Perioperative NLR/PLR dynamic phenotypes and prediction of postoperative recurrence in locally advanced gastric cancer: a retrospective cohort study

Abstract

Background

Systemic inflammation plays a critical role in tumor progression and postoperative recurrence in gastric cancer. The neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR) are readily accessible biomarkers of systemic inflammatory and immune responses, but their dynamic changes around surgery remain underexplored. This study aimed to assess the prognostic value of principal component–derived phenotypes based on perioperative NLR and PLR—specifically, principal component 1 (PC1; preoperative inflammatory burden) and principal component 2 (PC2; postoperative inflammatory recovery dynamics)—for predicting disease-free survival (DFS) in patients with locally advanced gastric cancer (LAGC).

Methods

We retrospectively analyzed data from 375 patients with LAGC who underwent radical gastrectomy at the Affiliated Hospital of Jiangnan University between January 2018 and December 2022, with follow-up through March 2025. NLR and PLR were measured one week before and one month after surgery. Principal component analysis (PCA) was applied to derive two independent inflammatory phenotypes (PC1 and PC2). These components, together with T stage, N stage, and preoperative carcinoembryonic antigen (CEA), were incorporated into a Cox proportional hazards regression model, and internal validation was performed using bootstrap resampling (1000 times). Model performance was evaluated using Kaplan–Meier survival analysis, log-rank testing, and time-dependent receiver operating characteristic (ROC) curve analysis.

Results

PCA identified two independent components: PC1 (preoperative inflammatory burden; variance explained 40.24%) and PC2 (postoperative inflammatory recovery; variance explained 36.16%). Multivariate Cox regression revealed that N stage, T stage, preoperative carcinoembryonic antigen (CEA), PC1, and PC2 were independent predictors of DFS (all p < 0.05). Stratified analysis demonstrated that patients with the “Dual-High” phenotype (high PC1 and high PC2) had a markedly shorter median DFS (13.1 months) compared with those with the “Dual-Low” phenotype (74.6 months; p < 0.001). Internal validation showed good discriminative ability of the model (bias-corrected concordance index [C-index] = 0.762), and time-dependent ROC curves indicated stable predictive performance at 12, 36, and 60 months postoperatively.

Conclusions

PCA-derived perioperative inflammatory phenotypes effectively stratify recurrence risk in LAGC. The “Dual-High” phenotype indicates poor prognosis, and combining these phenotypes with T stage, N stage, and preoperative CEA enhances individualized postoperative surveillance and adjuvant therapy decision-making.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16431-2.

Background

Gastric cancer remains one of the most prevalent malignancies worldwide, and its high recurrence rate and poor prognosis continue to pose major challenges for clinical management. Although radical surgery combined with adjuvant therapy has improved overall survival (OS), postoperative recurrence and metastasis remain the principal determinants of disease-free survival (DFS) in patients with gastric cancer [1]. Data show that the postoperative recurrence rate is approximately 42% in patients undergoing radical gastrectomy, with distinct early recurrence characteristics: about 79% are detected within 2 years after surgery, and recurrence patterns mainly include locoregional recurrence, peritoneal metastasis, and distant organ metastasis [2].

Identifying effective prognostic predictors—particularly biomarkers that reflect inflammation and the tumor microenvironment—is essential for early detection of high-risk patients and the optimization of individualized treatment strategies. In recent years, systemic inflammatory markers such as the neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR) have demonstrated significant prognostic value across various malignancies [3, 4]. These markers reflect the balance between host immunity and systemic inflammation and are closely associated with tumor aggressiveness, immune evasion, and angiogenesis [5].

However, most existing studies have evaluated inflammatory markers at a single preoperative time point. In contrast, dynamic perioperative changes—potentially influenced by surgical trauma, infection, residual tumor burden, or micrometastases—have not been systematically investigated in relation to prognosis. Moreover, because NLR and PLR are interrelated through shared lymphocyte counts, traditional univariate analyses may fail to capture their synergistic prognostic effects. This limitation underscores the need for more integrated analytic approaches.

Principal component analysis (PCA), a dimensionality reduction method, can transform correlated variables into a smaller set of independent components while retaining most of the original information. By mitigating multicollinearity, PCA enhances model stability and interpretability, providing a methodological foundation for constructing robust and clinically interpretable recurrence risk stratification tools [6]. Therefore, this study applied PCA to integrate perioperative dynamics of NLR and PLR. Combined with clinicopathological characteristics and tumor markers, we aimed to develop and validate a robust predictive model for DFS in patients with locally advanced gastric cancer (LAGC), providing a theoretical basis for identifying high-risk patients and guiding individualized postoperative management.

Methods

Study design and patient population

This retrospective cohort study included 375 patients with LAGC who underwent radical gastrectomy at the Affiliated Hospital of Jiangnan University between January 2018 and December 2022.

Inclusion criteria were as follows: (1) pathologically confirmed gastric adenocarcinoma consistent with the American Joint Committee on Cancer (AJCC) 8th edition TNM staging system (T3–T4 invasion with or without regional lymph node metastasis and no distant metastasis) [7]; (2) no history of other malignancies; (3) standard D2 radical gastrectomy performed in accordance with the Japanese Gastric Cancer Treatment Guidelines [8]; (4) complete clinical, pathological, and follow-up data; (5) no preoperative neoadjuvant chemotherapy or radiotherapy; and (6) receipt of postoperative adjuvant chemotherapy with complete documentation of the regimen.

Exclusion criteria included: (1) severe comorbidities (e.g., liver cirrhosis or cardiac insufficiency); (2) distant metastasis (M1) or palliative surgery; (3) survival of less than 3 months after surgery; and (4) receipt of preoperative neoadjuvant chemotherapy or radiotherapy.

The final cohort comprised 269 men and 106 women with a median age of 65 years (interquartile range [IQR], 11; range, 26–80). Pathological characteristics were as follows: histologic differentiation (poor, 167; moderate–poor, 137; moderate, 62; well–moderate, 6; well, 3), histologic type (non-specific, 337; mucinous/signet-ring cell, 38), lymphovascular invasion (negative, 119; positive, 256), and perineural invasion (negative, 110; positive, 265). TNM staging included T3 (n = 89), T4a (n = 261), and T4b (n = 25); and N0 (n = 82), N1 (n = 64), N2 (n = 90), and N3 (n = 139). Postoperative complications graded by the Clavien-Dindo classification [9] included grade I (n = 135), grade II (n = 228), and grade III (n = 12). Chemotherapy regimens were grouped as XELOX (n = 83), SOX (n = 131), monotherapy (n = 28), and others (n = 133).

This study was approved by the Institutional Ethics Committee of the Affiliated Hospital of Jiangnan University(LS2025428). All procedures were conducted in accordance with the Declaration of Helsinki.

Laboratory measurements

Peripheral venous blood samples (2 mL) were collected from all patients at the following two time points after an overnight fast: (1) within one week before surgery (i.e., the last examination prior to radical surgery); and (2) approximately one month after surgery (at which time patients had not yet started adjuvant chemotherapy, serving as a routine postoperative follow-up or pre-adjuvant chemotherapy assessment). Complete blood counts—including neutrophil, platelet, and lymphocyte counts—were measured using an automated hematology analyzer. The NLR and PLR were then calculated.

Preoperative serum levels of tumor markers, including carcinoembryonic antigen (CEA), carbohydrate antigen 19 − 9 (CA19-9), and carbohydrate antigen 125 (CA125), were quantified using a chemiluminescence immunoassay. Preoperative albumin levels were measured using a biochemical analyzer, and a cutoff value of 35 g/L was used for subgroup analysis.

Follow-up and endpoint definition

Patients were followed up through outpatient visits, telephone interviews, and review of electronic medical records. Imaging evaluations included contrast-enhanced computed tomography (CT) of the chest, abdomen, and pelvis every 3–6 months. Suspected recurrent lesions were further evaluated by positron emission tomography–CT (PET–CT) or endoscopic ultrasound–guided biopsy.

Tumor recurrence was defined according to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1 [10] as one or more of the following: (1) radiologic confirmation of new lesions or progression of existing lesions on CT or magnetic resonance imaging (MRI); (2) histopathologic confirmation of malignancy by biopsy; or (3) clinical evidence of recurrence supported by imaging findings.

Follow-up continued until March 2025. The primary endpoint was DFS, defined as the time from surgery to tumor recurrence, diagnosis of a second primary malignancy, or death from any cause [11]. Patients without events were censored at the date of the last follow-up.

Statistical analyses

Normality tests of all continuous variables confirmed their non-normal distribution. Continuous variables were expressed as medians (IQR) and compared using the Mann–Whitney U test, while categorical variables were presented as numbers (percentages) and compared using the χ² test. PCA was applied to perioperative NLR and PLR values using the Kaiser criterion (eigenvalue > 1) for component extraction, followed by varimax rotation. Survival curves were generated using the Kaplan–Meier method, and differences were evaluated using the log-rank test. Prognostic variables were analyzed using univariate and multivariate Cox proportional hazards regression models. All the above statistical analyses were performed using SPSS software (version 25.0; IBM Corp., Armonk, NY, USA).

To further evaluate the internal stability and predictive performance of the model, additional analyses were performed using R software (version 4.3.1). A prognostic model was constructed based on the independent prognostic factors identified by multivariate Cox regression. Bootstrap resampling (1000 times) was used to calculate the bias-corrected concordance index (C-index) to assess model discrimination and correct for optimism. The optimal cutoff value was determined by time-dependent receiver operating characteristic (ROC) curves, and patients were divided into high-risk and low-risk groups. DFS was compared between the two groups using the Kaplan–Meier method and the log-rank test. Time-dependent ROC curves were used to evaluate the predictive accuracy of the model at 12, 36, and 60 months, and the corresponding areas under the curve (AUC) were calculated. In addition, a nomogram was constructed to visualize the predictive model, and calibration curves were used to assess the agreement between the model-predicted DFS probabilities at 12, 36, and 60 months and the actual observed probabilities.

A two-sided p-value < 0.05 was considered statistically significant.

Results

Comparison of inflammatory indicators and tumor markers

Shapiro–Wilk tests confirmed that all continuous variables violated the assumption of normality (p < 0.05); therefore, nonparametric Mann–Whitney U tests were used for comparisons. Patients in the recurrence group showed significantly higher preoperative NLR (NLR1), postoperative NLR (NLR2), preoperative PLR (PLR1), postoperative PLR (PLR2), and preoperative CA125 levels compared with the non-recurrence group (all p < 0.05). No significant differences were observed in preoperative CEA or CA19-9 levels between the two groups (Table 1).

Table 1.

Comparison of inflammatory indicators and tumor markers between recurrence and non-recurrence groups [median (interquartile range)]

Indicator	Recurrence group (n = 123)	Nonrecurrence group (n = 252)	U-value	Z-value	p-value
NLR1	2.36 (1.64)	2.00 (1.21)	12,113.5	-3.434	0.001
PLR1	169.38 (104.00)	127.51 (67.42)	9,771.0	-5.811	< 0.001
NLR2	1.71 (1.56)	1.50 (0.87)	12,308.0	-3.237	0.001
PLR2	151.82 (96.19)	118.51 (54.36)	10,091.0	-5.487	< 0.001
Preoperative CEA	2.50 (4.24)	2.28 (3.32)	13,863.5	-1.659	0.097
Preoperative CA19-9	10.25 (21.89)	9.63 (16.26)	14,722.5	-0.787	0.431
Preoperative CA125	10.04 (8.61)	9.73 (6.70)	12,720.5	-2.819	0.005

Data are presented as median (interquartile range). Statistical comparisons were performed using the Mann-Whitney U test

Comparison of clinical, pathological, and treatment characteristics

Chi-square (χ²) analysis demonstrated that T stage, N stage, and preoperative albumin level were significantly associated with postoperative recurrence. Patients with T4b tumors or N3 lymph node involvement exhibited the highest recurrence risk, and patients with preoperative albumin < 35 g/L had a higher recurrence rate than those with albumin ≥ 35 g/L. Lymphovascular invasion showed borderline significance (p = 0.058), with a higher recurrence rate among patients with positive invasion compared with those without. No statistically significant differences were observed in other characteristics, including age, sex, tumor differentiation, pathological type, perineural invasion, Clavien-Dindo complication grade, or chemotherapy regimen (Table 2).

Table 2.

Comparison of clinical, pathological, and treatment characteristics between recurrence and non-recurrence groups [n (%)]

Factor	Category	Total cases	Recurrence group (n = 123)	Non-recurrence group (n = 252)	χ²-value	p- value
Sex	Men Women	269 106	91 (33.8) 32 (30.2)	178 (66.2) 74 (69.8)	0.457	0.499
Age group	< 60 years ≥ 60 years	114 261	42 (36.8) 81 (31.0)	72 (63.2) 180 (69.0)	1.214	0.271
Differentiation	Poor Moderate-poor Moderate Well-moderate Well	167 137 62 6 3	55 (32.9) 48 (35.0) 18 (29.0) 1 (16.7) 1 (33.3)	112 (67.1) 89 (65.0) 44 (71.0) 5 (83.3) 2 (66.7)	1.420	0.841
Pathological type	Non-specific Mucinous/signet-ring	337 38	114 (33.8) 9 (23.7)	223 (66.2) 29 (76.3)	1.594	0.207
Lymphovascular invasion	No Yes	119 256	31 (26.1) 92 (35.9)	88 (73.9) 164 (64.1)	3.603	0.058
Perineural invasion	No Yes	110 265	31 (28.2) 92 (34.7)	79 (71.8) 173 (65.3)	1.506	0.220
T stage	T3 T4a T4b	89 261 25	27 (30.3) 80 (30.7) 16 (64.0)	62 (69.7) 181 (69.3) 9 (36.0)	11.833	0.003
N stage	N0 N1 N2 N3	82 64 90 139	7 (8.5) 14 (21.9) 25 (27.8) 77 (55.4)	75 (91.5) 50 (78.1) 65 (72.2) 62 (44.6)	58.595	< 0.001
Preoperative albumin	≥ 35 g/L < 35 g/L	309 66	90 (29.1) 33 (50.0)	219 (70.9) 33 (50.0)	10.751	0.001
Clavien-Dindo grade	I II III	135 228 12	43 (31.9) 75 (32.9) 5 (41.7)	92 (68.1) 153 (67.1) 7 (58.3)	0.484	0.785
Chemotherapy regimen	XELOX SOX Monotherapy Others	83 131 28 133	24 (28.9) 43 (32.8) 12 (42.9) 44 (33.1)	59 (71.1) 88 (67.2) 16 (57.1) 89 (66.9)	1.858	0.602

Data are expressed as number (percentage). Statistical comparisons were performed using the chi-square (χ²) test

Principal component analysis (PCA)

Data suitability and component extraction

All continuous variables, including NLR and PLR were standardized using z-scores before analysis. Bartlett’s test of sphericity indicated significant intercorrelations among variables (χ² = 316.05, p < 0.001), confirming the appropriateness of PCA (Table 3). The variable loading plot and sample factor score distribution further demonstrated good consistency among the four indices (Fig. 1 and Fig. S1). According to the Kaiser criterion (eigenvalue > 1), two principal components were extracted. The scree plot exhibited a clear “elbow” inflection point after the second component (Fig. 2). Together, these two components explained 76.41% of the total variance, suggesting that they adequately represented the underlying structure of the perioperative inflammatory data (Table 3).

Table 3.

Key results of principal component analysis

Indicator	Value/Parameter
Bartlett’s test	χ² = 316.05, p < 0.001
Cumulative variance explained	76.41%

Data were derived from principal component analysis of perioperative inflammatory indicators (NLR and PLR)

Fig. 1.

Variable loadings plot of PCA for perioperative inflammatory indices. The PCA variable plot presents the loadings of four perioperative inflammatory indices (preoperative NLR and PLR, postoperative NLR and PLR) on the first two principal components. Close clustering of preoperative and postoperative indices respectively confirms favorable internal consistency, supporting the rationality of PCA for dimensionality reduction. Abbreviations: PCA, principal component analysis; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; PC, principal component

Fig. 2.

Scree plot of PCA for perioperative inflammatory indicators. The scree plot demonstrates that two principal components (PC1 and PC2) with eigenvalues > 1 were retained according to the Kaiser criterion, explaining the majority of total variance in the data. Abbreviations: PCA, principal component analysis; PC, principal component

Clinical interpretation of components

The first component (PC1), termed preoperative inflammatory burden, was primarily driven by preoperative PLR (loading = 0.89) and NLR (loading = 0.83), reflecting baseline systemic inflammation. The second component (PC2), termed postoperative inflammatory recovery, showed strong correlation with postoperative NLR (loading = 0.91) and PLR (loading = 0.76), representing inflammatory recovery dynamics after surgery (Table 4).

Table 4.

Rotated component matrix (varimax rotation)

Variable	PC1 (Preoperative inflammatory burden)	PC2 (Postoperative inflammatory recovery)
PLR1	0.89	0.12
NLR1	0.83	0.14
NLR2	-	0.91
PLR2	0.36	0.76

Rotation method: varimax with Kaiser normalization. Loadings < 0.10 are omitted

Orthogonality and component coefficients

The covariance matrix revealed zero covariance between PC1 and PC2, confirming their statistical independence. The component score coefficients for each variable are presented in Table 5. These coefficients were used to calculate the PC1 and PC2 scores for subsequent survival analysis using the following formulas (where zNLR1, etc., are z-scores calculated based on the mean and standard deviation of the cohort):

Table 5.

Component score coefficients

Variable	PC1 coefficient	PC2 coefficient
NLR1	0.544	−0.086
PLR1	0.589	−0.118
NLR2	−0.219	0.704
PLR2	0.066	0.506

Coefficients represent the weighting of each variable used to compute PC1 and PC2 scores in principal component analysis

Summary

PCA identified two statistically independent and biologically meaningful components—PC1 (preoperative inflammatory burden) and PC2 (postoperative inflammatory recovery)—which served as composite predictors for subsequent survival modeling.

Univariate and multivariate Cox regression analyses of prognostic factors

Univariate and multivariate Cox proportional hazards regression analyses were performed to evaluate the associations between clinicopathological characteristics and recurrence in patients with gastric cancer. Univariate analysis identified N stage, T stage, preoperative CA125, preoperative CEA, lymphovascular invasion, preoperative albumin, PC1, and PC2 as significant prognostic factors (all p < 0.05). In the multivariate model, N stage, T stage, preoperative CEA, PC1, and PC2 remained independent predictors of recurrence (all p < 0.05), whereas CA125, lymphovascular invasion, and preoperative albumin lost significance after adjustment (Table 6; robustness of T4b effect: Fig. S2, Table S1).

Table 6.

Univariate and multivariate cox regression analyses of prognostic factors in patients with gastric cancer

Factor	Univariate analysis HR (95% CI)	p-value	Multivariate analysis HR (95% CI)	p-value
Sex (Ref: Men)	0.86 (0.57–1.28)	0.457	–	–
Age group (Ref: <60 years)	0.90 (0.62–1.30)	0.571	–	–
Differentiation (Ref: Moderately/Well)	1.32 (0.82–2.13)	0.260	–	–
Pathological type (Ref: Non-specific)	0.84 (0.43–1.66)	0.615	–	–
N stage	1.93 (1.59–2.34)	< 0.001	1.93 (1.56–2.38)	< 0.001
T stage (Ref: T4b) T3 T4a	0.35 (0.19–0.66) 0.38 (0.22–0.65)	0.001 0.001 < 0.001	0.45 (0.24–0.84) 0.45 (0.26–0.80)	0.017 0.012 0.006
Preoperative CEA (Ref: <5 ng/mL)	1.79 (1.21–2.66)	0.004	1.93 (1.27–2.93)	0.002
Preoperative CA125 (Ref: <35 U/mL)	3.03 (1.63–5.63)	< 0.001	1.16 (0.57–2.37)	0.681
Preoperative CA19-9 (Ref: <25 U/mL)	1.32 (0.89–1.97)	0.165	–	–
Lymphovascular invasion (Ref: Absent)	1.58 (1.05–2.38)	0.028	0.93 (0.60–1.44)	0.746
Perineural invasion (Ref: Absent)	1.36 (0.90–2.04)	0.140
Preoperative albumin(Ref: ≥35 g/L)	1.76 (1.18–2.63)	0.005	1.25 (0.81–1.95)	0.316
Clavien-Dindo grade (Ref: I) II III	1.02 (0.70–1.47) 1.26 (0.50–3.20)	0.889 0.939 0.629	–	–
Chemotherapy regimen (Ref: SOX) XELOX Monotherapy Others	1.14 (0.69–1.88) 1.84 (0.97–3.50) 0.99 (0.65–1.51)	0.262 0.616 0.064 0.950	–	–
PC1 (per 1 SD increase)	1.38 (1.22–1.56)	< 0.001	1.28 (1.12–1.46)	< 0.001
PC2 (per 1 SD increase)	1.40 (1.24–1.59)	< 0.001	1.44 (1.26–1.64)	< 0.001

Differentiation groups: Poor (including Poor and Moderate-poor) vs. Moderately/Well (including Moderate, Well-moderate, and Well). PC1 and PC2 represent standardized principal component scores derived from perioperative inflammatory markers

HR  Hazard ratio, CI  Confidence interval, SD  Standard deviation

Survival analysis and prognostic validation of inflammatory principal components (PC1 and PC2)

Based on optimal ROC-derived cut-off values (PC1: − 0.099; PC2: 0.393), patients were stratified into high and low subgroups for each component. Kaplan–Meier survival analysis revealed that the low PC1 group had significantly longer median DFS compared with the high PC1 group (74.6 vs. 29.6 months; p < 0.001) (Fig. 3A). Similarly, the low PC2 group showed a significantly longer median DFS than the high PC2 group (56.7 vs. 17.1 months; p < 0.001) (Fig. 3B).

Fig. 3.

Kaplan–Meier curves for disease-free survival (DFS) according to perioperative inflammatory principal components. A DFS by PC1 (preoperative inflammatory burden). Patients with high PC1 scores (red line) had significantly shorter median DFS (29.6 months) compared with those with low PC1 scores (blue line; 74.6 months; log-rank test, p < 0.001). B DFS by PC2 (postoperative inflammatory recovery). High PC2 scores (red line) were associated with markedly reduced median DFS (17.1 months) versus low PC2 scores (blue line; 56.7 months; log-rank test, p < 0.001). C DFS stratified by combined inflammatory phenotypes. The “Dual-High” phenotype (yellow line; high PC1 + high PC2) was associated with the worst prognosis (median DFS = 13.1 months), whereas the “Dual-Low” group (blue line; low PC1 + low PC2) had the most favorable outcome (median DFS = 74.6 months; overall log-rank test, p < 0.001). The intermediate phenotypes — “PC1-High Only” (red line) and “PC2-High Only” (green line) — did not differ significantly from each other (p = 0.494), but each conferred intermediate recurrence risk between the “Dual-Low” and “Dual-High” groups. Abbreviations: DFS, disease-free survival; PC, principal component; PC1, preoperative inflammatory burden; PC2, postoperative inflammatory recovery

To assess the combined effects of these components, patients were further divided into four groups: Dual-Low (low PC1 + low PC2), PC1-High Only (high PC1 + low PC2), PC2-High Only (low PC1 + high PC2), and Dual-High (high PC1 + high PC2) (Fig. 3C). Pairwise log-rank tests demonstrated that the Dual-Low group had the longest median DFS (74.6 months), which was significantly longer than that of all other groups (p < 0.001). In contrast, the Dual-High group had the shortest median DFS (13.1 months), indicating the poorest prognosis (p < 0.05). No significant difference was observed between the PC1-High Only (34.2 months) and PC2-High Only (23.9 months) groups (p = 0.494), although both had significantly better DFS than the Dual-High group (Table 7).

Table 7.

Log-rank test results for median disease-free survival among patient subgroups

Comparison groups	χ²-value	p-value
Dual-Low vs. PC1-High Only	24.015	< 0.001
Dual-Low vs. PC2-High Only	20.585	< 0.001
Dual-Low vs. Dual-High	75.361	< 0.001
PC1-High Only vs. PC2-High Only	0.467	0.494
PC1-High Only vs. Dual-High	14.902	< 0.001
PC2-High Only vs. Dual-High	4.538	0.033

Group definitions: Dual-Low (low PC1 + low PC2), PC1-High Only (high PC1 + low PC2), PC2-High Only (low PC1 + high PC2), Dual-High (high PC1 + high PC2). Statistical method: log-rank test

DFS  Disease-free survival, PC  Principal component

Cox proportional hazards regression confirmed that the Dual-High group had the highest risk of recurrence (HR = 7.24, 95% CI: 4.36–12.01), followed by PC2-High Only (HR = 3.72, 95% CI: 2.00–6.91) and PC1-High Only (HR = 3.04, 95% CI: 1.89–4.88). All three groups had significantly higher recurrence risks compared with the Dual-Low group (all p < 0.001) (Table 8).

Table 8.

Cox regression analysis comparing risk among patient subgroups

Group comparison	B (coefficient)	SE (standard error)	Wald statistic	p-value	HR (95% CI)
Dual-High vs. Dual-Low	1.980	0.258	58.746	< 0.001	7.24 (4.36–12.01)
PC2-High Only vs. Dual-Low	1.313	0.316	17.285	< 0.001	3.72 (2.00–6.91)
PC1-High Only vs. Dual-Low	1.110	0.243	20.913	< 0.001	3.04 (1.89–4.88)

Group definitions: Dual-Low (low PC1 + low PC2), PC1-High Only (high PC1 + low PC2), PC2-High Only (low PC1 + high PC2), Dual-High (high PC1 + high PC2). Statistical method: Cox proportional hazards regression

HR  Hazard ratio, CI Confidence interval, SE  Standard error

Internal validation of the prognostic model and risk score formula

A prognostic model was constructed based on multivariate Cox regression, incorporating T stage, N stage, preoperative CEA, PC1, and PC2 as independent prognostic factors. The individualized risk score (linear predictor) for each patient was calculated using the following formula:

risk_score = 0.0057 × T4a + 0.8422 × T4b + 0.6499 × N + 0.6406 × CEA1 + 0.2734 × PC1 + 0.3772 × PC2.

where T4a and T4b are binary dummy variables (1 = yes, 0 = no, with T3 as reference), CEA1 is a binary variable (1 = elevated, 0 = normal), N is an ordinal variable (N0 = 0, N1 = 1, N2 = 2, N3 = 3), and PC1 and PC2 are continuous principal component scores.

Internal validation was performed using 1000 bootstrap resamples, yielding a bias-corrected C-index of 0.762 (95% CI: 0.723–0.811), indicating good discriminative ability of the model with no significant overfitting.

Risk stratification and model discrimination

Based on the continuous risk score, the optimal cutoff value of 0.847 was determined by time-dependent ROC curves, and patients were divided into low-risk and high-risk groups. Kaplan–Meier survival analysis showed that DFS was significantly shorter in the high-risk group than in the low-risk group (HR = 5.60, 95% CI: 3.84–8.15, log-rank P < 0.001) (Fig. 4).

Fig. 4.

Kaplan–Meier curves for disease-free survival (DFS) stratified by the model-derived risk score. Kaplan–Meier curves for DFS stratified by the model-derived risk score. Patients were divided into low-risk (blue line) and high-risk (red line) groups using the optimal cutoff (0.847) determined by the time-dependent ROC curve. The high-risk group had significantly shorter DFS compared with the low-risk group (log-rank p < 0.001; HR = 5.60, 95% CI: 3.84–8.15). The number at risk is shown below the plot. Abbreviations: DFS, disease-free survival; CI, confidence interval; ROC, receiver operating characteristic

Time-dependent ROC analysis demonstrated stable predictive performance of the model at 12, 36, and 60 months: AUC at 12 months = 0.813 (95% CI: 0.739–0.877), at 36 months = 0.823 (95% CI: 0.759–0.884), and at 60 months = 0.817 (95% CI: 0.706–0.907) (Fig. 5).

Fig. 5.

Time-dependent receiver operating characteristic (ROC) curves of the prognostic model. Time-dependent ROC curves of the prognostic model at 12, 36, and 60 months. The blue curve represents the 12-month ROC (AUC = 0.813), the red curve represents the 36-month ROC (AUC = 0.823), and the green curve represents the 60-month ROC (AUC = 0.817). The model demonstrated good and stable discriminative ability across all time points. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve

Nomogram construction and calibration curve analysis

A nomogram for predicting DFS was constructed based on T stage, N stage, preoperative CEA, PC1, and PC2 (Fig. 6). The nomogram visually displays the relative contribution of each factor to the risk score and provides individualized predictions of DFS probabilities at 12, 36, and 60 months, facilitating rapid clinical assessment.

Fig. 6.

Nomogram for predicting disease-free survival (DFS). Nomogram for predicting 1-year, 3-year, and 5-year disease-free survival (DFS) based on the final Cox model. The nomogram integrates T stage, N stage, preoperative CEA, PC1 (preoperative inflammatory burden), and PC2 (postoperative inflammatory recovery). To use the nomogram, locate each patient’s variable values on the corresponding axis, draw a vertical line upward to the “Points” axis to obtain points for each variable, sum the points, find the total points on the “Total Points” axis, and draw a vertical line downward to the DFS probability axes. The contribution of each factor to the risk score is reflected by the length of its point scale. Abbreviations: DFS, disease-free survival; CEA, carcinoembryonic antigen; PC, principal component

Calibration curve analysis showed good agreement between nomogram-predicted and actual observed probabilities for 1-year DFS, while mild overestimation was observed for 3-year and 5-year predictions (Fig. 7). These findings suggest that the model is more accurate for short-term predictions, and long-term predictions should be interpreted cautiously in conjunction with clinical circumstances.

Fig. 7.

Calibration curves of the nomogram for 1-, 3-, and 5-year disease-free survival (DFS). Calibration curves of the nomogram at 1 year (yellow circles), 3 years (blue squares), and 5 years (red diamonds). The x-axis represents the predicted DFS probability, and the y-axis represents the observed DFS probability. The diagonal dashed line indicates perfect prediction. The calibration curves show good agreement between predicted and observed probabilities at 1 year, with mild overestimation at 3 and 5 years, suggesting that the model is more accurate for short-term predictions. Abbreviations: DFS, disease-free survival

Discussion

Patients with LAGC remain at high risk of postoperative recurrence even after radical surgery, underscoring the need for reliable prognostic indicators. The systemic inflammatory response is closely linked to tumor progression and recurrence. In this study, elevated NLR and PLR were associated with poorer prognosis, which may be explained by several mechanisms within the tumor microenvironment.

First, lymphocytes play a pivotal role in antitumor immunity [12, 13]. Lymphopenia may weaken host immune surveillance, allowing tumor cells to evade immune recognition and metastasize [14].

Second, platelets facilitate tumor growth and dissemination by promoting angiogenesis, tumor cell migration, and distant metastasis [15, 16]. They directly interact with tumor cells through receptors such as P-selectin and glycoprotein Ibα (GPIbα) [17] and release cytokines and growth factors that enhance proliferation, invasion, and metastatic potential [18].

Third, neutrophil extracellular traps (NETs) also contribute to metastasis. Their formation involves mitochondrial reactive oxygen species (ROS) and peptidylarginine deiminase 4 (PAD4), and proteases within NETs (e.g., neutrophil elastase [NE] and matrix metalloproteinase-9 [MMP-9]) degrade the extracellular matrix and activate pro-metastatic signaling pathways [19].

In the present study, we systematically analyzed dynamic perioperative changes in NLR and PLR. Patients who developed recurrence exhibited significantly higher preoperative (NLR1, PLR1) and postoperative (NLR2, PLR2) levels, indicating that perioperative inflammatory burden is closely associated with recurrence. PCA identified two independent components: PC1 (preoperative inflammatory burden) and PC2 (postoperative inflammatory recovery).

Univariate and multivariate Cox regression analyses confirmed that conventional clinical factors—including T stage, N stage, and preoperative CEA—remained significant prognostic predictors. Importantly, PC1 and PC2 also emerged as independent predictors after adjustment, suggesting that dynamic inflammatory changes confer additional prognostic value beyond traditional indicators.

To rigorously control for confounders, we systematically evaluated the impact of postoperative complications, adjuvant therapy, and preoperative albumin levels. Postoperative complications were graded according to the Clavien–Dindo classification; their severity was not independently associated with DFS and did not alter the prognostic effect of PC2, suggesting that the prognostic value of postoperative inflammatory recovery is not driven by complications such as surgical trauma, infection, or anastomotic leakage. The blood sampling time point at approximately one month after surgery already avoided the peak of acute inflammatory response following surgical trauma (typically 7–14 days postoperatively). In this cohort, mild-to-moderate complications (Clavien–Dindo grades I and II, accounting for 96.8%) were mostly resolved clinically within two weeks after surgery and did not persist until the blood sampling time point. All patients received adjuvant chemotherapy, but univariate analysis showed no significant association between chemotherapy regimen and DFS, and blood collection at one month postoperatively was performed before the initiation of chemotherapy. In summary, the postoperative inflammatory recovery trajectory reflected by PC2 is neither a transient effect of complications nor a chemotherapy effect, but rather represents the patient’s natural recovery and immune reconstitution process, further supporting its reliability as an independent prognostic indicator. In addition, low preoperative albumin was associated with poor prognosis in univariate analysis but lost independent significance after multivariate adjustment, suggesting that its nutritional-inflammatory effects are largely captured by PC1 and other clinicopathological factors.

A key contribution of this study is moving beyond static indicators and confirming the independent prognostic significance of postoperative inflammatory recovery dynamics. Surgical trauma can induce transient physiological inflammation and short-term immunosuppression, creating a perioperative “immune window”. The natural resolution of postoperative inflammation and the efficiency of immune reconstitution directly affect the body’s ability to monitor and control potential micrometastases. Survival stratification based on PC1 and PC2 demonstrated that patients with the “Dual-High” phenotype (high PC1 + high PC2) had a markedly elevated recurrence risk, suggesting that high preoperative inflammatory burden combined with poor postoperative immune reconstitution synergistically exacerbates risk and identifies this phenotype as a key high-risk population. The “Dual-High” phenotype indicates a profoundly immunosuppressive tumor microenvironment, which may underlie its unfavorable prognosis. For patients with this high-inflammatory phenotype, future strategies could explore perioperative immune microenvironment remodeling and sequential interventions to optimize antitumor immunity and improve prognosis.

To further evaluate the robustness of the model, internal validation was performed using 1000 bootstrap resamples. The results showed good discriminative ability and low overfitting, and the individualized risk score effectively distinguished low-risk from high-risk patients, providing a quantitative basis for developing differentiated follow-up strategies in clinical practice. Time-dependent ROC curves demonstrated stable predictive performance of the model at 12, 36, and 60 months postoperatively. The nomogram enabled model visualization, facilitating rapid clinical assessment of individual recurrence probability. Calibration curves indicated higher accuracy for short-term predictions, with mild overestimation for long-term predictions, which may be related to long-term effects of adjuvant therapy or competing risk events, suggesting that long-term prognosis should be interpreted in conjunction with subsequent treatment response.

In summary, dynamic perioperative changes in NLR and PLR are valuable predictors of recurrence in LAGC. PCA-derived inflammatory components (PC1 and PC2), together with T stage, N stage, and preoperative CEA, form a robust prognostic model capable of identifying high-risk patients with the “Dual-High” phenotype. This model provides an actionable tool for individualized postoperative management. Nevertheless, this study has limitations, including its single-center retrospective design and modest sample size. Moreover, the absence of external validation means the model’s generalizability remains unproven. Although internal bootstrap validation suggests robustness, independent validation in multicenter cohorts is mandatory before clinical application. Therefore, multicenter prospective studies with larger cohorts are needed for external validation and recalibration, as well as to explore perioperative intervention strategies targeting the high-inflammatory phenotype.

Abbreviations

AUC

Area under the curve

C-index

Concordance index

CA19-9

Carbohydrate antigen 19 − 9

CA125

Carbohydrate antigen 125

CEA

Carcinoembryonic antigen

CI

Confidence interval

CT

Computed tomography

DFS

Disease-free survival

GPIbα

Glycoprotein Ib alpha

HR

Hazard ratio

IQR

Interquartile range

LAGC

Locally advanced gastric cancer

LOO

Leave-one-out

MMP-9

Matrix metalloproteinase-9

MRI

Magnetic resonance imaging

NE

Neutrophil elastase

NETs

Neutrophil extracellular traps

NLR

Neutrophil-to-lymphocyte ratio

OS

Overall survival

PAD4

Peptidylarginine deiminase 4

PCA

Principal component analysis

PC

Principal component

PC1

Principal component 1 (preoperative inflammatory burden)

PC2

Principal component 2 (postoperative inflammatory recovery)

PET–CT

Positron emission tomography–computed tomography

PLR

Platelet-to-lymphocyte ratio

ROC

Receiver operating characteristic

ROS

Reactive oxygen species

SD

Standard deviation

TNM

Tumor–node–metastasis

Authors’ contributions

Y.W. was responsible for the study design, data collection, data processing, analysis and interpretation, as well as drafting and reviewing and editing the manuscript. D.C. was responsible for the study conception and design. J.Z. was responsible for the acquisition and interpretation of data. S.L. and Y.X. were responsible for the critical revision of the manuscript. All authors have read and approved the final submitted manuscript.

Funding

No funding was received.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the guidelines of the Declaration of Helsinki and approved by the Medical Ethics Committee of the Affiliated Hospital of Jiangnan University (LS2025428). Informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.