Optimizing postoperative prognostication in duodenal adenocarcinoma: a multicenter study with external validation

Abstract

Objective

To develop and externally validate a prognostic nomogram for overall survival (OS) in resected duodenal adenocarcinoma (DA) using routinely available perioperative variables, thereby clarifying risk profiles and supporting clinical management.

Methods

Multicenter analysis of 2289 consecutive DA patients undergoing curative surgery (2012-2022) from China’s National Cancer Center database. External validation used 335 patients from Zhejiang Provincial People’s Hospital (2022-2024). LASSO-Cox regression selected variables from 89 perioperative factors to construct the nomogram, with web tool implementation.

Results

The LASSO-Cox model achieved 1-, 3-, and 5-year AUCs of 0.72 (95% CI, 0.68-0.77), 0.75 (95% CI, 0.72-0.77), and 0.76 (95% CI, 0.73-0.79), outperforming traditional Cox models (P < .01). External validation yielded AUCs of 0.76 (95% CI, 0.66-0.86) and 0.79 (95% CI, 0.74-0.86) for 1- and 3-year OS, and 0.81 (95% CI, 0.74-0.89) for estimated 5-year OS. The model stratified patients into low- and high-risk groups (cutoff 0.40), with low-risk patients showing superior survival. Eight predictors were selected, including modifiable surgical factors such as transfusion and operative time.

Conclusions

We developed and externally validated a postoperative prognostic nomogram for DA using routinely available perioperative variables. In the present study, the model improved postoperative risk stratification and may support counseling, follow-up planning, and multidisciplinary discussion regarding adjuvant therapy; however, it should be viewed as complementary to standard staging and clinical judgment, and broader clinical implementation will require further validation.

Implications for Practice.

This postoperative prognostic tool may complement routine clinicopathological assessment after DA resection. It is intended to support postoperative counseling, follow-up planning, and multidisciplinary discussion, rather than replace AJCC/pathological staging, clinician judgment, multidisciplinary review, or evidence-based treatment guidelines.

Introduction

Primary duodenal malignancies are rare, comprising approximately 0.3%-0.5% of gastrointestinal cancers.¹ Duodenal adenocarcinoma (DA), the predominant histologic subtype (52%-57% of cases), arises from the intestinal mucosa, most commonly in the periampullary region of the descending duodenum, with a rising incidence attributed to aging populations and enhanced diagnostics.^2–4 DAs are classified by origin as ampullary DA (a-DA) or nonampullary DA (Na-DA).⁵ Although these subtypes differ in molecular and immunohistochemical profiles, emerging evidence, including from our cohort, suggests comparable overall survival (OS),⁶ justifying pooled analyses for prognostic modeling. Despite surgical resection offering the best chance for cure, the 5-year OS for resectable DA remains suboptimal at 40%-60%, outperforming pancreatic head adenocarcinoma but aligning with other periampullary malignancies like distal cholangiocarcinoma.^1,7,8 This underscores the urgent need for refined prognostic tools to guide personalized therapy amid diagnostic delays from nonspecific symptoms and surgical complexities such as pancreatoduodenectomy.

Due to DA’s rarity, prognostic risk factors remain incompletely defined, with most studies focusing on a-DA subtypes and neglecting Na-DA cases.⁹ Established predictors of poor OS include non-R0 resection,^10,11 postoperative complications,¹² perineural invasion,¹³ lymph node metastasis,^14,15 microvascular invasion,^8,13 tumor size, and poor differentiation. However, accurate predictive models are scarce. Existing ones often rely on univariate analyses followed by Cox regression, overlooking multivariate interactions and risking bias per PROBAST guidelines.¹⁶ Moreover, they typically stem from small, single-center cohorts without external validation, limiting generalizability and clinical utility.

Machine learning approaches, such as least absolute shrinkage and selection operator (LASSO) regression, have gained traction for high-dimensional data by mitigating multicollinearity and overfitting,^9,17 with our prior work demonstrating enhanced performance in pancreatic surgery models.¹⁸ Building on this, we developed a LASSO-Cox-based nomogram and interactive web tool to predict surgical outcomes in DA patients, using multicenter data from China’s National Cancer Center (NCC). External validation in an independent cohort from Zhejiang Provincial People’s Hospital (ZJPPH) confirms its generalizability, advancing risk stratification, personalized treatment, and preventive strategies for this challenging malignancy.

Patients and methods

Study design

This multicenter cohort study developed and validated a prognostic nomogram for DA using LASSO-Cox regression. The primary nomogram was designed for postoperative prognostic assessment during early postoperative management. Candidate predictors were therefore limited to variables available at the intended time of model application. We followed TRIPOD guidelines for prediction model design, validation, and reporting,¹⁶ and reported data per STROCSS 2021 criteria.¹⁹ The study was registered with the Chinese Clinical Trial Registry (ChiCTR ID: ChiCTR2500099111), a WHO ICTRP Primary Registry. Both the training and validation cohorts were analyzed retrospectively under approval of the National Cancer Center ethics committee (No. 22/213-3415) as the lead institution of the multicenter study, with local oversight and consent waiver procedures following institutional requirements.

Study cohorts

The training cohort included 2289 consecutive DA patients who underwent curative-intent surgery (2012-2022) from 16 centers in China’s NCC database. This retrospective database aggregates quality-controlled perioperative data audited centrally. For external validation, we used an independent cohort of 335 DA patients treated at ZJPPH in China (2022-2024), a multi-campus system with 8 affiliates. Inclusion required histologically confirmed DA and radical resection (eg, pancreatoduodenectomy). We excluded cases without radical surgery, non-adenocarcinoma histology, or incomplete follow-up. The study flowchart is shown in Figure 1.

Figure 1.

Study flowchart. AJCC, American Joint Committee on Cancer (8th edition); DA, duodenal adenocarcinoma.

Indicators and definitions

After excluding duplicate or redundant codes, nonclinical administrative fields, and variables with >10% missing data, 90 clinically relevant perioperative variables were retained for analysis (49 preoperative, 5 intraoperative, 36 postoperative; Figure S1). Adjuvant therapy data (chemotherapy and/or radiotherapy) had >75% nonrandom missingness linked to geographic variation, treatment era, and evolving guideline recommendations, and were therefore excluded from the primary model; an exploratory analysis of adjuvant therapy was conducted separately in the external validation cohort (n = 309). For those with <10% missing data, we applied multiple imputation following Rubin’s rules for pooled estimates to minimize bias. The primary outcome was OS from surgery to death or last follow-up. Tumor recurrence was defined as radiologically, pathologically, or multidisciplinary-confirmed local, regional, or distant relapse after curative-intent resection. As a post-baseline event, recurrence was excluded from the primary nomogram and analyzed separately as a time-dependent covariate in secondary follow-up analyses. Age was recorded in years, operative time in minutes, and intraoperative transfusion as volume (mL) without unit conversion. Aspartate aminotransferase was measured in IU/L, with the upper limit of normal standardized as 40 IU/L across centers. Progression-free survival (PFS) was defined as the time from surgery to progression or death, and was assessed only in DA subgroup (a-DA/Na-DA) comparisons. For consistency, Na-DAs were classified as intestinal or pancreatobiliary/mixed phenotypes according to predominant histologic differentiation, with immunohistochemical markers (CK7, CK20, CDX2, MUC1) applied when necessary (4). Postoperative pancreatic fistula followed 2016 International Study Group of Pancreatic Surgery (ISGPS) guidelines,²⁰ including only clinically relevant Grades B/C. Delayed gastric emptying and postpancreatectomy hemorrhage used 2007 ISGPS definitions,^21,22 excluding Grade A. AJCC staging followed the 8th edition.²³

Statistical analysis

We used R software (version 4.2.3) for analyses. Continuous variables were reported as mean ± standard deviation and analyzed without dichotomization to preserve nonlinear relationships, per PROBAST.¹⁶ Categorical variables were percentages. We assessed baseline differences with the compareGroups package and confirmed sample size adequacy via pmsampsize for model performance. To select variables and address multicollinearity, we applied LASSO regression with 10-fold cross-validation, shrinking less important coefficients to zero. Selected variables entered Cox proportional hazards regression to build the nomogram. Bonferroni correction was omitted during modeling. LASSO-Cox includes built-in penalization, and analyses focused on a single primary survival outcome. Additional multiplicity adjustment would overpenalize continuous effects and increase type II errors.²⁴ Risk stratification used the Youden index on 5-year OS. Model performance metrics included concordance index (C-index), area under the receiver operating characteristic (ROC) curve (AUC), integrated discrimination improvement (IDI), net reclassification improvement (NRI), Akaike information criterion (AIC), decision curve analysis, and calibration curves. We compared the LASSO-Cox model to traditional Cox models (univariate, full-variable, AJCC-based) via time-dependent AUC. In the external validation cohort, exploratory analysis was performed among 309 patients with complete adjuvant therapy data, using multivariable Cox regression within high- and low-risk subgroups and testing interaction by adding a risk group × adjuvant therapy term; as therapy was nonrandomly assigned, results are hypothesis-generating. A 2-sided P < .05 indicated significance. We used DynNom package for an interactive web tool (available at https://zongtingteamcode.shinyapps.io/dynamic-da-os/).

Results

Characteristics of the study cohort

We used multiple imputation for variables with <10% missing data. This resulted in clean, complete datasets. The quality of data before and after imputation showed strong agreement (Figure S2). The training cohort included 2289 DA patients. We first divided them into a-DA and Na-DA groups. However, analyses revealed no significant differences in OS (hazard ratio [HR] 1.03, 95% CI, 0.90-1.19; P = .64) or PFS (HR 1.03, 95% CI, 0.91-1.16; P = .65) between groups (Figure 2A; Figure S3A). Thus, we pooled all patients for further analysis (Figure 2B). The training cohort had 1337 males and 952 females. Their mean age was 61.0 years, and mean body mass index (BMI) was 23.2 kg/m². The validation cohort included 335 patients: 189 males and 146 females. Their mean age was 59.6 years, and mean BMI was 23.1 kg/m². Both cohorts met sample size requirements for model development. Follow-up was from surgery until May 31, 2024 in the training cohort and January 31, 2026 in the validation cohort, or death. Median follow-up was 42 months (range, 5-146) and 36 months (range, 13-49), respectively. Median survival was 72 months in the training cohort, while not reached in the validation cohort. In the training cohort, 1-, 3-, and 5-year survival rates were 91.7%, 68.8%, and 54.0%. In the validation cohort, 1- and 3-year survival rates were 90.9% and 75.4%, while the 5-year rate was an estimated 58.1% given the maximum follow-up of 49 months. In the training cohort, 67.9% of patients had laparotomy, 18.7% laparoscopic surgery, and 13.4% robotic surgery. The validation cohort had similar rates: 64.8%, 21.5%, and 13.7%. Preoperative biliary drainage occurred in 30.6% of training cohort patients and 32.2% of validation cohort patients. Intestinal type, the predominant histologic subtype, comprised 50.0% of patients in the training cohort and 54.9% in the validation cohort, with the remainder classified as pancreatobiliary/mixed type. Baseline characteristics matched well between cohorts (Table S1).

Figure 2.

Overall prognosis and prognostic factor selection for duodenal adenocarcinoma (DA). (A) Overall survival (OS) curves for ampullary DA (a-DA) and non-ampullary DA (Na-DA). (B) OS and progression-free survival (PFS) curves for all DA patients (log-rank test, P > .05). (C) Coefficient profiles vs log(λ) in LASSO regression. (D) Area under the curve (AUC) vs log(λ). λ.min denotes the minimum MSE cutoff; λ.1SE denotes the cutoff at 1 standard error from the minimum.

Development of the novel predictive model

LASSO regression for variable selection

We applied LASSO regression to screen variables. This addressed multicollinearity and overfitting. The final nomogram included only variables available at the intended time of postoperative model application; tumor recurrence was not included as a predictor and was instead evaluated in separate follow-up analyses. After 10-fold cross-validation, we reduced variables to 34 at the minimum mean squared error (log(λ); Table S2). At one standard error of the minimum distance (1-fold SE), we selected 10 variables (best λ = 0.046; Figure 2C; Figure S3B; Table S3). We chose this simpler model due to its high C-index and interpretability (Figure 2D). The 10 variables were age, intraoperative blood transfusion, operative time, postoperative hemorrhage, T stage, perineural invasion, microvascular invasion, adjacent organ invasion, AJCC stage, and histological subtype.

LASSO-Cox model development

Next, we entered these 10 variables into Cox regression. Two were not independent prognostic factors (P > .05; Table S4). We excluded them and remodeled with the remaining 8. Each was an independent factor (Table 1; Figure S3C). There was no significant collinearity (VIF < 2) or interaction (Figure S3D). The model satisfied the Grambsch–Therneau test (Global P = .17, Figure S4). We refer to this novel model as the LASSO-Cox model. Age, intraoperative transfusion volume, and operative time were modeled as continuous variables in their natural units. For clinical interpretability, cumulative hazard ratios for representative increments are provided in Table S5.

Table 1.

Independent prognostic variables identified by LASSO-Cox regression in the training cohort.

LASSO-Cox model	VIF	HRa	95% CI	P value
Age, yr	1.03	1.0227	1.0155-1.0299	<.001
Intraoperative blood transfusion, mL	1.10	1.0004	1.0002-1.0005	<.001
Operative time, min	1.09	1.0009	1.0003-1.0014	.003
AJCC stage, n (%)b	1.08			<.001c
Stage II		1.5496	1.2754-1.8829	<.001
Stage III		1.8275	1.5053-2.2186	<.001
Stage IV		1.9936	1.3538-2.9357	.001
Perineural invasion, n (%)	1.23	1.4486	1.2227-1.7163	<.001
Microvascular invasion, n (%)	1.23	1.3557	1.1367-1.6170	.002
Postoperative hemorrhage, n (%)	1.04	1.5633	1.2528-1.9506	<.001
Histological subtype, n (%)	1.02	0.4856	0.4170-0.5655	<.001

Abbreviations: AJCC, American Joint Committee on Cancer; VIF, variance inflation factor; VIF ≥ 2, collinearity may exist between independent variables..

^aHRs for continuous variables are per-unit.

^bBased on the 8th edition.

^cWald test.

In secondary analyses (Table S6), tumor recurrence was strongly associated with subsequent mortality in the time-dependent Cox model (adjusted HR 18.67, 95% CI, 15.63-22.30; P < .001). Landmark analyses at 12, 24, and 36 months confirmed consistent associations with worse subsequent survival (all P < .001).

Risk stratification

The model’s time-dependent ROC showed strong accuracy: AUCs of 0.72 (95% CI, 0.68-0.77), 0.75 (95% CI, 0.72-0.77), and 0.76 (95% CI, 0.73-0.79) for 1-, 3-, and 5-year OS (Figure S5A). For stratification, we used the 5-year OS threshold (highest AUC). The Youden index gave a cutoff of 0.40 (Figure S5B). This divided patients into low-risk (≤0.40) and high-risk (>0.40) groups. OS differed significantly between groups (log-rank test; Figure S5C). The cumulative event probability was significantly higher in the high-risk group throughout the follow-up period, underscoring the effectiveness of risk stratification in predicting event occurrence (HR 3.41, 95% CI, 2.95-3.94, P < .01; Figure S5D).

Evaluation and validation of the novel predictive model

LASSO-Cox model evaluation

We compared the LASSO-Cox model to 3 traditional models: univariate-Cox, full-Cox, and AJCC-Cox (Figure S6A-C). At 60 months, the LASSO-Cox time-dependent AUC was 0.760, higher than 0.725 (univariate-Cox), 0.723 (full-Cox), and 0.693 (AJCC-Cox) (AUC, all P < .05). The LASSO-Cox had better fit (AIC 10595.5 vs 10654.5, 10665.0, 10836.0). Its IDI and NRI exceeded 0 compared to others (Table 2). Dynamic C-index plots confirmed consistent superiority (Figure 3A). Decision curves showed greater net benefit for 1-, 3-, and 5-year OS (Figure 3B; Figure S7A-B). Clinical impact and calibration curves further demonstrated superior performance of the LASSO-Cox model compared with the 3 traditional models (Figure 3C and D; Figure S7C–F).

Table 2.

Performance comparison of 4 models in the training cohort.

Models	Variable inclusion methods	Independent variables number (P < .05)	AIC	Time-dependent AUC （%）a	IDIa,b	NRIa,b	P-valueb,c
LASSO-Cox model	Penalized regression	8	10595.49	76.0	NA	NA	NA
Univariate-Cox model	Univariate analysis	10	10654.48	72.5	0.032	0.176	.001
Full-Cox model	Full variables	12	10664.97	72.3	0.036	0.174	.002
AJCC-Cox model	Reference23	3	10836.02	69.3	0.088	0.241	<.001

Abbreviations: AIC, Akaike information criterion. AIC is a standard to measure the goodness of model fit. The smaller the AIC value, the better the model; IDI, integrated discrimination improvement; NRI, net reclassification improvement.

^aTime cutoff was set to 60 with 100 iterations.

^bComparison with the LASSO-Cox model.

^cComparison by time-dependent AUC.

Figure 3.

Evaluation of the LASSO-Cox model. (A) Dynamic C-index plots over time across models; higher curves indicate superior discrimination. (B) Decision curve analysis comparing net benefit of different models for 3-year OS across threshold probabilities; higher curves denote greater clinical utility. (C) Clinical impact curve comparing high-risk event predictions of different models for 3-year OS. (D) Calibration curves comparing predicted and observed 3-year OS; closer alignment with the ideal line indicates greater accuracy.

LASSO-Cox model external validation

In the validation cohort, the LASSO-Cox model maintained strong performance. Time-dependent ROC analysis yielded AUCs of 0.76 (95% CI, 0.66-0.86) and 0.79 (95% CI, 0.74-0.86) for 1- and 3-year OS, respectively, and 0.81 (95% CI, 0.74-0.89) for estimated 5-year OS given the limited follow-up (Figure 4A). Dynamic C-index plots confirmed superiority (Figure S8A). The LASSO Cox model achieved comparable time-dependent AUCs to the AJCC Cox model, but yielded a lower AIC (1034.7). Positive IDI and NRI further supported its superiority (Table S7). Decision curves showed higher net benefit (Figure 4B; Figure S8B-C). Using the pre-specified cutoff of 0.40 from the training cohort, risk groups showed clear OS differences (Figure 4C). Discrimination remained strong in both risk groups (HR 3.73, 95% CI, 2.51-5.54, P < .01; Figure 4D). Overall, the LASSO-Cox model demonstrated excellent discrimination and stability in external validation.

Figure 4.

External validation of the LASSO-Cox model. (A) Time-dependent ROC curves for predicting 1-, 3-, and 5-year OS. The 5-year OS was estimated from Kaplan–Meier analysis with limited follow-up beyond 48 months. (B) Decision curve analysis comparing net benefit of different models for 5-year OS across threshold probabilities; higher curves denote greater clinical utility. (C) Kaplan–Meier survival curves comparing OS between high- and low-risk groups (log-rank test). (D) Cumulative incidence curves comparing death event probabilities between risk groups (log-rank test).

Due to substantial missingness in adjuvant therapy data (Figure S9A), an exploratory analysis was conducted in the external validation cohort (n = 309). In the overall cohort, adjuvant therapy was not significantly associated with improved survival (HR 1.42, 95% CI, 0.97-2.09, P = .11). Subgroup analyses showed benefit in high-risk patients (adjusted HR 0.54, 95% CI, 0.32-0.93, P = .02; Figure S9B), but no difference in low-risk patients (adjusted HR 0.81, 95% CI, 0.36-1.87, P = .62; Figure S9C). No significant interaction between risk group and adjuvant therapy was observed (adjusted HR 0.68, 95% CI, 0.25-1.82, P = .45).

Nomogram development and model visualization

Finally, we constructed a nomogram from the LASSO-Cox model to predict 1-, 3-, and 5-year OS and risk stratification (Figure 5A). It included one preoperative variable (age), 2 intraoperative (blood transfusion, operative time), and 5 postoperative (postoperative hemorrhage, perineural invasion, microvascular invasion, AJCC stage, histological subtype). To aid clinical use, we visualized the nomogram as an interactive web tool (Figure 5B; https://zongtingteamcode.shinyapps.io/dynamic-da-os/).

Figure 5.

Development and visualization of the LASSO-Cox nomogram for resected DA. (A) Nomogram predicting 1-, 3-, and 5-year OS, with variable point assignments and risk thresholds. Intraoperative blood transfusion was modeled as transfusion volume (mL) without conversion from blood-product units. (B) Interactive web-based nomogram with sample patient data (accessible online).

Discussion

DA is a rare gastrointestinal malignancy with poor prognosis despite resection. To address this gap, we developed and externally validated a LASSO-Cox nomogram and web tool for postoperative outcome prediction. From 89 perioperative variables in a large multicenter cohort, 8 independent OS predictors were identified: age, intraoperative transfusion, operative time, postoperative hemorrhage, perineural invasion, microvascular invasion, AJCC stage, and histological subtype. The model demonstrated strong discrimination, enabling effective risk stratification. In the era of precision oncology, perioperative DA management still relies on morphological staging rather than molecular profiling. Our nomogram incorporates routinely available perioperative metrics and reflects real-world practice, where adjuvant therapy benefits remain debated and largely limited to advanced disease^25,26; its contribution is to generate individualized risk estimates rather than propose novel prognostic factors.

Our findings align with established predictors like age, perineural invasion, microvascular invasion, and AJCC stage, which reflect tumor aggressiveness and patient frailty.^13,27–29 Histological subtype was a strong independent prognostic factor. Na-DAs were classified by differentiation phenotype to allow standardized comparison across anatomical sites.^4,30 Patients with intestinal type histology had about half the mortality risk of those with pancreatobiliary or mixed types (HR 0.49, 95% CI, 0.42-0.57), consistent with prior evidence that intestinal differentiation confers more favorable biology.^25,31 Routine subtype classification should be incorporated into DA pathology reporting, as it refines risk stratification beyond TNM staging and may inform subtype-specific therapeutic strategies. A postoperative prognostic nomogram should rely exclusively on information available at its intended time of clinical use. Including tumor recurrence in the primary model would introduce temporal bias, as recurrence is a post-baseline event rather than a variable available when the nomogram is applied. We therefore excluded recurrence from the primary model and analyzed it separately as a time-dependent covariate in follow-up analyses. Although recurrence was strongly associated with subsequent mortality (adjusted HR 18.67), it is more appropriately incorporated into surveillance protocols and dynamic risk reassessment than into primary postoperative prognostication.³²

Importantly, perioperative factors including intraoperative blood transfusion, operative time, and postoperative hemorrhage were identified as modifiable contributors to poor outcomes. Heavy intraoperative transfusion and prolonged operative time induce proinflammatory, immunosuppressive states with elevated cytokines (eg, IL-6, TNF-α), reduced NK cell activity, and impaired antigen presentation. This increases recurrence risk.^33,34 Postoperative hemorrhage raises mortality directly and via ischemia–reperfusion, added interventions, and delayed adjuvant. All harm short- and long-term results.^35,36 Strategies to shorten operative time, improve surgical quality, and use restrictive transfusion can mitigate. Notably, hazard ratios per unit increase in intraoperative blood transfusion and operative time were close to 1.0 yet statistically significant. Given their wide ranges, cumulative effects can substantially elevate postoperative risk. In nomogram construction, predictor weight reflects both coefficient and observed range, so even near-unity HRs can yield clinically relevant differences. By incorporating these variables, our model outperforms traditional approaches and underscores the importance of surgical optimization for improved survival.

Exploratory analysis in the validation cohort, adjuvant therapy was not associated with improved survival, likely reflecting confounding by indication, as it tended to be administered to patients with more aggressive disease.^25,26 Stratification by the LASSO-Cox model suggested benefit in high-risk patients but no difference in low-risk patients, consistent with the rationale that aggressive tumors with greater micrometastatic potential derive more benefit from systemic treatment.^26,37 The interaction between risk group and adjuvant therapy did not reach statistical significance, and these exploratory findings should be interpreted cautiously given the limited sample size and nonrandomized treatment allocation, since interaction tests generally require larger cohorts than main-effect analyses.³⁸ Nonetheless, these results suggest that LASSO-Cox–based risk stratification may help identify patients most likely to benefit from adjuvant therapy in DA, a hypothesis warranting prospective validation.

Compared with prior small, retrospective a-DA studies lacking external validation,^7,9,13–15 our nationwide multicenter design represents the largest DA cohort to date (n = 2624). Traditional univariate models prone to overfitting performed worse,¹⁶ whereas the LASSO-Cox approach managed high-dimensional data effectively, yielding an interpretable and stable “white-box” model. External validation confirmed generalizability and clinical reliability, and the study further provides practical tools including a nomogram and an interactive web app for real-time risk assessment. These contributions address key gaps in rare cancer prognostication by enabling clinicians to identify high-risk patients for intensified follow-up or multidisciplinary discussion to refine adjuvant therapy decisions. By incorporating modifiable factors, the model also highlights preventive strategies that may improve overall survival. Although external validation often yields attenuated discrimination, the LASSO-Cox model achieved higher time-dependent AUCs in the validation cohort. This likely reflects methodological and cohort-level factors rather than overfitting: the shorter observation window supported more reliable 1- and 3-year estimates, whereas the 5-year AUC was extrapolated from limited follow-up and thus less stable. Early clustering of events also enhanced discrimination,³⁹ underscoring the need for extended follow-up to clarify whether the 5-year estimate remains consistent or converges toward the training cohort.

Despite these strengths, limitations exist. First, multicenter data and potential nonrandom exclusion may introduce heterogeneity in surgical techniques and management, though centralized auditing mitigated this. Second, the long inclusion period could reflect evolving practices, addressed partially by AJCC 8th edition restaging. Third, incomplete availability of detailed systemic therapy data and molecular features (eg, MSI-H/MMR status) represents an important limitation. Although adjuvant therapy status was available for exploratory analysis, information on specific regimens, dose intensity, and treatment response was not uniformly captured. The nomogram should therefore be interpreted as a postoperative prognostic tool rather than a treatment-selection model. Finally, reliance on Chinese cohorts raises questions of ethnic generalizability, and multiple imputation, while standard, could introduce minor biases. This prediction model relies on associations to support clinical decisions, not to infer causation. Clinicians should apply it judiciously. Future research should validate the model in international cohorts, incorporate genomic data and adjuvant therapy details for refined predictions, and assess impacts on clinical outcomes through randomized trials.

Conclusions

We developed and externally validated a postoperative prognostic model for resected DA using routinely available perioperative variables. The model demonstrated prognostic utility in the present study and may support postoperative counseling and follow-up planning. However, it should be considered complementary to standard staging and clinical judgment, and broader clinical implementation will require additional validation in independent cohorts.

Author contributions

Zongting Gu (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing), Qifeng Xiao (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Writing—original draft, Writing—review & editing), Xin Wu (Data curation, Formal analysis, Methodology, Project administration, Resources, Writing—original draft, Writing—review & editing), Chunhui Yuan (Data curation, Investigation, Methodology, Resources, Supervision, Writing—original draft, Writing—review & editing), Xiaolong Tang (Data curation, Formal analysis, Methodology, Resources, Writing—original draft, Writing—review & editing), Fanbin Meng (Data curation, Methodology, Project administration, Resources, Writing—original draft, Writing—review & editing), Dong Wang (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Ren Lang (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Kaiqing Guo (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Xiaodong Tian (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Yu Zhang (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Enhong Zhao (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Zhenzhou Wu (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Jingyong Xu (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Ying Xing (Data curation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Feng Cao (Conceptualization, Data curation, Investigation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), Chengfeng Wang (Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing), and Jianwei Zhang (Conceptualization, Formal analysis, Investigation, Project administration, Resources, Supervision, Writing—original draft, Writing—review & editing)

Supplementary material

Supplementary material is available at The Oncologist online.

Funding

This work was supported by the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (CIFMS; 2022-I2M-1-010) and the Medical and Health Science Program of Zhejiang Province (2025HY0116).

Conflicts of interest

None declared.

Data availability

The original data presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Ethical approval

The study procedures were approved by the National Cancer Center Ethics Committee (No. 22/213-3415).

Code availability

An open-source version of the code for the model can be accessed at the following URL: https://github.com/ZongtingCode/DA-OS.