Our multicenter retrospective observational study was approved by the Ethics Committee of the Second Affiliated Hospital of Anhui University of Chinese Medicine and the Hefei Second People’s Hospital and followed the Declaration of Helsinki, which was a multiple-center, retrospective, and observational analysis on CAG patients admitted to two hospitals. The clinical data of 317 patients from September 2018 to January 2025 in the Department of Gastroenterology of the Second Affiliated Hospital of Anhui University of Chinese Medicine were utilized for our predictive model as the training dataset, and the clinical data of 92 patients from November 2023 to January 2025 in the Department of Gastroenterology of the Hefei Second People’s Hospital were utilized for our predictive model as the validation dataset.

The inclusion criteria for our clinical data collection were as follows: a) age over 18 years old, b) patients diagnosed through endoscopy or pathology demonstrating CAG, c) patients having complete and searchable clinical information such as blood biomarkers and CAG classification data, and d) patients participating in our retrospective observational study voluntarily. The exclusion criteria were a) patients with CAG who received medical treatment in the past, b) patients having incomplete clinical data, and c) patients who were not willing to participate in our retrospective observational study.

The diagnosis of gastric LGIN requires comprehensive evaluation through multimodal methods. First, endoscopic examination is the core means. Conventional gastroscopy can initially identify mucosal abnormalities (such as erythema, erosion, or mucosal roughness), while enhanced imaging techniques (such as narrow band imaging, magnifying endoscopy, or chromoendoscopy) can further observe microvascular and glandular structural changes and assist in locating suspicious areas. Secondly, histopathological analysis is the key to diagnosis, and multiple biopsies (at least three to five pieces) are required to cover the range of lesions to avoid missed diagnosis; microscopic features include mild nuclear atypia and disordered arrangement, but the lesions are limited to the lower half of the mucosa and need to be differentiated from inflammation or reparative hyperplasia. In addition, HP detection is indispensable because HP infection is an important cause of LGIN, and eradication therapy may reverse some lesions. Finally, it is necessary to combine clinical follow-up to dynamically evaluate the condition. Some cases may progress to high-grade lesions or cancer, and timely intervention is required. During the diagnostic process, attention should be paid to the consistency differences between pathologists, and multidisciplinary consultation is recommended to improve accuracy. In general, the diagnosis of LGIN relies on the close integration of endoscopy and pathology, combined with etiological evaluation and dynamic monitoring, to provide a basis for subsequent treatment decisions.

In our retrospective analysis, the least absolute shrinkage and selection operator (LASSO) and multivariate logistic regression were utilized to establish our predictive model. In order to prevent overfitting of our predictive model, the number ratio (317:92) between the training dataset and the validation dataset, approximately 7:3 or 8:2, is reasonable (15).

Similar to the aforementioned literature (16, 17) and the aim of our retrospective analysis, the following characteristics were analyzed and studied: a) sociodemographic characteristics [gender of patients (male or female), age at diagnosis, education (less than primary school, middle school, or upper college), and marital status (single or married)], b) questionnaire information [obesity (no, yes), hypertension (no, yes) (systolic blood pressure 140 or diastolic blood pressure 90), depression (no, yes), frailty (no, yes), alcohol consumption (no, yes), smoking (no, yes), diabetes (no, yes), family history (no, yes), dyslipidemia (no, yes) (cholesterol 6.19 or low-density lipoprotein, LDL 4.14)], c) laboratory data [HP infection (no, yes), glucose (mmol/L), cholesterol (mmol/L), LDL (mmol/L), pepsinogen I (μg/L), pepsinogen II (μg/L), gastrin 17 (pmol/L), alpha fetoprotein (AFP) (μg/L), carcinoembryonic antigen (CEA) (μg/L), carbohydrate antigen 125 (CA125) (U/mL), carbohydrate antigen 199 (CA199) (U/mL), and D-dimer (mg/L)], and d) endoscopic characteristics [Kimura–Takemoto (KT) classification (C1, C2 or C3) and bile reflux (no, yes)].

Data were analyzed using R software 4.4.2 in our retrospective analysis. Continuous characteristics were presented as median (interquartile range), p-values were calculated via the Mann–Whitney U test, and comparisons between groups were analyzed by the rank sum test. Categorical characteristics were presented as number (N) or proportion (%), and comparisons between groups were analyzed using the chi-squared test or Fisher’s exact test. We employed the correlative analysis of independent characteristics via the function “cor()” and a heatmap of a correlation matrix was obtained, showing a graphical representation that utilized color-coding to visualize the strength and direction of the relationships between characteristics in a dataset. Each cell in the matrix showed the correlation coefficient ranging from −1 to 1 based on Pearson’s method. Furthermore, a nomogram was plotted to illustrate the risk of LGIN in patients with CAG, and the LASSO regression analysis was utilized to select relevant characteristics to establish our predictive model of high-dimensional data (18). After selecting predictors of LGIN in patients with CAG, the 10-fold cross-validation based on the LASSO regression analysis was utilized to confirm the suitable tuning parameters (λ), and the coefficients of a sparse matrix with non-zero as selected characteristics via the minimum λ were considered (19). Selected characteristics were applied via the multivariate logistic regression analysis, and p-values less than 0.05 were considered as independent variables of the radiomics nomogram (20).

The performance of our predictive model was assessed in both the training and validation datasets, including the assessment of the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity. Moreover, the calibration curve and Hosmer–Lemeshow goodness of fit were used to evaluate the effectiveness (21). Decision curve analysis (DCA) was conducted to validate the accuracy of the predictive model by quantifying the net benefits at different threshold probabilities (22).

In the construction of the clinical model and in the subsequent analysis, R version 4.4.2 (http://www.r-project.org, R Foundation for Statistical Computing) was utilized. For baseline characteristics, we utilized the “tableone” package, and the table was drawn via both the “flextable” and “officer” packages. In the LASSO regression analysis, the “glmnet” R package was utilized. For establishing the linear regression model, the “rms” package was used to plot the radiomics nomogram for subsequent analysis. We plotted the ROC and DCA curves based on the “pROC” and “rmda” packages, respectively, and the R package “ResourceSelection” was used to perform the Hosmer–Lemeshow goodness of fit. A two-bilateral p-value less than 0.05 was considered statistically significant.

Detailed baseline information of the sociodemographic and clinicopathologic characteristics of both the training and validation datasets is shown in Table 1 . For the training dataset, with a total of 317 CAG patients with non-LGIN (N = 92, 29%) and LGIN (N = 225, 71%), consisting of 144 male patients (45.4%) and 173 female patients (54.6%), the mean (SD) age was 62.03 ± 12.11 in non-LGIN patients and 61.28 ± 11.91 in LGIN utilized. For the validation dataset, with a total of 92 CAG patients with non-

LGIN (N = 27, 29.3%) and LGIN (N = 65, 70.7%), consisting of 39 male patients (42.4%) and 53 female patients (57.6%), the mean (SD) age was 60.59 ± 15.36 in non-LGIN patients and 59.88 ± 12.30 in LGIN patients, as shown in Supplementary Table S1 and Table 1 . Among the characteristics in both the training and validation datasets, nine characteristics showed significant differences between the training and validation datasets: hypertension (p < 0.001), dyslipidemia (p < 0.001), HP infection (p = 0.008), glucose (p = 0.001), pepsinogen II (p < 0.001), AFP (p < 0.001), CEA (p = 0.001), CA125 (p < 0.001), and CA199 (p = 0.003).

As depicted in Figure 1 , the correlation matrix heatmap of our training dataset revealed the causal associations between different predictive characteristics. In our heatmap of our correlation matrix, warm colors (e.g., red) often indicate strong positive correlations, cool colors (e.g., blue) represent negative correlations, and neutral colors (e.g., white) denote weak or no correlation. A correlation analysis was performed on the predictive variables of our model, which included gender, age, level of education, marital status, obesity, hypertension, depression, frailty, alcohol consumption, smoking, diabetes, dyslipidemia, family history, HP infection, glucose, cholesterol, pepsinogen I, pepsinogen II, gastrin 17, AFP, CEA, CA125, CA199, D-dimer, Kimura–Takemoto classification, and bile reflux. Considering that all characteristics have weak correlation and no multicollinearity in the visualization of the heatmap, we utilized these recorded characteristics for the subsequent LASSO regression analysis in our retrospective study.

In our LASSO logistic regression analysis, we leveraged 10-fold cross-validation to obtain the optimal parameter λ for our predictive model and finally screened nine characteristics: alcohol consumption, smoking, family history, HP infection, glucose, pepsinogen I,

pepsinogen II, gastrin 17, bile reflux, and Kimura–Takemoto classification (p < 0.05), as depicted in Figure 2 . Two LASSO result figures have chosen non-zero coefficients as the underlying factors of LGIN. Based on the results of the LASSO logistic regression analysis, the fitting equation of our predictive model is as follows:

where x 1 represents alcohol consumption (No denotes 0, Yes denotes 1), x 2 represents smoking (No denotes 0, Yes denotes 1), x 3 represents family history (No denotes 0, Yes denotes 1), x 4 represents HP infection (No denotes 0, Yes denotes 1), x 5 represents pepsinogen I value, x 6 represents pepsinogen II value, x 7 represents gastrin 17 value, x 8 represents bile reflux (No denotes 0, Yes denotes 1), x 9 represents Kimura–Takemoto classification (C1, C2, C3), and the constant term of the formula (0.601) means reference intercept.

Multivariate logistic regression analysis further showed that family history (OR = 3.111, 95% CI = 1.620–6.039), HP infection (OR = 4.810, 95% CI = 2.335–10.16), pepsinogen I (OR = 0.982, 95% CI = 0.970–0.993), pepsinogen II (OR = 0.832, 95% CI = 0.755–0.912), bile reflux (OR = 2.388, 95% CI = 1.212–4.727), and Kimura–Takemoto classification (C3 vs. C1) (OR = 3.874, 95% CI = 1.693–9.264) are independent risk factors for LGIN in CAG patients (p < 0.05), respectively, as presented in Table 2 . The β values of three characteristics, encompassing pepsinogen I, pepsinogen II, and gastrin 17, were all less than zero, and the OR values of the aforementioned characteristics were also less than zero, which means that these three characteristics are the protective factors for LGIN in CAG patients.

Meanwhile, the predictive model was plotted as a radiomics nomogram, constructed using family history, HP infection, pepsinogen I, pepsinogen II, bile reflux, and Kimura–Takemoto classification based on the aforementioned risk factors, as shown in Figure 3 . Our radiomics nomogram provides a visual representation of the impact of each factor, helping doctors in conducting individualized risk evaluations in clinical practice. For example, if a patient with CAG had family history, HP infection with pepsinogen I (170) and pepsinogen II (17), bile reflux, and C2 classification, then the patient’s corresponding scores would be approximately 43, 60, 0, 0, 30, and 30, respectively, with a total score of 133. This would reveal that the estimated probability of CAG patients with LGIN is approximately 23%.

As shown in Figures 4A, B , the discrimination power of our predictive model was assessed by AUC

values calculated in ROC figures by analyzing the indication of LGIN in CAG patients in both the training and validation datasets. The ROC figure in the training dataset calculated an AUC value of 0.838 (95% CI = 0.789–0.887) with a specificity of 0.761 and a sensitivity of 0.791 as well as an AUC value of 0.941 (95% CI = 0.893–0.989) with a specificity of 0.852 and a sensitivity of 0.908 in the validation dataset. The AUC in the validation dataset (0.941) had a higher score than in the training dataset (0.838), which means that our predictive model has a good effect. Furthermore, the Hosmer and Lemeshow goodness-of-fit (GOF) test and the calibration curve were utilized to evaluate our predictive model, and a p-value of the Hosmer and Lemeshow GOF test greater than 0.05 indicates that the predictive model has a good degree of fit. The results showed that our model had a good fit for the training (χ ² = 4.1407, df = 8, p-value = 0.8442) and validation datasets (χ ² = 3.3873, df = 8, p-value = 0.9078). The calibration curves for the radiomics nomogram based on our multivariate logistic regression analysis in the training and validation datasets are depicted in Figures 5A, B and demonstrated good agreement between the prediction results and the observational outcomes. In clinical practice, calibration curves are commonly used to evaluate and optimize predictive models such as the probability of postoperative complications for LGIN in CAG patients. By comparing the predicted probability with the actual occurrence probability, doctors can develop more accurate postoperative monitoring and intervention plans. Moreover, for decision curve analysis, the DCA in both the training and validation datasets indicated that the net benefit of our predictive model was consistently better than the two extreme strategies (all treatment and no treatment) across a wide range of threshold probabilities, representing its underlying clinical ability, as shown in Figure 6 . According to the DCA, doctors can choose the most appropriate intervention threshold based on changes in net benefit. This helps avoid excessive or inappropriate intervention and improve the quality of medical decision-making for LGIN in CAG patients. Good calibration curve and the DCA revealed that our model has good calibration, clinical application, and generalization.