This editorial refers to "Risk assessment of type I gastric neuroendocrine tumors based on endoscopic and clinical features of autoimmune gastritis" by Li et al, 2025; https://doi.org/10.3748/wjg.v31.i41.111449.
INTRODUCTION
Neuroendocrine neoplasms (NENs) comprise a heterogeneous group of tumors that originate from peptidergic neurons and neuroendocrine cells, exhibiting a wide range of clinical behavior from indolent, low-grade malignancies to aggressive, highly metastatic forms[1,2]. Gastroenteropancreatic NENs account for more than 70% of all NENs, and their detection has increased markedly in recent years due to enhanced awareness and diagnostic advancements[3]. Among these, the incidence of gastric neuroendocrine tumors (G-NETs) has shown a consistent upward trend[4]. G-NETs are classified into three subtypes - type I, type II, and type III - based on cellular origin and associated underlying conditions. Types I and II arise from enterochromaffin-like (ECL) cells located in the gastric fundus and body. Type I G-NETs are the most common, representing 70%-80% of cases, and show a strong association with autoimmune gastritis (AIG)[3,5].
Studies have reported that the prevalence of type I G-NETs among patients with AIG ranges from 1.0% to 12.5%[6]. Accordingly, risk stratification of G-NETs based on AIG-related factors is essential for accurate diagnosis and effective management. Despite this recognition, current risk assessment strategies for AIG-associated G-NETs remain limited, as they often rely on single biomarkers (e.g., serum gastrin or vitamin B12 levels) or isolated endoscopic findings. Such approaches fail to capture the multifactorial nature of tumorigenesis[7-9]. For example, although serum vitamin B12 deficiency is common in AIG[10], it lacks specificity for G-NET risk stratification, as it may arise from other etiologies or manifest only in advanced stages of gastritis[11,12]. Similarly, endoscopic evaluation of ECL cell hyperplasia, while critical, is operator-dependent and may underestimate pre-neoplastic changes during early disease stages.
To address these limitations, integrated models combining clinical, laboratory, and endoscopic features have been proposed as a more robust approach. Notably, in a recent issue of the World Journal of Gastroenterology, Li et al[13] applied the Boruta algorithm to identify five key predictors of G-NETs risk in AIG patients: Age, Helicobacter pylori (H. pylori) infection history, vitamin B12 levels, gastric corpus atrophy severity, and serum gastrin concentration. This data-driven selection of variables marks a paradigm shift toward precision risk stratification, as it accounts for both host factors (age) and disease-specific mechanisms (atrophy severity, hypergastrinemia). Such integrated models hold promise for improving early detection and guiding personalized surveillance strategies, thereby addressing the unmet need for accurate risk assessment in AIG patients at risk of developing G-NETs.
PATHOPHYSIOLOGICAL MECHANISM OF AIG-INDUCED G-NETS
G-NETs associated with AIG develop through a multistep pathophysiological process initiated by autoimmune-mediated destruction of the gastric oxyntic mucosa. This leads to impaired acid secretion, resulting in hypergastrinemia and the subsequent proliferation and neoplastic transformation of ECL cells[7,9]. This cascade involves complex interactions among immune dysregulation, hormonal imbalances, and alterations in the mucosal microenvironment, which collectively and synergistically promote tumorigenesis.
AIG is characterized by immune-mediated destruction of parietal cells in the gastric corpus and fundus, primarily driven by interferon-γ-secreting CD4+ Th1 cells. This immune response leads to chronic inflammation and oxyntic atrophy[14,15]. Progressive parietal cell loss results in hypochlorhydria or achlorhydria[16], which disrupts normal acid-dependent feedback inhibition and induces hypergastrinemia[17], with serum gastrin levels frequently exceeding 1000 pg/mL[18]. Chronically elevated gastrin stimulates ECL cells via cholecystokinin-2 receptors, promoting ECL cell hyperplasia through trophic signaling pathways[19]. Animal models have demonstrated that AIG progression includes stages ranging from chronic gastritis to mucous neck cell hyperplasia, metaplasia, and intraepithelial neoplasia, highlighting the condition’s precancerous potential. Sustained acid suppression, such as that observed in pernicious anemia or long-term PPI use, consistently induces ECL cell hyperplasia[20].
ECL cells in the oxyntic mucosa exhibit marked sensitivity to the trophic effects of gastrin[21]. Prolonged hypergastrinemia, particularly in the context of AIG, drives a progression from hyperplasia to dysplasia and eventually to the formation of type I G-NETs[22,23]. This neoplastic transformation is associated with molecular alterations such as loss of the MEN1 gene, overexpression of Bcl-2, and upregulation of various growth factors[21]. Gastrin functions as a tumor promoter, but additional cofactors, including chronic inflammation, are required for malignant transformation. The extent of oxyntic atrophy, largely determined by parietal cell loss and sustained hypergastrinemia, correlates closely with the risk of G-NET development.
ECL cells, the predominant endocrine cell population in the oxyntic mucosa, demonstrate pronounced responsiveness to gastrin-mediated trophic signaling. Under prolonged hypergastrinemic conditions, these cells undergo a stepwise pathological sequence - from physiological hyperplasia to dysplasia and, ultimately, neoplasia[21]. This progression typically begins with diffuse or linear hyperplasia, followed by micronodular and adenomatoid hyperplasia, and may culminate in the development of well-differentiated type I G-NETs. These tumors represent approximately 75% of all gastric NENs and are strongly associated with AIG[22,23]. The underlying molecular mechanisms involve several contributing factors, including loss of the MEN-1 tumor suppressor gene located at 11q13, overexpression of anti-apoptotic proteins such as Bcl-2 (which prolongs ECL cell survival and exposure to mitogenic stimuli), and upregulation of growth factors such as basic fibroblast growth factor and transforming growth factor-α[21]. Notably, gastrin functions primarily as a tumor promoter rather than a direct transforming agent, requiring additional cofactors - such as the chronic mucosal inflammation characteristic of AIG - to facilitate malignant progression[21].
In AIG, the extent of oxyntic atrophy correlates closely with the risk of G-NET development, primarily due to progressive parietal cell loss and reduction in functional mucosal tissue[24-26]. Severe corpus atrophy results in pronounced hypergastrinemia, which promotes ECL cell dysregulation and increases tumorigenic potential[27-29]. Endoscopic features such as mucosal thinning and the presence of visible vessels are indicative of advanced atrophy and are frequently associated with ECL cell hyperplasia or early-stage G-NETs[30]. Accurate differentiation between hyperplasia and neoplasia requires histopathological evaluation of ECL cells. Diagnostic accuracy in identifying ECL cell abnormalities is significantly higher when assessments are performed by specialized gastrointestinal pathologists.
METHODOLOGICAL EVALUATION OF INTEGRATED CLINICAL-ENDOSCOPIC MODELS
Feature selection is a critical component in developing integrated clinical-endoscopic models, as its methodological rigor directly affects model stability and predictive accuracy. Previous studies have shown that both the Boruta algorithm and the least absolute shrinkage and selection operator (LASSO) regression possess distinct advantages in reducing high-dimensional variable spaces. The Boruta algorithm determines feature importance by analyzing the z-score distribution of shadow attributes, effectively classifying features as confirmed (green), rejected (red), or tentative (blue). This classification is based on comparisons with the minimum, average, and maximum shadow feature thresholds, providing an interpretable visual framework for feature selection[31,32]. In comparative evaluations involving multiple algorithms, Boruta achieved an area under the curve (AUC) of 0.82 in identifying ferroptosis-related genes associated with diabetic osteoporosis - slightly lower than LASSO’s AUC of 0.84. However, Boruta demonstrated greater robustness in managing complex and interdependent feature sets[33].
The rigor of model validation is essential for ensuring clinical translational value. Most existing studies rely on single-center data, such as a prospective cohort of consecutive patients with autoimmune atrophic gastritis enrolled between 2000 and 2018, which established staging criteria for atrophy severity based on 1- to 3-year endoscopic follow-up. However, dependence on single-center datasets may introduce referral bias - such as the overrepresentation of complex cases - and may reflect region-specific population characteristics, both of which can limit the generalizability of predictive models[34]. To address these limitations, multicenter external validation is critical. Incorporating independent cohorts from diverse geographic regions and clinical settings enables a more robust assessment of model performance and stability across heterogeneous populations. For instance, a multicenter study by Lenti et al[35] reported that externally validated models demonstrated significantly greater consistency in risk stratification compared to those developed in single-center settings, although specific quantitative metrics were not disclosed in that study. Accordingly, integrated clinical-endoscopic models should undergo multicenter, cross-regional validation to reduce selection bias and enhance real-world applicability.
The clinical utility of a model is defined by its ability to enhance risk stratification relative to existing guidelines. Receiver operating characteristic (ROC) curves are widely recognized as the gold standard for evaluating diagnostic performance, providing a systematic representation of the trade-off between sensitivity and specificity. The AUC serves as a quantitative measure of discriminative accuracy[36]. The integrated model achieves an AUC of 0.830, which is comparable to high-performing diagnostic models developed from similar datasets, such as the LASSO model (AUC: 0.84) and the Boruta model (AUC: 0.82), indicating robust diagnostic capability[33]. When selecting optimal thresholds under conditions where sensitivity and specificity are equally weighted, the minimum sum of squares method - minimizing the squared deviations of (1-sensitivity) and (1-specificity) - provides a more precise determination of the optimal cutoff point in the upper-left region of the ROC curve, thereby reducing inter-method variability in threshold selection[37]. Compared to current clinical guidelines, such as the 2021 American Gastroenterological Association recommendations by Shah et al[12], the proposed model - which integrates endoscopic features (e.g., yellow-white cobblestone-like elevations in the fundic gland region, parietal cell pseudohypertrophy) and key clinical markers (e.g., anti-parietal cell antibodies) - demonstrates improved resolution for early risk detection. This makes it particularly well-suited for dynamic monitoring of patients at risk for AIG-associated G-NETs[38].
CLINICAL SIGNIFICANCE OF RISK STRATIFICATION MODELS IN PRACTICE
Risk stratification models for G-NETs associated with AIG are critical to optimizing clinical decision-making, offering substantial clinical value across three key dimensions: Risk-stratified management, optimization of therapeutic strategies, and prognostic assessment. These models not only improve the precision of clinical interventions but also enhance efficient resource utilization, particularly in primary care settings.
The heterogeneity of G-NETs necessitates individualized follow-up strategies tailored to tumor subtype and underlying pathophysiology. Type I G-NETs, which are strongly associated with AIG and hypergastrinemia, generally demonstrate benign behavior with low metastatic potential. In contrast, type III tumors are gastrin-independent, sporadic, and frequently metastatic, placing them in the highest risk category[39,40]. For high-risk individuals - such as younger patients or those with low pepsinogen I/II ratios - the model recommends endoscopic surveillance every 6-12 months to facilitate early detection of aggressive variants, including well-differentiated high-grade (G3) tumors characterized by a Ki-67 proliferation index ≥ 30%. These tumors may progress to liver metastasis and are associated with poor prognosis[41]. Conversely, low-risk patients - such as those with stable type I tumors and no evidence of proliferative activity - may safely extend surveillance intervals to 2-3 years. This approach is supported by evidence indicating that selected patients with synchronous G-NETs and duodenal gastrinomas remain free of recurrence or metastasis over at least two years of follow-up[42].
The association between AIG and vitamin B12 deficiency represents a promising target for chemoprevention. Clinical evidence indicates that vitamin B12 deficiency is one of the predominant hematologic manifestations in AIG, observed in 37 out of 109 patients, and individuals with AIG accompanied by iron or vitamin B12 deficiency commonly exhibit nonspecific gastrointestinal symptoms[43,44]. Given that AIG is a recognized precursor to G-NETs - with 4 out of 109 AIG patients progressing to gastric carcinoid tumors - vitamin B12 supplementation may not only correct hematologic abnormalities associated with deficiency but also potentially reduce the risk of G-NETs by ameliorating the underlying gastric atrophy[44]. Furthermore, endoscopic management can be prioritized for low-risk lesions, thereby reducing the need for more invasive interventions[45].
Integrating clinicopathological factors - such as proliferative status and tumor differentiation - into risk stratification models improves prognostic accuracy. Gastric ECL-cell NETs exhibit substantial variability in clinical outcomes depending on proliferative activity and the underlying pathological context. Combining these parameters enables more precise classification into distinct prognostic groups, which is particularly valuable when assessing small biopsy specimens[46]. For example, targeted sequencing of aggressive G-NETs in AIG has identified pathogenic variants, including ATM mutations, that are associated with metastatic potential and can be incorporated into prognostic frameworks to enhance risk assessment[41]. In contrast, the indolent behavior observed in most type I tumors - illustrated by the absence of recurrence in selected cases - supports favorable risk categorization. This, in turn, helps alleviate patient anxiety and minimizes unnecessary medical interventions and healthcare expenditures[42].
Variations in pathological evaluation - for example, only 23% of non-academic centers use triple immunostain panels for AIG - combined with the frequent use of nonspecific terminology such as "atrophic gastritis", often result in inappropriate testing and suboptimal follow-up[30]. Risk stratification models improve diagnostic consistency by integrating clinical factors (e.g., vitamin B12 deficiency), endoscopic findings (e.g., ECL cell hyperplasia), and histopathological markers (e.g., Ki-67 index), enabling primary care providers to accurately identify high-risk patients who require specialist referral and low-risk individuals appropriate for conservative monitoring. This approach minimizes unnecessary invasive procedures and supports current guidelines advocating long-term endoscopic surveillance in AIG patients due to their elevated risk of gastric cancer[38].
LIMITATIONS AND FUTURE DIRECTIONS
The current integrated clinical-endoscopic feature model for risk stratification of G-NETs in AIG presents several limitations that must be addressed to improve clinical utility and generalizability. Methodologically, reliance on single-center data introduces potential biases, such as referral bias, which may inflate performance estimates when the model is applied in broader, real-world settings. This aligns with existing evidence suggesting that single-center studies often lack external validity, highlighting the importance of multicenter collaborations and rigorous external validation to confirm model robustness[47].
Clinically, the model is limited by incomplete incorporation of key biological variables. For instance, the detection rate of gastrin levels remains suboptimal, with only 22 cases documented in a pivotal study[13], despite established associations between elevated gastrin and G-NET pathogenesis[38]. Furthermore, challenges in interpreting neuroendocrine immunostains and variability in diagnostic reporting practices for AIG hinder accurate risk stratification[30]. Methodological inconsistencies in statistical analyses - such as arbitrary selection of cut-points in ROC curve analysis - may result in divergent minimally important change thresholds, thereby influencing responder definitions and potentially altering trial outcomes[37].
Technologically, progress toward multi-modal models represents a promising direction. Integrating imaging omics - such as narrow-band imaging features[13] - with molecular markers, including anti-thyroid peroxidase antibodies[48], may improve predictive accuracy by simultaneously capturing morphological and biological heterogeneity. Furthermore, early endoscopic characteristics of AIG remain insufficiently characterized, highlighting the need for systematic investigation to facilitate early detection of G-NETs[38]. Clarifying the etiological role of H. pylori infection in the initiation of AIG is also essential, as current evidence supports the existence of infection-independent autoimmune gastric disease, which may have distinct implications for G-NETs risk[15]. Addressing these knowledge gaps will enhance the translational relevance of the model in guiding personalized clinical management of AIG-induced G-NETs.
CONCLUSION
This study systematically addresses the clinical challenge of risk assessment for AIG-related G-NETs by developing an integrated clinical-endoscopic feature model in combination with the Boruta algorithm. Beyond enabling precise risk stratification for G-NETs development in patients with AIG, the model's core value lies in addressing a critical gap in existing risk assessment systems, which often lack deep integration of specific biomarkers with clinical phenotypes. The model provides scientifically rigorous yet clinically applicable tools for decision-making. Methodologically, the study introduces an innovative approach by combining the Boruta feature selection algorithm with multi-dimensional clinical-endoscopic indicators, leveraging machine learning to uncover potential risk associations. This research framework presents a reusable paradigm for risk prediction in gastrointestinal tumors, overcoming the limitations of traditional single-factor analyses by emphasizing the synergistic value of multi-modal data. It also offers a valuable reference for risk modeling of other mucosa-associated tumors, such as colorectal adenomas and esophageal squamous dysplasia. Although validated in a single-center retrospective cohort, clinical translation of the model necessitates further validation. Future research should prioritize three areas: First, confirming external validity and temporal stability through multicenter, large-sample prospective cohorts; second, evaluating the added value of genomic markers (e.g., ATM gene mutations) in risk stratification by integrating abnormalities in DNA damage repair pathways, as suggested by the 2023 study on pathogenic ATM variants, to construct a comprehensive "clinical-imaging-molecular" three-tier risk assessment system; and third, developing clinical decision support systems based on the model to enable personalized screening and intervention strategies for AIG-related G-NETs, thereby advancing from risk stratification toward precision management.
Peer review: Externally peer reviewed.
Peer-review model: Single blind
Specialty type: Gastroenterology and hepatology
Country of origin: China
Peer-review report’s classification
Scientific quality: Grade A, Grade B, Grade B
Novelty: Grade A, Grade B, Grade C
Creativity or innovation: Grade B, Grade B, Grade D
Scientific significance: Grade B, Grade B, Grade B
P-Reviewer: Haruma K, MD, PhD, Professor Emerita, Japan; Yu HG, MD, Professor, China S-Editor: Li L L-Editor: A P-Editor: Wang WB
