All patients diagnosed with invasive ductal carcinoma–type pancreatic cancer were identified from the SEER database (http://seer.cancer.gov/) using SEER*Stat Software (version 8.3.9.2). The SEER research data included SEER incidence and population data associated with age, sex, race, year of diagnosis, and geographic areas (including SEER registry and county). The clinical features included age, gender, tumor size, lymph nodes (LNs) examined, regional nodes positive, histological grade, histologic type, American Joint Committee on Cancer (AJCC) stage, total number of in situ/malignant tumors in a patient, TNM stage, primary tumor site, surgery, survival time, and survival status. Surgery is not effective in treating stage III or IV pancreatic cancer patients, so we only selected patients with stage I or II pancreatic cancer. Stage I and II patients who underwent surgery and had complete clinical information were chosen for further analysis.

The eligibility criteria for the study included the following: (1) patients with histological codes of 8500 and 8521, (2) patients subjected to the first surgical excision, (3) patients receiving lymph node dissection surgery, and (4) patients with a diagnosis of stage I or II pancreatic cancer. The exclusion criteria encompassed the following: (1) patients failing to record ELN and complete clinical information, (2) stage IIIB and IV patients due to surgery not being prioritized for them, (3) stage N2, N3, and M1 patients, and (4) patients of unknown or incomplete survival data and clinical features (unrecorded number of lymph nodes before preoperative examination and irradiation).

In this study, the ELNs were divided into low and high subgroups, and an attempt was made to evaluate all possible divisions of the ELN data. The function “surv_cutpoint” from the R packages “survminer” and “survival” was utilized to discover the optimal cut-off value of ELNs. The cut-off point was defined as the value at which the survival prognosis of the two groups differs most significantly among all possible subgroups of ELNs. We collected the cut-off points of ELNs and the ELN/RNP ratio from patients with stage I and II IDCP. In addition, we analyzed the threshold of ELNs and the ELN/RNP ratio for patients at the T3N0M0 or T3N1M0 stage of the disease.

Before survival analysis, patients at stage I and II underwent PSM for the purpose of adjusting potential biases by selecting statistically different variables in the propensity model. A caliper—defined as the maximum tolerated difference between matched subjects in terms of “non-perfect” matching—was set at 0.01. The selected variables of stage I and II patients included age, gender, T-staging, N-staging, and stage of cancer. For these patients, the selected variables in the propensity model covered age, gender, radiation, and stage of cancer.

Finally, we assessed the correlation between the ELN/RNP ratio and IDCP mortality using restricted cubic spline (RCS).

We had previously collected the optimal cut-off values of ELNs for survival of stage I and II patients. These data were subjected to subgroup survival analysis to evaluate the survival benefit of each group (age > 69 years, age ≤ 69 years, females, males, N0, N1, T1, T2, T3, stage I, and stage II).

Based on the optimal cut-off values for ELNs associated with survival benefit, model 1 included clinical factors such as ELNs, age, sex, T-staging, and N-staging. Adjusted model 2 covered ELNs, age, gender, T-staging, N-staging, radiation, and chemotherapy.

The forest plot shows the Odds ratio (OR) and P-values for each clinical factor in this prediction model.

R software’s rms package generated an alignment diagram of a multivariate model, indicating the 1-, 2-, and 3-year survival rates of each subgroup. We used the C-index to quantify the discriminative ability of the nomogram, which estimated the difference between predicted and actual survival. We also performed a decision curve analysis (DCA) to evaluate the clinical benefit of our model.

SPSS (Version 26.0) and R 3.4.1 software (http://www.r-project.org) were employed to analyze the data. Pearson's chi-square test and the independent t-test were used to compare the baseline pathological characteristics. The Kaplan–Meier method was utilized to calculate the cumulative survival rate; a Student’s t-test was used for the logarithms to compare the survivorship curve. The prognostic factors that proved statistically significant in univariate analysis were analyzed via the multivariate Cox proportional hazard model. Results were measured with 95% confidence intervals (CI), and a two-tailed P-value of <0.05 was considered statistically significant.

We assessed the correlation between the cut-off value of ELNs and the survival rate of IDCP patients using a method of exclusion. For ELN values of ≥7, there were significant survival differences between groups. For ELN values of ≥24, there were survival differences. Thus, 24 ≥ ELNs ≥ 6 was recommended to excise in clinical. When 10 was used as the divider, the two groups showed the most significant differences, with maximum chi-square values of 12.206, P = 0.00048 (Supplementary Table S1). Accordingly, ELN values of ≥10 were considered the minimum number for optimal survival for IDCP patients (Figures 1A,B). To test the sensitivity of the optimal cut-off value of ELNs, we randomly extracted 30%, 50%, and 80% of patients from the entire patient cohort. According to the results, the optimal cut-off value of ELNs was 9 (Supplementary Figure S1).

To ensure the accuracy and reliability of results, we used a PSM analysis to reduce confounding factors; ELN values of >10 and<10 were analyzed as baseline values for each group. For pre-PSM patients, age (P = 0.005), T-stage (P < 0.001), N-stage (P < 0.001), and AJCC stage (P < 0.001) showed a maldistribution between groups (Supplementary Table S2). The maldistribution of influence factors between the two groups could potentially lead to bias in the survival analysis. For uneven baselines, the data were subjected to PSM to exclude influence factors. To guarantee the largest sample size on even baselines, our study used different caliper values and matching for data of different groups. To warrant equal numbers of patients in both groups, our study matched pancreatic cancer patients with ELN >10 and <10 in a 1:1 ratio. After PSM matching, there were no significant differences in age, T-stage, and N-stage between the two groups. As shown in Supplementary Table S3, a total of 1,447 pairs were successfully matched when the PSM model used 1:1 matching with 0.01 calipers. After matching, there were no statistical differences in the variables of age (P = 0.911), gender (P = 0.882), T-staging (P = 0.944), N-staging (P = 0.882), and AJCC stage (P = 1.000) between the two groups.

According to the post-PSM results of IDCP patients, ELNs values of ≥10 and <10 were still important prognostic factors (Figure 1C). Then, we examined the effect of optimal cut-off values for ELNs on the prognosis of each subgroup of IDCP. According to the survival analysis for each subgroup (age > 69 years, age ≤ 69 years, female, male, N0, N1, T3, and stage II), patients with ELNs values of ≥10 had better estimated median survival rates (Figure 2, Supplementary Figure S2, Supplementary Table S4). However, no significant prognostic differences were found in the T1, T2, and stage I subgroups (Supplementary Figure S2). We also presented the estimated OR risk value for each subgroup based on the optimal cut-off point of ELNs in IDCP patients (post-PSM) (Supplementary Figure S3). ELN values of ≥10 remained beneficial for survival in stage T2 patients with OR >1 and P-values close to 0.05. Univariate and multivariate Cox regression analysis indicated that the ELN cut-off point of 10 was an independent prognostic factor (univariate Cox regression: HR, 1.241; 95% CI: 1.143–1.349; P < 0.001; multivariate Cox regression: HR, 1.263; 95% CI: 1.162–1.372; P < 0.001) (Supplementary Table S5).

A multivariate Cox regression prognostic model was constructed based on the optimal cut-off value for ELNs. Model 1 (base model) incorporated ELNs, age, sex, TNM stages, and stage of cancer. Model 2 (adjusted model) included ELNs, age, sex, stage, TNM stages, radiation, and chemotherapy. Figure 3 presented the estimated OR risk values for each factor, which indicated ELNs was a significant prognostic factor [OR = 1.24 (95% CI = 1.14–1.34), P < 0.001]. Supplementary Figure S4 shows the visualized multifactor Cox regression model with a C-index value of 0.6597. DCA showed that model 2 had a higher predictive efficacy than model 1, with better mortality prediction ability (Supplementary Figure S5).

We employed both baseline and adjusted XGBoost models to predict mortality in stage I and II IDCP patients. The baseline model incorporated the following variables: age, sex, surgery, tumor size, T-stage, N-stage, primary site labeled, regional nodes positive, and total number of in situ/malignant tumors per patient. The adjusted model included all variables from the baseline model, with the addition of regional nodes examined. The results demonstrated that the adjusted model exhibited higher sensitivity in predicting patient prognosis (P < 0.05) (Figure 4). The machine learning model (XGBoost) indicated that ELN risk categories serve as a significant prognostic factor for patients with IDCP.

In addition, given the small number of patients in T1 and T2 stages, further analysis of ELN cut-off values for optimal survival benefit was conducted only for patients in the T3 stage. We calculated the optimal cut-off point for T3N0M0 patients with IDCP. The result showed an ELN value of 7 was the most significant division point for the prognostic differences between the two groups (ELNs > 7 vs. ELNs ≤ 7) (Figure 5, Supplementary Figure S6). Likewise, the optimal cut-off point for patients with T3N1M0 IDCP was calculated. An ELN value of 12 was the most significant division point for the prognostic differences between the two groups (ELNs > 12 vs. ELNs ≤ 12).

Patients with T3N1M0 IDCP were selected as the target population, and the outcome variable was set to “death.” This XGboost model included the following variables: age, sex, surgery, tumor size, T-stage, N-stage, primary site labeled, regional nodes examined, regional nodes positive, and total number of in situ/malignant tumors for a patient. The model provided the following results: ELNs were the most significant predictor (Figure 6).

As mentioned earlier in this study, the ELNs of IDCP are dependent on RNP, and thus the survival of pancreatic cancer patients is influenced by the ELN/RNP ratio. Therefore, we evaluated the minimum ratio of ELNs to RNP for optimal survival benefit of IDCP patients. As shown in Figure 1, an ELN/RNP ratio of 9.25 is the threshold for survival benefit in IDCP patients, with the risk of death due to IDCP gradually decreasing with an increasing ELN/RNP ratio (Figure 7). Therefore, an ELN/RNP ratio of 9 is the optimal cut-off value for clinical benefit.