Construction and validation of a risk prediction model for perineural invasion in early gastric cancer based on multivariate analysis

Abstract

Objective

We aimed to identify independent risk factors for perineural invasion (PNI) in early gastric cancer (EGC) and to construct the first individualized nomogram for predicting PNI risk.

Methods

We retrospectively analyzed clinicopathological data from 416 EGC patients who underwent radical gastrectomy between December 2019 and August 2025. The optimal set of risk predictors for PNI was selected using the LASSO regression model with ten-fold cross-validation. Independent risk factors were subsequently identified via multivariable logistic regression analysis. For internal validation, we randomly selected 30% of the sample as a validation set using R software (version 4.4.2). The model’s performance was comprehensively evaluated by assessing its discrimination (area under the receiver operating characteristic curve, AUC), calibration (Hosmer-Lemeshow test and calibration curve), and clinical utility (decision curve analysis, DCA).

Results

A total of 416 patients were included in the final analysis, among whom 30 (7.21%) had PNI. LASSO regression analysis identified eight predictors for PNI: age, CEA level (ng/mL), tumor location, maximum tumor thickness, tumor differentiation, lymphovascular invasion, Lauren classification, and pT stage. These variables were subsequently incorporated into a multivariable logistic regression model. The analysis revealed that age (OR = 1.105, 95% CI: 1.029–1.187, P = 0.006), CEA level (OR = 59.489, 95% CI: 3.456–1023.871, P = 0.005), maximum tumor thickness (OR = 38.807, 95% CI: 3.408–441.872, P = 0.003), and lymphovascular invasion (OR = 4.131, 95% CI: 1.337–12.768, P = 0.014) were independent risk factors for PNI in EGC (all P < 0.05). The nomogram demonstrated strong discriminative ability, with AUC values of 0.895 (95% CI: 0.839–0.950) in the training cohort and 0.783 (95% CI: 0.625–0.940) in the validation cohort. The Hosmer–Lemeshow test indicated good model calibration in both the training (χ² = 11.994, P = 0.152) and validation cohorts (χ² = 3.833, P = 0.872). DCA showed substantial clinical net benefits across a wide range of threshold probabilities.

Conclusion

In conclusion, this study identified age, CEA level, maximum tumor thickness, and lymphovascular invasion as independent predictors of PNI in EGC. We developed the first nomogram for individualized PNI risk assessment, which demonstrated strong predictive performance, good calibration, and clinical usefulness. Although this tool offers a reliable approach for personalized risk evaluation, further multicenter validation is necessary to enhance its clinical applicability.

Background

The prognosis of EGC is generally favorable. However, perineural invasion (PNI)—a key marker of aggressive tumor behavior—is strongly associated with tumor recurrence and poor patient prognosis [1]. Current evidence on the incidence and clinical impact of PNI in EGC remains limited, and no reliable predictive tools are currently available. While previous studies have primarily focused on PNI in advanced gastric cancer, risk factors and predictive models specific to EGC have not yet been established [2]. This study aims to systematically identify independent risk factors for PNI in EGC and to develop the first clinicopathology-based nomogram for PNI prediction. By integrating preoperative serological markers, imaging features, and pathological characteristics, the model may support personalized treatment planning and improve prognostic assessment for EGC patients.

Materials and methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of the Weifang People’s Hospital. Due to its retrospective nature, the requirement for informed consent was waived by the Institutional Review Board of the Weifang People’s Hospital.

Patients

This retrospective cohort study included 416 EGC patients who underwent radical gastrectomy at the First Affiliated Hospital of Shandong Second Medical University between December 2019 and August 2025. Inclusion criteria: (1) Pathologically confirmed EGC (pT1 stage); (2) No preoperative anticancer therapy; (3) No history of other malignancies; (4) Complete clinicopathological records. Exclusion criteria: (1) Postoperative pathological confirmation of advanced gastric cancer (pT2−4 stage); (2) Concurrent systemic malignancies; (3) Incomplete clinical data.

Data collection

Baseline parameters included: Age, Sex, body mass index (BMI), Hypertension, Diabetes_mellitus, Smoking_history, CEA, CA19−9, CA25, CA72−4, AFP, preoperative albumin, preoperative hemoglobin, maximum tumor thickness, tumor location (gastric cardia/fundus, body, antrum), tumor differentiation (well/moderately/poorly), histological classification (pure/mixed type), endoscopic tumor morphology (non-ulcerated/ulcerated type), tumor diameter (cm), perineural invasion, lymphovascular invasion, total lymph nodes dissected, Lauren classification (intestinal/mixed/diffuse), pT stage (T1a/T1b), and pN stage (N0/N1+).

Statistical analysis

To leverage all available data and enhance the stability of the coefficient estimates, the final multivariable logistic regression model (for the nomogram) was constructed based on the entire cohort (n = 416), considering the limited number of PNI-positive cases.

All statistical analyses were performed using R software (version 4.4.2). The optimal set of risk predictors for PNI was selected using the LASSO regression model with ten-fold cross-validation. Independent risk factors were subsequently identified via multivariable logistic regression analysis [3–5]. A nomogram model for PNI risk prediction was constructed based on multivariate analysis outcomes [6]. Model calibration was evaluated using the Hosmer-Lemeshow test, while discriminative performance was assessed via receiver operating characteristic (ROC) curve analysis, quantified by the area under the curve (AUC). For internal validation, 30% of the cohort was stratified and randomly selected as a validation set [7]. P < 0.05 was considered statistically significant.

Results

A total of 416 patients with EGC who underwent radical gastrectomy were included in this study. The mean age of the cohort was 63.1 ± 9.1 years, and 71.4% (297/416) were male. The most common tumor location was the gastric antrum (42.3%, 176/416), and the majority of tumors were poorly differentiated (54.3%, 226/416). According to the Lauren classification, 65.4% (272/416) were intestinal type and 33.4% (139/416) were diffuse type. The overall incidence of lymphovascular invasion (LVI) was 26.7% (111/416), and perineural invasion (PNI) was present in 7.21% (n = 30) of the cohort. Comparative analysis showed statistically significant differences (P < 0.05) between PNI-positive and PNI-negative groups in several clinicopathological features, including serum CEA levels, maximum tumor thickness, tumor location, differentiation grade, endoscopic tumor morphology, LVI, Lauren classification, pT stage, and pN stage. For detailed data, see Table 1.

Table 1.

Comparison of clinicopathological characteristics between PNI-Negative and PNI-Positive patient groups

Variable	Description/Values	PNI(–) (n = 386)	PNI(+) (n = 30)	χ2 / t value	P value
Sex	Female/Male	110/276	9/21	0.031	0.861
Age (years)		62.91 ± 9.25	66.17 ± 6.45	−1.891	0.059
Body Mass Index (BMI)		23.48 ± 3.36	23.70 ± 4.21	−0.342	0.732
Hypertension	No / Yes	271/115	20/10	0.166	0.684
Diabetes mellitus	No / Yes	327/59	24/6	0.469	0.493
Smoking history	No / Yes	226/160	16/14	0.311	0.577
CEA (ng/mL)	< 5 / ≥5	377/9	25/5	17.589	0.000
CA19−9 (U/mL)	< 34 / ≥34	367/19	27/3	1.433	0.231
CA72−4 (U/mL)	< 6.9 / ≥6.9	343/43	29/1	1.794	0.180
AFP (ng/mL)	< 7 / ≥7	372/13	28/2	0.865	0.352
Preoperative albumin (g/L)		40.42 ± 4.73	39.77 ± 4.07	0.733	0.464
Preoperative hemoglobin (g/L)		132.54 ± 19.78	133.40 ± 21.33	−0.228	0.820
Maximum tumor thickness		0.33 ± 0.19	0.65 ± 0.30	−8.297	0.000
Tumor location	Gastric cardia/fundus, body, and antrum	83/146/157	3/8/19	6.096	0.049
Tumor differentiation	Well / Moderately / Poorly	33/155/198	1/1/28	20.046	0.000
Histological classification	Pure/Mixed type	179/207	8/22	4.369	0.039
Endoscopic Tumor Morphology	Non-ulcerated/Ulcerated type	174/212	4/26	11.459	0.001
Tumor diameter (cm)	< 2/ ≥2	138/248	6/24	3.051	0.081
lymphovascular invasion	No / Yes	297/89	8/22	35.967	0.000
Total lymph nodes dissected		29.05 ± 11.11	30.10 ± 10.30	−0.500	0.617
Lauren classification	Intestinal / Mixed / Diffuse	264/4/118	8/1/21	21.587	0.000
pT stage	T1a/ T1b	173/213	1/29	19.689	0.000
pN stage	N0 / N1 +	323/63	17/13	13.603	0.000

LASSO regression for selecting predictors of perineural invasion in EGC

The perineural invasion (PNI) status (negative vs. positive) was designated as the dependent variable, with 24 candidate predictors included as independent variables. To mitigate multicollinearity and select the most informative predictors, least absolute shrinkage and selection operator (LASSO) regression was employed. As illustrated in Fig. 1A, as the penalty parameter λ increased, the coefficients of the variables were progressively shrunk toward zero, with some ultimately eliminated, thus reducing model overfitting. Through ten-fold cross-validation, eight variables retained non-zero coefficients (Fig. 1B). These selected predictors were: age, CEA level, tumor location, maximum tumor thickness, tumor differentiation, lymphovascular invasion, Lauren classification, and pT stage.

Fig. 1.

Selection of potential predictors by the least absolute shrinkage and selection operator(LASSO) regression

Multivariate logistic regression analysis of PNI in patients with EGC

The eight predictors selected from the LASSO regression analysis were included in a multivariate logistic regression model. The analysis identified the following independent risk factors for PNI in EGC (all P < 0.05): age (OR = 1.105, 95% CI: 1.029–1.187, P = 0.006), CEA level (OR = 59.489, 95% CI: 3.456–1023.871, P = 0.005), maximal tumor thickness (OR = 38.807, 95% CI: 3.408–441.872, P = 0.003), and lymphovascular invasion (OR = 4.131, 95% CI: 1.337–12.768, P = 0.014). Detailed results are provided in Table 2.

Table 2.

Multivariate analysis of factors associated with PNI in EGC

Variable	B	SE	Wald χ2	P	OR	95% CI
Age	0.100	0.036	7.543	0.006	1.105	1.029–1.187
CEA (ng/mL)	4.086	1.452	7.920	0.005	59.489	3.456−1023.871
Tumor location
Gastric cardia	-	-	-	-	1	-
body	0.123	0.844	0.021	0.884	1.131	0.216–5.921
antrum	0.868	0.757	1.312	0.252	2.381	0.540−10.504
Maximum tumor thickness	3.659	1.241	8.690	0.003	38.807	3.408−441.872
Tumor differentiation
Well	-	-	-	-	1	-
Moderately	−1.696	5534.489	0.000	1.000	0.183	0.000-.
Poorly	18.105	4932.199	0.000	0.997	72906465.9	0.000-.
lymphovascular invasion	1.419	0.576	6.071	0.014	4.131	1.337–12.768
Lauren classification
Intestinal	-	-	-	-	1	-
Mixed	−0.833	0.571	2.126	0.145	0.435	0.142–1.332
Diffuse	−0.532	1.287	0.171	0.680	0.588	0.047–7.325
pT stage	17.779	2452.218	0.000	0.994	52633408.4	0.000-.

Development of a nomogram model for predicting PNI in EGC

A predictive nomogram for PNI in EGC was developed based on four independent risk factors identified by multivariate logistic regression analysis: age, CEA ≥ 5 ng/mL, maximum tumor thickness, and lymphovascular invasion. The total points derived from the nomogram demonstrated a positive correlation with PNI risk, with higher scores corresponding to increasing probabilities of PNI occurrence (Fig. 2).

Fig. 2.

Nomogram prediction model for perineural invasion (PNI) in EGC. [Note: The nomogram was constructed based on the multivariable logistic regression analysis of the entire cohort (n = 416).]

Validation of the nomogram model for PNI in EGC

ROC curve analysis demonstrated area under the curve (AUC) values of 0.895 (95% CI: 0.839–0.950) in the training cohort (Fig. 3A. ) and 0.783 (95% CI: 0.625–0.940) in the validation cohort (Fig. 3B. ), indicating high discriminative ability of the nomogram model. The Hosmer-Lemeshow test and calibration curves revealed no statistically significant deviation between observed outcomes and predicted probabilities in both the training cohort (χ² = 11.994, P = 0.1515) and validation cohort (χ² = 3.8334, P = 0.8718), confirming excellent calibration performance of the model.(Figure 4A and B).

Fig. 3.

ROC curves of the nomogram prediction model. (A) Training cohort, (B) Validation cohort

Fig. 4.

Calibration curves of the nomogram prediction model. (A) Training cohort, (B) Validation cohort

Clinical decision curve analysis of the nomogram model

Clinical decision curve analysis confirmed the clinical utility of the nomogram, which demonstrated superior net benefits over both treat-all and treat-none strategies across threshold probabilities of 0.01–0.99 (training cohort) and 0.01–0.95 (validation cohort), indicating its strong potential for clinical application (Fig. 5) [8, 9].

Fig. 5.

Clinical decision curve analysis (DCA) of the nomogram prediction model. (A) Training cohort DCA, (B) Validation cohort DCA

Discussion

Perineural invasion (PNI) is a significant aggressive biological behavior in gastric cancer, closely associated with tumor recurrence, metastasis, and poor prognosis. However, in patients with EGC, the incidence, clinical significance, and prediction of PNI remain underexplored, creating a critical gap in precision clinical management. Through a retrospective analysis of 416 EGC patients, this study established and validated for the first time a nomogram model for predicting PNI risk in EGC. The model incorporates four independent predictors—age, serum CEA level, maximum tumor thickness, and lymphovascular invasion (LVI)—and demonstrated favorable discriminative ability, calibration, and clinical utility in both the training and internal validation cohorts. It thus provides the first quantitative tool for individualized risk assessment and treatment decision-making in EGC.

Clinical and biological interpretation of key predictors

The four independent predictors identified in this study elucidate the risk of perineural invasion (PNI) from distinct perspectives. Advanced age as a risk factor may reflect declining immune surveillance or accumulated genetic damage, creating a microenvironment conducive to tumor cells acquiring an invasive phenotype, including PNI [10]. Elevated serum CEA level (≥ 5 ng/mL) represents one of the most robust predictors in the model. Carcinoembryonic antigen (CEA) is not merely a classical tumor marker but also functions as a cell adhesion molecule that actively participates in tumor progression [11]. Studies indicate that CEA overexpression may disrupt E-cadherin-mediated intercellular junctions and activate signaling pathways such as PI3K/Akt and Wnt/β-catenin, thereby promoting epithelial-mesenchymal transition (EMT). As a crucial step in tumor cells acquiring migratory and invasive capabilities, EMT finds its manifestation in the perineural space as PNI [12–14]. Maximum tumor thickness directly quantifies the depth of vertical invasion within the gastric wall. Within EGC ( pT1 stage), greater thickness typically indicates deeper submucosal invasion, which not only increases the physical probability of contacting the submucosal plexus but is also associated with more active proliferative and invasive biological characteristics [15, 16]. The co-occurrence of lymphovascular invasion (LVI) with PNI demonstrates significant synergistic predictive value. As parallel micro-pathways for tumor invasion and metastasis, LVI and PNI often share similar molecular regulatory mechanisms (e.g., upregulation of angiogenesis and neuroguidance factors like VEGF and ANGPT), reflecting the activation of a comprehensive invasive phenotype [17–21]. Thus, the presence of LVI serves as a strong indicator of the tumor’s potential for dissemination through multiple routes, including PNI.

Innovations and clinical implications of the model

The principal innovation of this study resides in its dedicated focus on developing a prediction model specifically tailored to EGC. To date, the majority of predictive models for perineural invasion (PNI) in gastric cancer have been established based on cohorts of advanced gastric cancer (AGC). These models typically incorporate late-stage features, such as deep T-stage invasion (T3/T4), lymph node metastasis (N+), large tumor diameter, and serosal involvement. However, EGC differs fundamentally from AGC in terms of tumor burden, biological behavior, and treatment strategies. Applying AGC-derived predictors directly to EGC populations thus lacks biological plausibility and clinical utility.In contrast, our model was developed based on clinicopathological characteristics intrinsic to EGC, identifying predictors with relevance at this earlier disease stage—such as tumor thickness (as opposed to T-stage) and preoperative CEA levels. This effort fills a critical gap in EGC-specific tools for individualized PNI risk assessment. Moreover, the intended clinical applications of our model—including pre-operative risk evaluation before endoscopic resection and informing adjuvant therapy decisions following radical EGC surgery—differ substantially from those of existing AGC-based models, underscoring its targeted clinical relevance and distinctiveness.

Limitations and future directions

We acknowledge the limitations of this study. First, The key limitation of this study is the relatively low number of PNI-positive events (n = 30, 7.2%). This resulted in a low events-per-variable (EPV) ratio, which may compromise the stability of the multivariable logistic regression model and its statistical power.The predictive accuracy of the model will require further validation and calibration in future large-scale cohorts. Second, the model’s partial reliance on postoperative pathological variables (e.g., LVI and precise tumor thickness) limits its immediate application in purely preoperative settings. Nevertheless, the model holds dual clinical value: on one hand, it provides a tool for precise postoperative risk stratification, where high-risk scores may warrant closer follow-up or prompt multidisciplinary discussion regarding adjuvant therapy; on the other hand, it identifies core predictors, establishing a theoretical foundation and clear direction for developing a purely preoperative model—for instance, by incorporating preoperative imaging-based tumor thickness, radiomic features, and serum biomarkers.

In conclusion, this study successfully developed and initially validated the first nomogram for predicting perineural invasion (PNI) risk in patients with EGC. The model incorporates four readily available clinicopathological variables—age, CEA level, tumor thickness, and lymphovascular invasion (LVI)—and demonstrated favorable predictive performance. Not only does it provide the first individualized tool for assessing PNI risk in EGC patients, but it also shifts research attention toward a previously overlooked area, highlighting the importance of evaluating aggressive biological behaviors like PNI even at the earliest stages of gastric cancer. Future efforts involving large-scale, multi-center external validation and the development of purely preoperative prediction models will further refine this tool, ultimately helping to advance EGC management toward more precise and personalized care.

Authors’ contributions

Guangxu Zhu , Xuguang Jiao, Yanpeng Gao and Xiaomeng Wang conjectured the study and reviewed the paper; Guangxu Zhu, and Shengjie Zhou analyzed the data and wrote the draft; Jianjun Qu selected the patients and collected the data.

Funding

① Supported by the Weifang Youth Medical Talent lift project.

② Shandong Province Medical Health Science and Technology Project (No. 202304010544).

Data availability

The datasets utilized and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of the Weifang People’s Hospital. Due to its retrospective nature, the requirement for informed consent was waived by the Institutional Review Board of the Weifang People’s Hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.