Extrapancreatic extension modifies chemotherapy benefit in resected pancreatic ductal adenocarcinoma: a population-based cohort study

Abstract

The effect of extrapancreatic extension (EPE) on the benefit of chemotherapy remains unclear. We analyzed 22,045 surgically treated patients with resected PDAC from the SEER database (2004–2022). Variable importance was assessed using Random Survival Forest models with SHAP and permutation methods. Cox proportional hazards models evaluated survival associations and EPE-chemotherapy interactions on multiplicative and additive scales, with Fine-Gray competing risk models for sensitivity analyses. Variable importance analysis identified chemotherapy, EPE, and N stage as the most influential prognostic factors. Chemotherapy absence increased disease-specific mortality (adjusted HR 1.45, 95% CI 1.39–1.52, p < 0.001), while EPE conferred additional risk (adjusted HR 1.25, 95% CI 1.19–1.31, p < 0.001). Notably, within the EPE stratum, chemotherapy absence increased mortality risk by 83% (adjusted HR 1.83, 95% CI 1.75–1.91) versus 34% (adjusted HR 1.34, 95% CI 1.23–1.46) in non-EPE patients. EPE patients receiving chemotherapy achieved comparable adjusted 3-year disease-specific survival to non-EPE patients without chemotherapy (39.5% vs 35.0%). Significant multiplicative interaction (HR 1.37, 95% CI 1.24–1.50, p < 0.001) and additive interaction (RERI 0.63, 95% CI 0.51–0.76, p < 0.001) were observed, with consistent patterns across tumor size categories validated in competing risk analyses. This population-based analysis revealed EPE as an independent adverse prognostic factor in resected PDAC and a significant EPE-chemotherapy interaction on both multiplicative and additive scales across tumor size categories. Despite conferring worse prognosis, EPE-positive patients receiving chemotherapy achieved survival comparable to that of EPE-negative patients without chemotherapy, underscoring the importance of ensuring receipt of chemotherapy in EPE-positive disease. These findings support the need to re-evaluate the role of EPE in staging systems; however, prospective validation is warranted.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-47990-w.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies worldwide, with a 5-year survival rate that has only recently surpassed 10% despite significant therapeutic advances^1,2. For the minority of patients presenting with resectable disease, adjuvant chemotherapy has become the standard of care following landmark trials including PRODIGE-24³ and ESPAC-4⁴, which established modern combination regimens including modified FOLFIRINOX (mFOLFIRINOX) and gemcitabine with capecitabine as superior to older monotherapies. However, substantial heterogeneity exists in treatment response ^5,6. This substantial variability underscores the critical need for identifying factors that predict chemotherapy effectiveness to optimize patient selection and treatment strategies.

Extrapancreatic extension (EPE), defined as tumor direct infiltration beyond the pancreatic parenchyma into peripancreatic soft tissues, represents a controversial yet potentially crucial prognostic factor in PDAC. The 8th edition of the American Joint Committee on Cancer (AJCC) staging system removed EPE from T-stage criteria⁷, transitioning to a purely size-based classification due to concerns about reproducibility and inter-observer variability⁸. However, accumulating evidence challenges this decision. Multiple studies demonstrate that EPE remains a powerful independent prognostic factor, with hazard ratios ranging from 1.5 to 2.0 for disease-specific mortality ^9–11. Furthermore, modified staging systems that retain EPE criteria have demonstrated superior prognostic discrimination compared to the AJCC 8th edition across several large validation studies^7,9–11, suggesting its continued clinical relevance in risk stratification^6,9,11,12.

Despite extensive research establishing both EPE status and chemotherapy as independent prognostic factors, no studies have formally investigated whether EPE modifies chemotherapy effectiveness—a critical knowledge gap that limits our ability to personalize treatment selection. We therefore conducted a comprehensive population-based analysis using the SEER database to investigate the interaction between EPE and chemotherapy in resected PDAC. This study aimed to: (1) quantify the independent prognostic impact of EPE and chemotherapy; (2) determine whether EPE modifies the effectiveness of chemotherapy through formal interaction analysis; and (3) examine whether this interaction varies across tumor size categories. This investigation represents the first systematic evaluation of EPE-chemotherapy interactions and may inform therapeutic decision-making in resected PDAC.

Methods

Patients and data collection

This retrospective cohort study utilized data from the Surveillance, Epidemiology, and End Results (SEER) database (Incidence—SEER Research Data, 17 Registries, Nov 2024 Sub [2000–2022]), accessed via the National Cancer Institute’s SEER*Stat software (version 9.0.40.1). Patients diagnosed with pancreatic cancer were identified using the International Classification of Diseases for Oncology, Third Edition (ICD-O-3 site codes: C25.0-C25.9). The study cohort comprised patients with histologically confirmed pancreatic ductal adenocarcinoma (PDAC; ICD-O-3 morphology codes: 8140, 8500) diagnosed between 2004 and 2022. Exclusion criteria comprised: (1) age less than 18 years at diagnosis; (2) T4 stage disease or unknown T stage; (3) distant metastases (M1 stage) or unknown M stage; (4) unknown tumor size, N stage, or county-level demographic data; (5) patients who did not undergo surgical treatment; (6) follow-up duration less than one month; and (7) unknown survival status. The final analysis included 22,045 eligible patients (Supplementary Fig. 1). The study protocol received exemption from the Institutional Review Board given the de-identified, publicly available nature of the SEER database.

Variable collection

Comprehensive variables were collected across three domains. Demographic variables included age at diagnosis, sex, race (White, Black, Other/Unknown), marital status (Married, Single, Divorced/Widowed/Separated, Unknown), household income (low, middle, or high), and county type (urban or rural). Clinicopathological characteristics encompassed year of diagnosis (2004–2010, 2011–2017, or 2018–2022), primary tumor site (head, body, tail, or overlapping/unknown), histological grade (I/II, III/IV, or unknown), tumor size categories (0-20 mm, 20-40 mm, or greater than 40 mm), N stage classification per AJCC 8th edition (N0, N1, or N2). Extrapancreatic extension (EPE) was defined as tumor infiltration beyond the pancreatic parenchyma, extracted from SEER variables "CS extension (2004–2015)" and "EOD Primary Tumor (2018 +)". Treatment variables included surgery type (partial pancreatectomy, Whipple procedure, or total/extended pancreatectomy), chemotherapy receipt (yes or no, determined from the SEER “Chemotherapy recode” variable, which records whether chemotherapy was administered as part of the first course of treatment following diagnosis), radiotherapy administration (no radiotherapy, beam radiation, or other radiation modalities), and receipt of radiotherapy (yes or no). Notably, the SEER database records chemotherapy as a binary indicator and does not provide information on treatment timing relative to surgery (neoadjuvant vs. adjuvant), specific chemotherapy regimens, number of cycles, or treatment completion rates. The primary outcome was disease-specific survival (DSS), defined as the interval from diagnosis to death attributed to pancreatic cancer or last follow-up.

Statistical analysis

To identify the most influential prognostic factors and overcome limitations of traditional regression approaches in handling complex interactions and multicollinearity, we employed machine learning-based variable importance assessment using Random Survival Forest (RSF) models^13,14. Variable importance was quantified using two complementary methods: (1) SHAP (SHapley Additive exPlanations) analysis^13,15, which computed mean absolute SHAP values to measure each variable’s average contribution to survival predictions across all possible feature combinations, thereby capturing both main effects and higher-order interactions; and (2) permutation importance¹³, which evaluated each variable’s contribution by measuring the decrease in model predictive performance (concordance index) after random permutation of that variable, providing an unbiased estimate of variable importance independent of model structure.

Univariable and multivariable Cox models were constructed to estimate hazard ratios (HRs) with 95% confidence intervals (CIs) for associations between clinical variables and disease-specific survival (DSS). To investigate whether EPE status modifies the therapeutic effect of chemotherapy on DSS, Survival differences between EPE and chemotherapy groups were visualized using Kaplan–Meier curves for unadjusted estimates and direct adjustment methods for covariate-adjusted estimates^16,17. Interaction effects between EPE and chemotherapy were examined on multiplicative and additive scales^18,19. Multiplicative interaction assesses whether the combined effect differs from the product of individual effects, which is calculated as HR₁₁/(HR₁₀ × HR₀₁). Additive interaction, measured by the Relative Excess Risk due to Interaction (RERI = HR₁₁-HR₁₀-HR₀₁ + 1), quantifies whether the combined effect exceeds the sum of individual effects.

Recognizing the prognostic significance of tumor size in pancreatic cancer and its potential interaction with EPE and chemotherapy⁶, we performed stratified interaction analyses across three tumor size categories: 20 mm or less, greater than 20 mm to 40 mm, and greater than 40 mm. To evaluate the robustness of our primary findings and account for competing risks of death from non-pancreatic cancer causes, we conducted sensitivity analyses using Fine-Gray competing risk regression models. To address potential confounding from temporal changes in chemotherapy regimens, particularly following the PRODIGE 4 trial in 2011 that led to increased adoption of FOLFIRINOX, we performed sensitivity analyses stratified by treatment era (before 2011 vs. 2011 onwards), examining the EPE-chemotherapy interaction separately within each era using both Cox and Fine-Gray models. To assess potential collinearity between EPE and tumor size, we calculated Cramér’s V statistic for their categorical association and examined generalized variance inflation factors (GVIF) for all covariates in the multivariable Cox model.

Baseline characteristics were compared between chemotherapy groups using χ² tests for categorical variables and Mann–Whitney U tests for continuous variables. Data were presented as frequencies (percentages) for categorical variables and medians (interquartile ranges) for continuous variables. All analyses utilized R version 4.4.3. Statistical significance was defined as a two-sided p-value < 0.05. Significance levels are denoted as follows: ***p < 0.001, **p < 0.01, and *p < 0.05.

Results

Patient characteristics

A total of 22,045 patients with resected non-metastatic pancreatic ductal adenocarcinoma were identified from the SEER database spanning the years 2004 through 2022 (Supplementary Fig. 1). Of these, 16,682 patients (75.7%) received chemotherapy, while 5,363 (24.3%) did not (Table 1). Patients receiving chemotherapy were significantly younger (median 67 vs 72 years, p < 0.001), more likely to be married (66.0% vs 54.5%, p < 0.001), and had higher household income (38.5% vs 32.6% in high-income category, p < 0.001). Extrapancreatic extension (EPE) was present in 78.0% of patients overall (78.9% chemotherapy vs 75.3% no-chemotherapy, p < 0.001). Most tumors were located in the pancreatic head (73.7%) and measured 20-40 mm (61.7%). Well to moderately differentiated tumors (Grade I/II) comprised 55.4% of cases but were less common in the chemotherapy group (53.0% vs 62.8%, p < 0.001). Lymph node metastases were present in 63.7% of patients, with higher rates of N1 (39.7% vs 36.7%) and N2 disease (25.3% vs 22.9%) in the chemotherapy group (p < 0.001). Beam radiation therapy (35.8% vs 4.2%, p < 0.001) was predominantly utilized in the chemotherapy group. After a median follow-up of 20 months, disease-specific mortality occurred in 66.1% of patients (73.6% no-chemotherapy vs 63.7% chemotherapy, p < 0.001).

Table 1.

Patient characteristics stratified by chemotherapy in pancreatic ductal adenocarcinoma.

Variable	Overall N = 22,045a	No N = 5,363a	Yes N = 16,682a	p-valueb
Age (year)	68.00 (60.00, 75.00)	72.00 (64.00, 79.00)	67.00 (59.00, 73.00)	< 0.001
Sex				0.018
Female	10,954 (49.7%)	2,740 (51.1%)	8,214 (49.2%)
Male	11,091 (50.3%)	2,623 (48.9%)	8,468 (50.8%)
Race				0.736
White	18,156 (82.4%)	4,403 (82.1%)	13,753 (82.4%)
Black	2,018 (9.2%)	491 (9.2%)	1,527 (9.2%)
Other/Unknown	1,871 (8.5%)	469 (8.7%)	1,402 (8.4%)
Marital status				< 0.001
Married	13,924 (63.2%)	2,921 (54.5%)	11,003 (66.0%)
Single	2,725 (12.4%)	747 (13.9%)	1,978 (11.9%)
Divorced/Widowed/Separated	4,852 (22.0%)	1,534 (28.6%)	3,318 (19.9%)
Unknown	544 (2.5%)	161 (3.0%)	383 (2.3%)
Household income				< 0.001
Low	5,334 (24.2%)	1,386 (25.8%)	3,948 (23.7%)
Middle	8,544 (38.8%)	2,226 (41.5%)	6,318 (37.9%)
High	8,167 (37.0%)	1,751 (32.6%)	6,416 (38.5%)
County type				0.808
Urban	19,641 (89.1%)	4,783 (89.2%)	14,858 (89.1%)
Rural	2,404 (10.9%)	580 (10.8%)	1,824 (10.9%)
Year of diagnosis				< 0.001
2004–2010	6,106 (27.7%)	2,096 (39.1%)	4,010 (24.0%)
2011–2017	8,497 (38.5%)	2,038 (38.0%)	6,459 (38.7%)
2018–2022	7,442 (33.8%)	1,229 (22.9%)	6,213 (37.2%)
Primary tumor site				< 0.001
Head	16,246 (73.7%)	3,799 (70.8%)	12,447 (74.6%)
Body	1,716 (7.8%)	422 (7.9%)	1,294 (7.8%)
Tail	2,285 (10.4%)	593 (11.1%)	1,692 (10.1%)
Others(overlapping/unknown)	1,798 (8.2%)	549 (10.2%)	1,249 (7.5%)
Grade				< 0.001
Grade I/II	12,214 (55.4%)	3,368 (62.8%)	8,846 (53.0%)
Grade III/IV	6,491 (29.4%)	1,778 (33.2%)	4,713 (28.3%)
Unknown	3,340 (15.2%)	217 (4.0%)	3,123 (18.7%)
Tumor size category (mm)				< 0.001
(0,20]	3,573 (16.2%)	1,050 (19.6%)	2,523 (15.1%)
(20,40]	13,599 (61.7%)	3,068 (57.2%)	10,531 (63.1%)
(40,90 +]	4,873 (22.1%)	1,245 (23.2%)	3,628 (21.7%)
Extension				< 0.001
No EPE	4,850 (22.0%)	1,326 (24.7%)	3,524 (21.1%)
EPE	17,195 (78.0%)	4,037 (75.3%)	13,158 (78.9%)
N stage				< 0.001
N0	7,997 (36.3%)	2,164 (40.4%)	5,833 (35.0%)
N1	8,595 (39.0%)	1,970 (36.7%)	6,625 (39.7%)
N2	5,453 (24.7%)	1,229 (22.9%)	4,224 (25.3%)
Surgery				< 0.001
Partial pancreatectomy	3,030 (13.7%)	797 (14.9%)	2,233 (13.4%)
Whipple procedure	15,578 (70.7%)	3,643 (67.9%)	11,935 (71.5%)
Total or extended pancreatectomy	3,437 (15.6%)	923 (17.2%)	2,514 (15.1%)
Radiotherapy				< 0.001
No	15,843 (71.9%)	5,141 (95.9%)	10,702 (64.2%)
Beam	6,121 (27.8%)	213 (4.0%)	5,908 (35.4%)
Other Radiation	81 (0.4%)	9 (0.2%)	72 (0.4%)
Follow-up (months)	20.00 (10.00, 37.00)	13.00 (6.00, 27.00)	21.00 (12.00, 39.00)	< 0.001
Cause of deaths				< 0.001
Alive	5,758 (26.1%)	774 (14.4%)	4,984 (29.9%)
Pancreatic cancer	14,575 (66.1%)	3,949 (73.6%)	10,626 (63.7%)
Other causes	1,712 (7.8%)	640 (11.9%)	1,072 (6.4%)

^aMedian (Q1, Q3); n (%).

^bWilcoxon rank sum test; Pearson’s Chi-squared test.

Variable importance and prognostic impact

SHAP analysis identified chemotherapy and EPE as variables with the highest predictive contributions in the RSF model (Fig. 1). The N stage showed the highest contribution (28.8%), followed by chemotherapy (17.2%) and EPE (12.0%). The permutation importance method yielded consistent rankings, with chemotherapy emerging as the variable with the largest predictive contribution (30.8%), followed by N stage (24.7%) and EPE (11.7%). Notably, variable importance rankings derived from the RSF model reflect each variable’s contribution to predictive accuracy and should not be interpreted as measures of causal or clinical relevance, as they may be influenced by data quality, variable coding consistency, and correlations among variables in the retrospective dataset.

Fig. 1.

Variable importance ranking for disease-specific survival prediction in resected pancreatic ductal adenocarcinoma. SHAP (SHapley Additive exPlanations) importance (left panel) and permutation importance (right panel) derived from random survival forest models. Variables are ranked by their percentage contribution to predictive performance.

Cox regression analyses validated these findings (Fig. 2). Absence of chemotherapy increased disease-specific mortality by 45% (adjusted HR 1.45, 95% CI 1.39–1.52, p < 0.001), while EPE conferred a 25% increased risk (adjusted HR 1.25, 95% CI 1.19–1.31, p < 0.001). Lymph node involvement demonstrated a dose-dependent mortality risk (N1: adjusted HR 1.48, 95% CI 1.42–1.54; N2: adjusted HR 1.92, 95% CI 1.84–2.01; both p < 0.001). Larger tumor size showed graded effects relative to tumors ≤ 20 mm (20-40 mm: adjusted HR 1.40, 95% CI 1.33–1.47; > 40 mm: adjusted HR 1.67, 95% CI 1.57–1.77, both p < 0.001), as did higher tumor grade (Grade III/IV: adjusted HR 1.36, 95% CI 1.31–1.41). Radiotherapy (adjusted HR 0.96, 95% CI 0.92–0.99) and high household income (adjusted HR 0.86, 95% CI 0.82–0.90) demonstrated protective effects. Sensitivity analyses using Fine-Gray competing risk models confirmed these associations, with comparable effect sizes (Supplementary Fig. 2). Notably, collinearity assessment between EPE and tumor size category yielded a Cramér’s V of 0.204, indicating a weak-to-moderate association. Generalized variance inflation factors in the multivariable Cox model were 1.12 for both EPE and tumor size, with no variable exceeding 1.88, confirming minimal multicollinearity across all covariates.

Fig. 2.

Univariable and multivariable Cox regression analysis for disease-specific survival.

Interaction effect between EPE and chemotherapy

Survival analyses revealed distinct prognostic patterns across the four EPE and chemotherapy groups (Fig. 3). Using No EPE with Chemotherapy (Yes) as the reference, adjusted analyses (Fig. 3B) demonstrated a hierarchical mortality risk: No EPE with Chemotherapy (No) (adjusted HR 1.34, 95% CI 1.23–1.46), EPE with Chemotherapy (Yes) (adjusted HR 1.17, 95% CI 1.11–1.24), and EPE with Chemotherapy (No) showing the poorest outcomes (adjusted HR 2.15, 95% CI 2.02–2.28; all p < 0.001). Within the EPE stratum, the absence of chemotherapy was associated with an 83% increased mortality risk (Table 2, adjusted HR 1.83, 95% CI 1.75–1.91) compared to a 34% increase in non-EPE patients (adjusted HR 1.34, 95% CI 1.23–1.46). EPE patients receiving chemotherapy achieved adjusted 3-year survival rates comparable to non-EPE patients without chemotherapy (39.5% vs 35.0%), suggesting that chemotherapy receipt may partially offset EPE’s adverse prognostic impact. Competing risk analyses corroborated these findings, with adjusted cumulative incidence curves showing 3-year death rates of 59.9% for No EPE without chemotherapy versus 58.3% for EPE with chemotherapy.

Fig. 3.

Survival curves for different EPE and chemotherapy groups based on Cox model. (A) Unadjusted survival curves and (B) adjusted survival curves showing disease-specific survival across four patient groups stratified by extrapancreatic extension (EPE) status and chemotherapy receipt. Adjusted analyses controlled for age, sex, race, marital status, household income, county type, year of diagnosis, primary tumor site, grade, tumor size, N stage, surgery, and radiotherapy.

Table 2.

Interaction analysis between different EPE and chemotherapy groups based on Cox model.

	Unadjusted Cox model			Adjusted Cox model
	Chemotherapy: Yes	Chemotherapy: No	Effect of chemotherapy within the strata of EPE	Chemotherapy: Yes	Chemotherapy: No	Effect of chemotherapy within the strata of EPE
	HR [95% CI]	HR [95% CI]	HR [95% CI]	HR [95% CI]	HR [95% CI]	HR [95% CI]
No EPE	1.00 (Reference)	1.27 (1.16, 1.38)***	1.27 (1.16, 1.38)***	1.00 (Reference)	1.34 (1.23, 1.46)***	1.34 (1.23, 1.46)***
EPE	1.53 (1.46, 1.61)***	2.85 (2.68, 3.02)***	1.86 (1.78, 1.93)***	1.17 (1.11, 1.24)***	2.15 (2.02, 2.28)***	1.83 (1.75, 1.91)***
Effect of EPE within the strata of chemotherapy	1.53 (1.46, 1.61)***	2.25 (2.08, 2.43)***		1.17 (1.11, 1.24)***	1.60 (1.48, 1.74)***
Multiplicative scale	1.47 (1.34, 1.61)***			1.37 (1.24, 1.50)***
RERI	1.05 (0.91, 1.19)***			0.63 (0.51, 0.76)***

Multivariable Cox model adjusted for age, sex, race, marital status, household income, county type, year of diagnosis, primary tumor site, grade, tumor size, N stage, surgery and radiotherapy.

RERI, relative excess risk due to interaction.

Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05.

Formal interaction tests confirmed a statistically significant interaction between chemotherapy and EPE status. The adjusted multiplicative interaction of 1.37 (95% CI 1.24–1.50, p < 0.001) indicated that the detrimental effect of chemotherapy absence was 37% greater in EPE-positive disease compared to EPE-negative disease. The adjusted additive interaction (RERI 0.63, 95% CI 0.51–0.76, p < 0.001) revealed substantial excess risk from the combined absence of chemotherapy and presence of EPE beyond their individual contributions. These findings remained robust in adjusted competing risk analyses (Supplementary Table 1: multiplicative interaction 1.26, 95% CI 1.14–1.39; RERI 0.41, 95% CI 0.28–0.54), consistently suggesting a greater relative benefit of chemotherapy in EPE-positive PDAC.

Sensitivity analysis

Stratified analyses across tumor size categories (≤ 20 mm, 20-40 mm, > 40 mm) revealed consistent EPE-chemotherapy interaction patterns (Fig. 4). In brief, the absence of chemotherapy exerted a disproportionately greater detrimental effect on EPE patients compared to non-EPE patients across all size strata. Additionally, EPE patients receiving chemotherapy demonstrated superior outcomes compared to non-EPE patients without chemotherapy across all size strata. Specifically in tumors ≤ 20 mm, adjusted analyses showed that EPE with chemotherapy (Yes) achieved comparable adjusted 3-year survival to No EPE without chemotherapy (53.9% vs 54.9%; adjusted HR 1.15 vs 1.12). This pattern intensified in intermediate tumors (20-40 mm), where EPE with chemotherapy substantially outperformed No EPE without chemotherapy (38.4% vs 31.4%; adjusted HR 1.20 vs 1.48). Even in large tumors (> 40 mm), chemotherapy maintained its protective effect, with EPE with Chemotherapy (Yes) patients demonstrating better survival than no EPE with chemotherapy (No) (31.8% vs 23.0%, adjusted HR 1.14 vs 1.50). Formal interaction testing revealed significant positive multiplicative and additive interactions across all tumor size categories (Supplementary Table 2). Competing risk analyses validated these findings (Supplementary Table 3 and Supplementary Fig. 4).

Fig. 4.

Survival curves for different EPE and chemotherapy groups stratified by tumor size based on Cox model. Left panel: unadjusted survival curves. Right panel: adjusted survival curves. Adjusted analyses controlled for age, sex, race, marital status, household income, county type, year of diagnosis, primary tumor site, grade, N stage, surgery, and radiotherapy.

To assess the robustness of the observed EPE-chemotherapy interaction to temporal changes in chemotherapy regimens, we performed sensitivity analyses stratified by the 2011 time point corresponding to the PRODIGE 4 trial. In the adjusted Cox model, significant positive interactions were observed in both treatment eras: before 2011 (multiplicative interaction 1.22, 95% CI 1.06–1.41, p < 0.01; RERI 0.40, 95% CI 0.22–0.58, p < 0.001) and after 2011 (multiplicative interaction 1.46, 95% CI 1.28–1.66, p < 0.001; RERI 0.84, 95% CI 0.66–1.03, p < 0.001). Fine-Gray competing risk models corroborated these findings in both eras (Supplementary Tables 4–5). The persistence of significant interactions across treatment eras indicates that the observed effect modification is robust to the major shift in prescribing practice following the PRODIGE 4 trial.

Discussion

This comprehensive population-based analysis represents the first systematic evaluation of interaction effects between extrapancreatic extension and chemotherapy in resected pancreatic ductal adenocarcinoma (PDAC). Our findings identify EPE as a significant adverse prognostic factor and demonstrate a statistically significant interaction between EPE status and chemotherapy on both multiplicative and additive scales, suggesting that the magnitude of mortality reduction associated with chemotherapy receipt may differ by EPE status. Importantly, non-EPE patients who received chemotherapy demonstrated the best overall survival, consistent with EPE being an adverse prognostic factor. Despite this, the interaction analyses suggest a greater relative mortality reduction associated with chemotherapy receipt among EPE-positive patients compared to EPE-negative patients. These observations, while hypothesis-generating, highlight the importance of EPE as a prognostic factor that warrants consideration in the clinical assessment of resected PDAC patients. Importantly, these findings should not be interpreted as suggesting any alteration to the universal recommendation of adjuvant chemotherapy in resected PDAC. Rather, given that 24.3% of surgically resected patients in our cohort did not receive chemotherapy—consistent with population-based studies reporting non-receipt rates of approximately 30–40% due to advanced age, postoperative complications, or patient refusal ²⁰—our results underscore the particular importance of ensuring chemotherapy receipt in EPE-positive patients, who may have the most to lose from treatment omission.

The difference in HR magnitude for chemotherapy absence between the EPE-positive and EPE-negative strata should be interpreted with caution, as hazard ratios are relative measures that may be influenced by differences in baseline risk, sample size, and event rate distribution across strata. Formal interaction tests on both multiplicative and additive scales provide more appropriate statistical evidence for effect modification. Notably, EPE-positive patients receiving chemotherapy achieved an adjusted 3-year disease-specific survival comparable to EPE-negative patients without chemotherapy, suggesting that chemotherapy receipt may partially offset the adverse prognostic impact of EPE. This observation underscores the importance of ensuring chemotherapy receipt in EPE-positive patients rather than indicating a differential treatment strategy.

The consistency of this interaction across tumor size strata reinforces its clinical utility across diverse patient populations. Our finding in small tumors (≤ 20 mm), where EPE patients receiving chemotherapy achieved adjusted 3-year disease-specific survival rates comparable to non-EPE patients who did not receive chemotherapy (53.9% vs 54.9%), suggesting that chemotherapy receipt may partially offset the prognostic disadvantage conferred by extrapancreatic extension. More strikingly, as tumor size increased, the survival advantage of EPE patients receiving chemotherapy over non-EPE patients without chemotherapy became progressively more pronounced. In intermediate-sized tumors (20-40 mm), EPE patients receiving chemotherapy demonstrated superior adjusted 3-year disease-specific survival compared to non-EPE patients who did not receive chemotherapy (38.4% vs 31.4%), and this difference further widened in (> 40 mm, 31.8% vs 23.0%). This size-dependent pattern is consistent with the hypothesis that chemotherapy may partially offset the adverse impact of EPE across tumor sizes, although the underlying mechanisms warrant further investigation. This consistency across tumor size categories further supports the robustness of the observed EPE-chemotherapy interaction and reinforces the prognostic importance of EPE across the spectrum of tumor sizes in resected PDAC.

Our results provide critical evidence to the ongoing debate surrounding the AJCC 8th edition’s removal of EPE from the T-staging criteria. The recent literature increasingly questions this decision, with multiple validation studies demonstrating EPE’s persistent prognostic significance ^9–11 despite inter-observer variability concerns ²¹. Importantly, the observed interaction between EPE and chemotherapy, if confirmed in prospective studies, provides additional support for EPE’s clinical relevance beyond its established role as a prognostic factor. Notably, the Japan Pancreas Society maintains EPE-based staging in their 8th edition²², reflecting international disagreement about optimal classification approaches. The observed EPE-chemotherapy interaction, if validated prospectively, could provide additional justification for reconsidering the role of EPE in clinical decision-making frameworks and staging systems.

Several biological mechanisms may contribute to the observed interaction between EPE and chemotherapy benefit. While EPE tumors frequently exhibit aggressive basal-like transcriptional signatures, recent evidence demonstrates that 17–43% harbor DNA damage response defects, particularly homologous recombination deficiency, rendering them exceptionally sensitive to platinum-based chemotherapy with progression-free survival extending from 12.2 to 44.1 months in responsive patients^23–25. In addition, the tumor microenvironment in EPE disease may paradoxically facilitate drug delivery through invasion-associated matrix degradation and increased vascular permeability, while the inflammatory milieu disrupts the dense stromal barriers that typically limit chemotherapy penetration in pancreatic cancer^26–28. Although speculative, this convergence of potential molecular vulnerabilities and microenvironmental alterations provides a plausible biological rationale for the observed EPE-chemotherapy interaction in our population-based analysis; direct mechanistic validation is warranted.

Several methodological strengths support our investigation, including a large sample size providing adequate statistical power for interaction detection; a population-based design reducing selection bias; and complementary analytical approaches (Cox and Fine-Gray models) that ensured consistency across different statistical assumptions. Nevertheless, several limitations warrant consideration. First, the SEER database records chemotherapy as a binary variable (yes/no) without information on treatment timing (neoadjuvant vs. adjuvant), specific regimens (e.g., FOLFIRINOX, gemcitabine-based, or gemcitabine plus capecitabine), number of cycles, or completion rates. This limitation precludes the analysis of regimen-specific interactions and may introduce heterogeneity in the observed chemotherapy effects. Although surgical resection was a mandatory inclusion criterion and the majority of chemotherapy recipients likely received adjuvant treatment, we cannot exclude a proportion of neoadjuvant recipients, particularly in later study years. Second, treatment protocols evolved substantially during the study period, from gemcitabine monotherapy to modern combination regimens including modified FOLFIRINOX following the PRODIGE-24 trial. Although our multivariable model adjusted for the year of diagnosis (categorized as 2004–2010, 2011–2017, and 2018–2022 to reflect these treatment eras), this adjustment serves as a surrogate and cannot fully account for regimen-specific effects. Sensitivity analyses stratified by treatment era (before vs. after 2011) demonstrated consistent EPE-chemotherapy interactions in both periods (Supplementary Tables 4–5), supporting the robustness of our findings to temporal changes in prescribing patterns; however, this era-based approach remains an indirect proxy for specific chemotherapy regimens. Third, the retrospective nature of SEER data precludes the assessment of chemotherapy dose intensity, dose modifications, and treatment adherence, which may influence the magnitude of observed interactions. Fourth, while we adjusted for key prognostic factors, unmeasured confounders such as performance status, surgical margin status, and comorbidities may partially account for treatment selection patterns. It should be noted that EPE and margin status represent distinct pathological concepts: EPE reflects the biological behavior of the tumor — direct infiltration beyond the pancreatic parenchyma — whereas margin status reflects the surgical adequacy of resection^7,8. The two are not interchangeable, as an EPE-positive tumor can achieve R0 resection, while an EPE-negative tumor may have positive margins. Future studies should incorporate detailed treatment protocols and standardized EPE assessment criteria to refine EPE-stratified treatment algorithms.

Conclusion

This population-based analysis confirms EPE as an independent adverse prognostic factor in resected pancreatic ductal adenocarcinoma and identifies a significant EPE–chemotherapy interaction on both multiplicative and additive scales across tumor size categories. Despite conferring worse prognosis, EPE-positive patients receiving chemotherapy achieved survival comparable to that of EPE-negative patients without chemotherapy, underscoring the importance of ensuring receipt of chemotherapy in EPE-positive disease. These findings support the need to re-evaluate the role of EPE in staging systems; however, prospective validation is warranted.

Abbreviations

PDAC

Pancreatic ductal adenocarcinoma

EPE

Extrapancreatic extension

AJCC

American Joint Committee on Cancer

DSS

Disease-specific survival

SEER

Surveillance, Epidemiology, and End Results

ICD-O-3

International Classification of Diseases for Oncology, Third Edition

RSF

Random Survival Forest

SHAP

SHapley Additive exPlanations

HR

Hazard ratio

CI

Confidence interval

RERI

Relative Excess Risk due to Interaction

Author contributions

Shuai Liang (SL) contributed to the conception and design of the study, performed data acquisition, and drafted the manuscript. Xiaolin Dou (XD), Xuejun Gong (XG), and Hongqin Jiang (HJ) contributed to data collection, analysis, and interpretation, and assisted in revising the manuscript. Wei Wei (WW) supervised the study, provided critical intellectual input, and was a major contributor in writing and revising the manuscript. SL and WW confirm the authenticity of all the raw data. All authors read and approved the final manuscript.

Data availability

The datasets used and/or analyzed for the present study are available from the corresponding author (Wei Wei,weiweicsu@outlook.com) on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

Since the SEER database is publicly available, this study was exempt from review by the Ethics Committee of the authors’ hospital.

Consent to participate

Written informed consent was waived due to the use of de-identified data from the SEER public database.