia.¹ Early GC (EGC), which is confined to the mucosa or submucosa, offers a significantly better prognosis if detected and treated appropriately.² However, the detection of lymph node metastasis (LNM), which occurs in approximately 20%–30% of the patients with EGC, is crucial for determining the appropriate therapeutic approach, including the need for extended lymphadenectomy or more conservative treatment.³ Therefore, accurate identification of LNM in the early stages of GC can help avoid over- or undertreatment, both of which are associated with poor prognosis.

Traditional diagnostic methods for assessing LNM in EGC, including computed tomography (CT), magnetic resonance imaging (MRI), and endoscopic ultrasound (EUS), have been widely utilized. While these techniques provide valuable imaging data, their sensitivity and accuracy for detecting small or microscopic metastatic LNs, especially in EGC, remain limited.⁴ CT and MRI are often incapable of detecting small LNs that may harbor microscopic metastasis, leading to either unnecessary surgeries or missed opportunities for more aggressive treatment.⁵ Similarly, EUS, although beneficial in assessing larger LNs, is less effective in detecting micrometastasis, and its accuracy heavily depends on the operator’s skill and experience.⁶ These limitations underscore an urgent need for more effective and precise diagnostic methods to assess LNM in EGC.

In recent years, artificial intelligence (AI), particularly deep learning (DL) and machine learning (ML), has emerged as a promising tool for revolutionizing cancer diagnostics.^7,8 AI can analyze vast amounts of data from various sources, including medical images, histopathological slides, and clinical data, to identify complex patterns that may not be immediately apparent to the human eye.⁹ This capability positions AI as an ideal technology for improving the accuracy of LNM detection in EGC.^10,11 Numerous studies have explored the application of AI in detecting LNM in GC, using a variety of imaging modalities, including CT, MRI, and EUS. For example, a recent study demonstrated that DL models could significantly improve the accuracy of histopathological diagnosis of EGC, achieving an overall area under the curve (AUC) of 0.75 in predicting LNM status.¹² Additionally, AI models trained on endoscopic images have shown promise in enhancing the detection of LNM, offering higher diagnostic performance than traditional methods.¹⁰ Despite these advancements, the reported performance of AI-based LNM prediction models remains heterogeneous, with considerable variability in study design, sample size, model architecture, input features, and validation strategies. Furthermore, the generalizability and clinical utility of these algorithms have not been systematically evaluated.

Aim

This systematic review and meta-analysis aimed to provide a comprehensive evaluation of AI applications for LNM detection in EGC. By reviewing the methodologies, performance metrics, and limitations of AI models used in current studies, we aimed to identify key trends, challenges, and gaps in the field. Furthermore, we discussed potential implications of AI for clinical practice and proposed future research directions to overcome existing barriers and improve the integration of AI into diagnostic workflows. Beyond diagnostic accuracy, accurate preoperative prediction of LNM is fundamental to minimally-invasive treatment planning in EGC. The decision to pursue endoscopic resection or gastrectomy with lymphadenectomy relies heavily on nodal status assessment, as inappropriate patient selection may result in noncurative endoscopic resection or unnecessary surgical morbidity. Therefore, AI-based prediction models should be evaluated not only as diagnostic tools, but also as decision-supporting instruments within minimally-invasive and endoscopic surgical workflows.

Materials and methods

This study was designed as a systematic review and meta-analysis to evaluate the diagnostic performance of AI-based models for predicting LNM in patients with EGC. The review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies guidelines.¹³ The study protocol was prospectively registered in the Open Science Framework (https://osf.io/grf8q/). Since all data were obtained from previously published studies, ethical approval and patient informed consent were not required.

Search strategy

To ensure a comprehensive review of the literature, a systematic search was conducted across 3 electronic databases: PubMed, Embase, and Web of Science. The search was performed from the inception of each database until August 2025 without any language limitation. The search terms included combinations of the following key words: “artificial intelligence,” “machine learning,” “deep learning,” “gastric cancer,” “early gastric cancer,” “lymph node metastasis,” “diagnosis,” and “detection.” Boolean operators (AND, OR) were used to refine the search and maximize retrieval of relevant studies. Additionally, manual searches were conducted through the reference lists of included articles to identify additional studies that may have been overlooked during the database search.

Inclusion and exclusion criteria

To be included in the analysis, the studies had to meet the following criteria: 1) study design: retrospective or prospective diagnostic accuracy studies; case reports, reviews, meta-analyses, and opinion pieces were excluded; 2) population: studies involving adult patients with EGC, including those with and without LNM; 3) index test: AI-based models, including ML, DL, or hybrid approaches, designed to predict the presence of LNM; and 4) outcome measures: studies that reported performance metrics, such as sensitivity, specificity, accuracy, AUC, or other relevant diagnostic performance metrics for AI-based methods in detecting LNM.

Studies were excluded if they focused on late-stage GC, AI was not used as the primary diagnostic tool for LNM detection, or the full text of the article was not available.

Data extraction

Two independent reviewers (CXP and YJP) performed the data extraction using a predesigned, standardized form. Discrepancies between them were resolved through consensus or with the assistance of a third reviewer (YGZ). The following data were extracted from each study: 1) study characteristics: first author, publication year, country, study design, sample size, patient demographics; 2) AI model details: algorithm type (eg, convolutional neural network, support vector machine), input features (clinical, pathological, imaging, or multimodal), and validation strategy; 3) validation cohorts: type of validation (internal vs external) and whether physician assessment was included as a comparator; and 4) diagnostic performance: true positives (TPs), false positives (FPs), false negatives (FNs), true negatives (TNs), sensitivity, specificity, and AUC values.

Since most of the included studies did not explicitly report diagnostic contingency Tables, we reconstructed the 2 × 2 Tables using 2 complementary approaches. First, when sufficient data were available, we derived the numbers of TPs, FPs, FNs, and TNs based on the reported sensitivity, specificity, total sample size, and the number of cases confirmed as positive according to the reference standard. Second, in the studies lacking sufficient numerical details, we estimated these values by extracting the optimal sensitivity and specificity from the receiver operating characteristic (ROC) curve using the Youden index. It is important to acknowledge that the latter approach may introduce potential bias, as the cutoff point determined by the ROC curve may not precisely mirror clinical practice. This discrepancy could result in case misclassification and, consequently, influence the calculated contingency Table values.

Quality assessment

The methodological quality of the included studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool.¹⁴ Four domains were assessed: patient selection, index test, reference standard, and analysis. Each domain was rated as having a “low,” “high,” or “unclear” risk of bias, and concerns regarding applicability were similarly evaluated. The assessments were performed independently by 2 reviewers (CXP and YJP), with disagreements resolved by consensus.

Statistical analysis

A bivariate random-effects model was employed to synthesize diagnostic accuracy estimates across the studies and evaluate the performance of AI-based models in predicting LNM in EGC. The pooled sensitivity, specificity, and corresponding 95% CIs were calculated separately for the internal validation datasets, external validation datasets, and physician assessments (endoscopic, radiological, or pathological), where applicable. Forest plots were constructed to visually summarize the pooled sensitivity and specificity estimates, while summary ROC (SROC) curves were generated to depict the overall diagnostic performance and provide combined AUC estimates with their 95% CIs and prediction intervals. Between-study heterogeneity was assessed using the Higgins I² statistic, with values of 25%, 50%, and 75% representing low, moderate, and high heterogeneity, respectively. In the instances where substantial heterogeneity was observed (I² >50%) and the number of included datasets exceeded 10, meta-regression analyses were performed to explore potential sources of heterogeneity. Covariates considered in the meta-regression included AI model type (eg, ML vs DL), validation strategy (internal vs external), type of input features (clinical, imaging, or multimodal), and study design (retrospective cohort vs case control). Additionally, univariate subgroup analyses were conducted to further examine their potential effects on diagnostic accuracy in the internal validation.

Potential publication bias was evaluated using the Deeks funnel plot asymmetry test, with a P value below 0.05 indicating significant small-study effects. Moreover, Fagan nomograms were constructed to estimate post-test probabilities and assess the clinical utility of AI-based models across different pretest probability scenarios.

All statistical analyses were conducted using the Midas and Metadata packages in Stata, version 14.0 (StataCorp, College Station, Texas, United States). All statistical tests were 2-sided, and a P value below 0.05 was considered significant. The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Results

Study selection

A total of 502 records were identified during the search phase. After eliminating 84 duplicates, 418 unique articles remained for screening. Upon careful evaluation of the titles and abstracts, 380 articles were excluded as not meeting the predefined criteria. Subsequent in-depth scrutiny of full texts led to the inclusion of 18 articles for analysis.^10-12,15-29 The included studies were conducted primarily in China, Korea, and Japan, spanning from 2021 to 2025. These investigations centered on the application of AI in predicting LNM in EGC. Various AI methodologies, encompassing ML and DL models, were utilized across the studies. The detailed literature screening process is presented in Figure 1.

Characteristics of included studies

A total of 18 studies involving 41 505 patients were included, with all studies adopting retrospective or prospective cohort designs. All included studies utilized datasets from East Asian populations, reflecting the higher incidence of EGC in these regions. Input features for the AI models varied, including clinical characteristics, imaging data, histopathological features, and genomic profiles. ML techniques, such as random forest, gradient boosting machines, and support vector machines, were commonly used, alongside DL methods, such as convolutional neural networks. Sample sizes ranged widely, with training datasets containing between 20 and 10 332 cases. Many studies lacked external validation, limiting their generalizability, while others employed robust validation strategies, achieving higher levels of reliability. Reported performance metrics, such as sensitivity, specificity, accuracy, and AUC, highlighted the potential of AI to enhance LNM detection. The detailed characteristics of the included studies are outlined in Table 1.

Table 1. Characteristics of the included studies on using artificial intelligence to detect lymph node metastasis in patients with early gastric cancer

Author, year	Country	Study type	Study design	Patient inclusion period	GC staging	Algorithm used in Al models	Variables used in Al models	Training set, n	Test set, n	Validation set, n
Wang et al, 202115	China	Single-center	Retrospective cohort	2012–2017	T1–T2	mRMR	Station 3 lymph nodes and primary tumor radiomics	80	79	NA
Zhu et al, 202220	China	Multicenter	Retrospective cohort	NA	T1	GBM, XGBoost, RF, DT, and NNET	Clinical features	1878	470	NA
Tian et al, 202214	China	Multicenter	Case control study	2010–2015	T1a–T1b	GLM, RPART, RF, GBM, SVM, RDA, and NNET	Clinical features	1839	458	227
Na et al, 202216	Korea	Single-center	Retrospective cohort	2005–2021	T1a	LR, SVM, and RF	Clinical features	10 332	NA	4428
Zeng et al, 202219	China	Single-center	Retrospective cohort	2016–2021	T1a–T1b	Pretrained deep learning networks	Deep transfer learning, radiomics, and clinical features	388	167	79
Yang et al, 202218	China	Multicenter	Retrospective cohort	2012–2021	T1a–T1b	Linear SVC, LR, XGBoost, LightGBM, and Gaussian process classification model	Clinical features	305	NA	35
Wei et al, 202217	China	Multicenter	Retrospective cohort	2015–2021	EGC	RFC, DT, SVM, XGBoost, GLM, and ANN	MRI parameters	368	158	NA
Lee et al, 202323	Korea	Single-center	Retrospective cohort	2012–2020	EGC	GBM and LR	Clinical features	2044	512	548
Hayashi et al, 202322	Japan	Single-center	Prospective cohort	2013–2018	T1b	XGBoost	Clinical and pathological variables	382	NA	140
Dong et al, 202321	China	Single-center	Retrospective cohort	2017–2022	T1–T2	10-lncRNA risk-prediction model	Genome-wide expression profiles of lncRNA	20	98	127
Seo et al, 202425	Korea	Multicenter	Retrospective cohort	2007–2017	T1b	Logistic regression, RF, XGBoost, and SVM	Clinical and pathological variables	2426	NA	1042
Kato et al, 202411	Japan	Multicenter	Retrospective cohort	2010–2021	EGC	Neural network	Clinical and pathological variables	3506	NA	536
Yang et al, 202410	China	Single-center	Retrospective cohort	2016–2023	EGC	CNN	Endoscopic images	54	24	30
Lee et al, 202424	Korea	Multicenter	Retrospective cohort	2018–2023	EGC	CNN	Endoscopic images and videos	4336 images and 153 videos	260 images and 10 videos	436 images and 89 videos
Sung et al, 202412	Korea	Multicenter	Retrospective cohort	NA	EGC	DeepLabV3+ and XGBoost	Hematoxylin and eosin–stained images	NA	NA	NA
Kang et al, 202528	Korea	Multicenter	Retrospective cohort	2010–2015	EGC	CNN and CNN with RF	Endoscopic images, demographic data, biopsy pathology, CT findings	2927	449	766
He et al, 202527	China	Multicenter	Retrospective cohort	2006–2019	pT1N0	2.5D MIL-based model	Preoperative portal venous phase CT images	1953	NA	1211
Gao et al 202526	China	Single-center	Retrospective cohort	NA	T1	VGG16, ResNet34, MobileNetV2, and PVTv2	Morphological features of collagen fibers from multiphoton microscopy	143	69	NA
Abbreviations: AI, artificial intelligence; ANN, artificial neural network; CNN, convolutional neural network; CT, computed tomography; DT, decision tree; EGC, early gastric cancer; FCNN, fully convolutional neural network; GBM, gradient boosting machine; GC, gastric cancer; GLM, generalized linear model; LR, logistic regression; MIL, multiple instance learning; MRI, magnetic resonance imaging; mRMR, minimum redundancy maximum relevance; NA, not applicable; NNET, neural network; PVT, pyramid vision transformer; RDA, regularized dual averaging; RF, random forest; RFC, random forest classifier; RPART, recursive partitioning and regression tree; SVC, support vector classifier; SVM, support vector machine; VGG, visual geometry group; XGBoost, extreme gradient boosting

Quality assessment

The methodological quality of the included studies, evaluated using the QUADAS-2 tool, is summarized in Table 2. Overall, the risk of bias was deemed low to moderate across most domains. However, patient selection bias was frequently rated as high (13/17 studies), largely due to retrospective designs and nonconsecutive enrollment. Index test bias was low in most cases, reflecting consistent application of AI algorithms. Concerns regarding applicability were generally low across all domains. These findings suggest that while the methodological rigor was acceptable, future studies should prioritize prospective designs and standardized validation protocols to further reduce bias. The detailed risk bias of the included studies is illustrated in Table 2.

Table 2. Quality Assessment of Diagnostic Accuracy Studies-2 evaluation of the risk of bias

Author, year	Risk of bias				Applicability concerns
	Patient selection	Index test	Reference standard	Analysis	Patient selection	Index test	Reference standard
Wang et al, 202115	High	Low	Unclear	Unclear	Low	Low	Low
Zhu et al, 202220	Unclear	Low	Unclear	Unclear	Low	Low	Low
Tian et al, 202214	High	High	Low	High	Low	Low	Low
Na et al, 202216	High	High	Low	Low	Unclear	Low	Low
Zeng et al, 202219	High	High	Low	Low	Low	Low	Low
Yang et al, 202218	High	High	Unclear	Unclear	Low	Low	Low
Wei et al, 202217	High	High	Low	High	Low	Low	Low
Lee et al, 202323	High	High	Low	Low	Low	Low	Low
Hayashi et al, 202322	High	High	Unclear	Unclear	Low	Low	Low
Dong et al, 202321	High	Unclear	Low	Low	Low	Low	Low
Seo et al, 202425	High	Unclear	Low	Low	Low	Low	Low
Kato et al, 202411	High	Unclear	Low	Low	Low	Low	Low
Yang et al, 202410	High	High	Low	High	Low	Low	Low
Lee et al, 202424	Low	Low	Low	Low	Low	Low	Low
Sung et al, 202412	High	Low	Low	Unclear	Low	Low	Low
Kang et al, 202528	High	Low	Low	High	Low	Low	Low
He et al, 202527	High	Low	Low	High	Low	Low	Low
Gao et al, 202526	High	Low	Low	High	Low	Low	Low

Diagnostic performance of artificial intelligence models Performance in internal validation cohorts

In 15 studies reporting internal validation results, AI-based models demonstrated robust diagnostic accuracy for predicting LNM in EGC.^{10-12,16,18-21,23-29} The pooled sensitivity was 0.81 (95% CI, 0.62–0.92), and the specificity was 0.82 (95% CI, 0.66–0.91; Figure 2). The corresponding summary area under the SROC curve was 0.88 (95% CI, 0.85–0.91; Figure 3), indicating strong discriminative capacity. Applying these values to a pretest probability of 20% increased the post-test probability to 53% following a positive result, and decreased it to 6% after a negative result (Figure 4). These results indicate that AI models can substantially refine clinical risk stratification beyond baseline estimates. Nevertheless, significant heterogeneity was observed (I² = 95.23% for sensitivity; I² = 99.35% for specificity), reflecting variability in algorithm type, input features, and reference standards.

Performance in external validation cohorts

External validation, available in 10 studies, further supported the generalizability and robustness of AI-based approaches.^{15-17,10,20,23-25,28,29} Pooled sensitivity and specificity were 0.81 (95% CI, 0.52–0.94) and 0.84 (95% CI, 0.67–0.94), respectively (Figure 5). The summary AUC was 0.9 (95% CI, 0.87–0.92; Figure 6), slightly outperforming the internal validation results and confirming consistent diagnostic power across independent patient populations. The positive likelihood ratio and negative likelihood ratio were 5 and 0.23, respectively, translating into a post-test probability of 56% after a positive test, and 6% after a negative test (Figure 7). These metrics underscore the clinical applicability of AI algorithms in diverse clinical settings and patient cohorts.

Comparison with clinician performance

Five studies directly compared the diagnostic performance of AI models with that of experienced clinicians, including endoscopists, radiologists, and pathologists.^{17,18,22,24,25} AI algorithms consistently outperformed human experts, particularly in sensitivity. The pooled sensitivity and specificity for clinician assessments were 0.68 (95% CI, 0.32–0.91) and 0.8 (95% CI, 0.65–0.9), respectively (Figure 8), with the AUC of 0.82 (95% CI, 0.79–0.85; Figure 9). Similarly, the post-test probability after a positive clinician assessment increased to only 46%, as compared with 53%–56% for AI predictions (Figure 10).

Subgroup analyses and meta-regression in internal validation cohorts

To explore potential sources of heterogeneity, subgroup analyses were conducted based on the type of study (single-center or multicenter), study design, and variable type (Table 3). Stratification by type of study suggested that AI models developed in single-center studies achieved slightly higher sensitivity (0.9; 95% CI, 0.75–0.96) than those from multicenter studies (0.69; 95% CI, 0.42–0.87; P = 0.33). Similarly, case-control studies demonstrated higher sensitivity (0.99; 95% CI, 0.38–1) than retrospective cohorts (0.72; 95% CI, 0.55–0.84; P = 0.12). Stratification by input variables indicated that models incorporating clinical and pathological variables showed slightly higher sensitivity (0.84; 95% CI, 0.54–0.96) than those using endoscopic and imaging features (0.78; 95% CI, 0.66–0.87). These results suggest that none of these factors may contribute to heterogeneity, although none of the subgroup differences were significant. Subgroup analyses by type of study, study design, and variable type showed no significant differences. Notably, within most subgroups, residual heterogeneity remained significant (P <⁠0.05), indicating that the examined factors did not fully account for variability across the studies. Furthermore, meta-regression analyses indicated that none of the tested covariates explained the substantial heterogeneity. Specifically, AI model type (P = 0.54), validation strategy (P = 0.32), type of input features (P = 0.27), and study design (P = 0.25) were not significant moderators.

Table 3. Subgroup analyses results based on internal validation data

Variable		Sensitivity			Specificity
		Summarized results	95% CI	I2	Summarized results	95% CI	I2
Overall pooled effect (n = 15)		0.18	0.62–0.92	95.23	0.82	0.66–0.91	99.35
Type of study	Single-center (n = 6)	0.9	0.75–0.96	95.12	0.71	0.39–0.9	99.25
	Multicenter (n = 9)	0.69	0.42–0.87	94.82	0.87	0.71–0.95	99.42
Study design	Retrospective cohort (n = 11)	0.72	0.55–0.84	94.22	0.87	0.76–0.93	97.72
	Case control study (n = 4)	0.99	0.38–1	97.29	0.62	0.15–0.94	99.74
Variables used in artificial intelligence models	Clinical and pathological variables (n = 10)	0.84	0.54–0.96	95.86	0.79	0.54–0.93	99.51
	Endoscopic and imaging variables (n = 5)	0.78	0.66–0.87	84.43	0.85	0.75–0.92	94.69

Publication bias

Potential publication bias was evaluated using the Deeks funnel plot asymmetry test for the included diagnostic accuracy studies. The test indicated no evidence of publication bias for the internal validation analysis (P = 0.38; Figure 11), external validation analysis (P = 0.35), or clinician comparison studies (P = 0.48; Figures 12 and 13). The symmetrical distribution of effect sizes further supports the reliability of the pooled estimates.

Discussion

Principal findings

This systematic review and meta-analysis comprehensively evaluated the diagnostic performance of AI-based models for predicting LNM in EGC. By synthesizing evidence from 18 studies involving 41 505 patients, we demonstrated that AI algorithms achieved high diagnostic accuracy for LNM detection. In internal validation cohorts, the pooled sensitivity and specificity were 0.81 and 0.82, respectively, with the AUC of 0.88. Importantly, similar performance was maintained in external validation datasets (sensitivity, 0.81; specificity, 0.84; AUC, 0.9), suggesting promising generalizability. Moreover, AI models consistently outperformed experienced clinicians, particularly in sensitivity, highlighting their potential to augment or even surpass human diagnostic capabilities. Overall, these findings suggest that AI-based prediction tools could play a pivotal role in preoperative risk stratification and clinical decision-making for EGC.

Comparison with previous studies

Our findings align with and extend previous literature demonstrating the potential of AI to improve oncological diagnostic accuracy. Conventional imaging modalities, such as CT, MRI, and EUS, have historically shown limited sensitivity for detecting micrometastatic disease, often resulting in FNs and suboptimal treatment decisions. For example, previous meta-analyses reported sensitivities of approximately 60% for CT and EUS in detecting NM in EGC—significantly lower than the pooled estimates observed in our analysis for AI models.^4,30 This suggests that, by integrating multidimensional data—including imaging, clinical, and pathological features—AI can extract latent patterns beyond the capabilities of conventional diagnostic approaches.

Moreover, our results corroborate earlier evidence from individual studies indicating that AI can outperform clinicians in nodal assessment. As compared with experts, the superior sensitivity of AI models observed in this meta-analysis reflects their ability to consistently detect subtle imaging and histopathological cues that may be overlooked due to human cognitive limitations or interobserver variability. Notably, several DL-based models achieved AUCs exceeding 0.9 in the external validation, underscoring their potential utility as reliable tools in routine clinical practice.

Despite the encouraging diagnostic performance, substantial heterogeneity was observed across the included studies, which likely arises from several methodological differences, including variations in study design, data sources, sample size, and feature selection. Our subgroup and meta-regression analyses, however, did not identify any single covariate—such as algorithm type, validation strategy, or input modality—as a significant source of heterogeneity. This suggests that multiple interacting factors contribute to the observed variability, underscoring the complexity of AI model development and evaluation in this context. One noteworthy finding is that models incorporating clinical and pathological variables tended to achieve slightly higher sensitivity than those relying primarily on imaging data. This observation reflects the multifactorial nature of NM, which is influenced not only by tumor morphology but also by biological aggressiveness and host–tumor interactions—features that are often captured in clinicopathological and genomic data. Therefore, future AI approaches may benefit from multimodal data integration, combining radiological, pathological, and molecular features to enhance predictive performance and clinical utility. Additionally, most included studies employed retrospective, single-center designs, and only 11 studies performed external validation. This limited diversity in patient populations and clinical settings may lead to spectrum bias and restrict generalizability of the findings. While the pooled performance in the external validation cohorts was similar to that in the internal datasets, larger multicenter prospective studies are needed to confirm these results across diverse clinical scenarios.

Implications for minimally-invasive and endoscopic surgery

Accurate assessment of LNM risk is a cornerstone of treatment selection in EGC, particularly in the era of minimally-invasive and endoscopic therapies. Patients with a low risk of nodal involvement may safely undergo endoscopic resection, thereby avoiding unnecessary gastrectomy and preserving gastric function, whereas those with a high risk of LNM require surgical resection with appropriate lymphadenectomy to ensure oncological safety. The present meta-analysis demonstrates that AI-based models achieve robust diagnostic accuracy and consistently outperform clinician assessment, highlighting their potential value as preoperative decision-supporting tools in this context.

From a minimally-invasive surgery perspective, AI-based LNM prediction may function as an effective triage instrument to stratify patients before treatment selection. By providing individualized probability estimates of NM, AI models may help identify optimal candidates for endoscopic resection and reduce the risk of noncurative procedures. This is particularly relevant given the clinical challenges of salvage surgery following noncurative endoscopic resection, which is associated with increased operative complexity, higher morbidity, and increased patient burden. Improved preoperative risk stratification using AI could therefore contribute to reducing unnecessary secondary surgery and optimizing initial treatment strategies.

Furthermore, AI-assisted prediction may support surgical planning beyond the binary choice between endoscopic and surgical treatment. For patients proceeding to minimally-invasive gastrectomy, preoperative estimation of LNM risk may inform the extent of LN dissection and facilitate more tailored operative strategies. In this regard, AI models align closely with the objectives of minimally-invasive surgery—minimizing surgical trauma while maintaining adequate oncological clearance.

Importantly, AI-based tools should be viewed as complementary to, rather than replacements for, clinical judgment. Integration of AI predictions into multidisciplinary decision-making processes may enhance diagnostic confidence, reduce interobserver variability, and standardize treatment selection, particularly in settings with variable expertise. As minimally-invasive and endoscopic techniques continue to evolve, AI-driven risk stratification holds promise for refining patient selection and improving the overall quality of surgical care in EGC.

Limitations

Several limitations of this meta-analysis should be acknowledged. First, the substantial heterogeneity across the studies limits the certainty of pooled estimates, despite the efforts to explore potential moderators. Second, the predominance of retrospective designs introduces risks of selection bias and confounding, while nonconsecutive patient enrollment in many studies may further compromise external validity. Third, the lack of standardized reference standards for nodal status—particularly regarding micrometastasis detection—may have contributed to variability in model performance and outcome definitions. Fourth, most models were developed and validated in East Asian populations, raising questions about their applicability to Western cohorts with differing disease epidemiology, histopathological features, and treatment strategies. Fifth, while AI models demonstrated high accuracy, few studies reported calibration metrics or clinical utility analyses (eg, decision-curve analysis), which are critical for translating diagnostic performance into meaningful clinical outcomes. Finally, the absence of an open-source model code and publicly available datasets in most studies limits reproducibility and hinders external validation by independent researchers.

Future directions

To fully harness the potential of AI in EGC management, several research priorities must be addressed. Prospective, multicenter studies with standardized patient inclusion criteria and outcome definitions are essential to validate model performance and enhance generalizability. Future efforts should also focus on integrating multimodal data—including radiomics, histopathomics, genomics, and proteomics—to capture the full biological complexity of LNM. Advances in federated learning and privacy-preserving AI architectures could facilitate collaborative model development across institutions without compromising patient data security.

Moreover, clinical implementation studies are needed to assess the real-world impact of AI-based LNM prediction tools on patient outcomes, health care utilization, and cost-effectiveness. Such studies should incorporate calibration analyses, net benefit evaluations, and decision-analytic modeling to quantify the added value of AI in clinical workflows. Finally, transparent reporting following Transparent Reporting of a Multivariable Prediction Model of Individual Prognosis or Diagnosis for AI and Prediction model Risk of Bias Assessment Tool for AI guidelines,^31,32 along with public release of code and anonymized datasets, will be crucial for advancing the field and fostering reproducibility.

Conclusions

This systematic review and meta-analysis demonstrated that AI-based models achieve high diagnostic accuracy for predicting LNM in EGC and consistently outperform clinician assessment. Beyond diagnostic performance, these models show considerable potential as decision-supporting tools for minimally-invasive and endoscopic management by assisting clinicians in selecting appropriate treatment strategies and tailoring surgical extent. Prospective multicenter validation and implementation studies are warranted to facilitate the integration of AI-assisted prediction into minimally-invasive GC care.