Prognostic Features in Surgically Resected Well-Differentiated Pancreatic Neuroendocrine Tumors: an Analysis of 904 Patients with 7882 Person-Years of Follow-Up

Abstract

The clinical behavior of well-differentiated pancreatic neuroendocrine tumors (PanNETs) is difficult to predict. In order to define, more accurately, prognosticators for patients with a surgically resected PanNET, the pathologic features and Ki-67 immunolabeling indexes of PanNETs resected from 904 consecutive patients at an academic tertiary care hospital were correlated with patient outcome. The mean patient age at surgery was 56.6 years (SD 14.0), 477 were male (52.8%), and 7882 person-years of follow-up were obtained (mean 8.8 years, SD 6.5). The 10-year survival was 81% (95% CI: 77,86%) for patients with G1 PanNETs (Ki-67 <3%), 68% (95% CI: 61,76%) for patients with G2a PanNETs (Ki-67 3 - <10%), 44% (95% CI: 29,66%) for patients with G2b PanNETs (Ki-67 of 10%- ≤20%), and 23% (95% CI: 8,61%) for patients with G3 PanNETs. Vascular invasion (HR 3.0, p <0.0001), tumor size ≥ 2 cm (HR 2.88, p <0.0001), perineural invasion (HR 2.42, p<0.0001), and positive margins (HR 2.18, p <0.0001) were associated with worse overall survival. Insulinoma (HR 0.34, p=3e-04), sclerosing variant (HR 0.47, p=0.05), and cystic variant (HR 0.61, p=0.05) were associated with improved overall survival. T, N and M stages were all statistically significant classifiers of overall survival. Similar associations were found with respect to disease relapse. There was a significant (P<0.001) increase in the proportion of patients diagnosed with stage I vs stage IV disease over time. This study supports the classification of PanNETs into four grades (G1, G2a, G2b, and G3) based on Ki-67 labeling, which allows a more accurate prognostic assessments of patients.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12022-025-09866-z.

Introduction

The incidence of well-differentiated pancreatic neuroendocrine tumors (PanNETs), the most prevalent non-ductal neoplasms of the pancreas, is increasing with increasing use of abdominal imaging [1–7]. Clinical management of the growing numbers of patients with PanNETs is challenging [8]. A number of new medical therapies are available for patients with advanced disease, and expert opinions vary on the optimal use of surgery, complicating the selection of the best therapy for patients [8–10]. These challenges are perhaps greatest for patients with small, <2 cm, PanNETs, as prognostic classification of these relatively indolent tumors is imperfect, and the risks of surgery are not trivial [11–14]. For example, Partelli and colleagues concluded, from the results of the ASPEN trial, that active surveillance is safe for patients with a small PanNET, while Lin and Huang, in an analysis of 1102 patients, concluded that surgical resection is recommended for these patients [12, 14]. As a result of this uncertainty, guidelines vary, and the treatment of a significant proportion of patients is not based on any guidelines [10, 11].

A number of clinical, pathological and genetic prognostic markers for patients with a PanNET have been identified [4, 7]. Male sex and increasing age are both associated with shorter survival [4, 15]. Clinically, insulinomas and cystic PanNETs have been reported to be associated with improved survival [7, 16–21]. The pathological factors associated with survival include tumor size, lymph node status, distant metastasis, the proliferation rate of the neoplastic cells, vascular invasion, and perineural invasion [22–27]. Margin status, necrosis, a sclerotic growth pattern with serotonin expression, an invasive growth pattern, and the type and intensity of any associated immune infiltrates have also been reported as significant prognosticators [22, 28–35]. The genes DAXX and ATRX are frequently somatically inactivated in PanNETs and inactivation of either one of these genes is associated with the alternative lengthening of telomeres (ALT) phenotype [36]. The ALT phenotype and inactivation of DAXX or ATRX are also poor prognosticators [37–40]. Despite these predictive features, prognostication remains imperfect, and there is still room for improvement [7].

Recently, several studies have proposed refinements to prognostic features [41, 42]. Currently, PanNETs are divided into three grades based on proliferation rate of the neoplastic cells: G1 (Ki-67 <3%), G2 (Ki-67 3–20%) and G3 (Ki-67 >20%) [7, 43]. Adsay and colleagues as well as others have suggested that grade 2 PanNETs should be subdivided into two grades (G2a for Ki-67 3- <10%, and G2b for Ki-67 of 10%- ≤20%), and that G2b PanNETs have a prognosis similar to G3 (Ki-67 >20%) PanNETs [41, 44, 45]. For small tumors, <2 cm, Pawlik and colleagues suggested that vascular invasion may be a particularly useful prognosticator [42].

To further refine the prognosticators for surgically resected PanNETs, particularly T1N0M0 tumors, and to specifically examine the most appropriate cut-off for proliferation rate, we reviewed the pathology of a large single-institution series of surgically resected PanNETs with up to three decades of follow-up and correlated findings with patient outcome. As it has been suggested that the incidence of clinically recognized PanNETs is increasing because of the incidental detection of asymptomatic lower-stage PanNETs, we also examined trends in age and stage at diagnosis over time [1, 46].

Methods

Study Population

This study was approved by the Institutional Review Board of the Johns Hopkins Hospital. The pathology and surgery files of the Johns Hopkins Hospital were searched for surgically resected PanNETs from January 1984 to January 2025. All available medical records were reviewed, as were all available microscope slides. Cause of death could be determined by chart review in the majority of deceased patients. In a minority of cases, death indexes were used to determine the date of death, although the cause is unknown for this group of patients.

Pathology Review

Five of the authors (ALK, EDY, GK, AK and RHH), as a group, reviewed all available slides. Neuroendocrine neoplasms metastatic to the pancreas, and neoplasms not meeting diagnostic definitions outlined in the 5th edition of the World Health Organization (WHO) Classification of Tumors of the Digestive System were excluded, leaving 904 patients [43]. Histologic slides were available from 883 of these 904 cases (Mean of 26, and median of 25 slides per case). Cystic PanNETs were defined by imaging and confirmed by gross appearance. Insulinoma was defined as a PanNET with associated clinical findings of hyperinsulinemic hypoglycemia [7]. Patients had the classical “Whipple triad” of symptoms of hypoglycemia, low blood glucose levels (below 3.0 mmol per liter), and relief of symptoms when given glucose [7, 47]. Margins and size were recorded as reported in the pathology report. The sclerosing variant of PanNET was defined as a PanNET composed of cords of cells embedded in dense stromal fibrosis, centered on a large pancreatic duct, often with upstream ductal dilatation, and frequent expression of serotonin by immunolabeling [33–35, 48–50].

Proliferation Rate

Tumor Ki-67 labeling index was available for 820 of the 904 patients. Ki-67 labeling was determined by one of two methods. For 302 of the cases, a labeling “hot spot” with the highest density of Ki-67 labeling cells was photographed, the photo printed and at least 500 neoplastic cells were manually counted [51]. For 534 of the cases the percentage of neoplastic cells labeling in a “hot spot” was determined by the pathologist [51]. As it has been suggested that stratification of G2 tumors may be of prognostic significance, we stratified tumors into four grades; G1 (Ki-67 <3%), G2a (Ki-67 3- <10%), G2b (Ki-67 10%- ≤20%), and G3 (Ki-67 >20%).

In order to compare the two methods of determining the Ki-67 labeling index, we selected 100 of the 534 cases for which the Ki-67 labeling was determined by a pathologist, and freshly immunolabeled the tumor (Ki-67 2μg/ml, Roche, Cat #790–4286, UV HRP Universal Multimer, UltraView DAB detection kit, Roche Ventana Benchmark Ultra Stainer). A labeling “hot spot” with the highest density of Ki-67 labeling cells was then photographed, and at least 500 neoplastic cells were manually counted [51]. The Ki-67 labeling index determined by manual counting was then correlated with the original pathologist’s count.

Statistics

The primary endpoint of this study was overall survival (OS), defined as the time from surgery to last follow-up or death and was estimated using the Kaplan-Meier method. Eleven of the 904 patients had more than one pancreas surgery. In these instances, we included their first surgery in the analyses. Twenty-nine of the 904 patients had <30 days of follow-up information and were excluded from recurrence and survival analyses. Hazard ratios (HR) for differences in OS according to patient subgroups were estimated with Cox proportional hazards models. Cause-specific survival was estimated similarly, except patients who died without relapse were censored on their date of death. Time to relapse was calculated as the time from surgery to date of relapse (event), death (competing event) or last follow-up, whichever came first. Patients with unknown disease status at last follow-up were excluded. The impact of margin status on outcome was calculated for the entire cohort and separately after excluding patients who had an enucleation procedure. Estimates of cumulative incidence of relapse at 2, 5, and 10 years and differences in time to relapse according to patient subgroups defined at the time of surgery were estimated using proportional sub-distribution hazards regression models [52].

A multivariable model selection approach for OS was performed using a Least Absolute Shrinkage and Selection Operator (LASSO) method. LASSO, a regularization procedure, selects optimal covariates by applying a penalty factor through cross-validation. It performs both variable selection and shrinkage, reducing some coefficients to zero and improving model generalization by controlling overfitting [53]. Age (per 5 years), sex, year of surgery, tumor size (per 1 cm), Ki-67, TNM staging, and presence of insulinoma, sclerosing variant, cystic variant, vascular invasion, perineural invasion and margins were offered as candidates to the model selection. Estimates, 95% confidence intervals, and p-values were bias-corrected using the selective Inference package in R [54].

Restricted cubic splines were used to assess the potential non-linear association between Ki-67 and tumor size with OS. To estimate changes in the prevalence of stage I disease over time, a multinomial regression model with stage (I, II, III, or IV) was fit with year of surgery included as the main independent variable, adjusting for age and gender. Due to the relatively small number of patients represented in some years, year of diagnosis was grouped into quintiles for this analysis. Predicted probabilities of being diagnosed with stage I disease were estimated from the model using the MNLpred package in R [55]. Changes in age at surgery over time were estimated with a simple linear regression model. For all analyses, p-values < 0.05 were considered statistically significant. Analyses were completed with R version 4.4.1 [56].

Results

Patient Characteristics and Follow-Up

Patient characteristics are shown in Table 1. A total of 7882 person-years of follow-up were obtained (mean 8.8 years, SD 6.5) on the 875 patients followed for > 1 month. Of the 875 patients, 251 were followed until death. Disease status at last follow-up was known on 822 (94%) of the 875 patients.

Table 1.

Descriptive summary of patient characteristics at the time of surgery and in follow-up

	N = 904
Age at Surgery, years
Mean (SD)	56.6 (14.0)
Median [Min, Max]	58.0 [12.3, 93.3]
No. observations	904
Age at Surgery - no. (%)
< 60	509 (56.3)
60+	395 (43.7)
Sex - no. (%)
Female	427 (47.2)
Male	477 (52.8)
Year of Surgery - no. (%)
1984 - 2000	106 (11.7)
2001 - 2010	306 (33.8)
2011 - 2020	390 (43.1)
2020 - 2024	102 (11.3)
Years of Follow-Up
Mean (SD)	8.8 (6.5)
Median (Min, Max)	7.7 [0.0, 40.2]
No. observations	900
Death or LFU < 30 d of surgery - no. (%)
Alive at 30 d	875 (97.2)
Death or LFU < 30 d	25 (2.8)
No. Missing	4
Type of Surgery - no. (%)
Whipple	324 (35.8)
Total Pancreatectomy	19 (2.1)
Central Pancreatectomy	13 (1.4)
Distal Pancreatectomy	471 (52.1)
Enucleation	76 (8.4)
No Surgery/Missing	1 (0.1)
Tumor Size, cm
Mean (SD)	3.4 (3.0)
Median (Min, Max)	2.5 [0.1, 27.0]
No. observations	904
Tumor Size - no. (%)
< 2 cm	338 (37.4)
2 cm +	566 (62.6)
T Stage - no. (%)
T1	338 (37.4)
T2	330 (36.5)
T3	214 (23.7)
T4	22 (2.4)
N Stage - no. (%)
N0	552 (61.1)
Nx or Unknown	83 (9.2)
N1	269 (29.8)
M Stage - no. (%)
M0	796 (88.1)
M1a	99 (11.0)
M1b	3 (0.3)
M1c	6 (0.7)
Stage Grouping - no. (%)
I	255 (28.2)
II	263 (29.1)
III	193 (21.3)
IV	108 (11.9)
Unstaged	85 (9.4)
Ki-67
Mean (SD)	4.5 (6.6)
Median (Min, Max)	2.0 [0.1, 58.6]
No. observations	820
Grade - no. (%)
G1	461 (56.2)
G2	335 (40.9)
G3	24 (2.9)
No. Missing	84
Grade, Proposed - no. (%)
G1	461 (56.2)
G2a	282 (34.4)
G2b	53 (6.5)
G3	24 (2.9)
No. Missing	84
Insulinoma - no. (%)
No	805 (89.0)
Yes	99 (11.0)
Functional - no. (%)
ACTH	1 (0.8)
Gastrinoma	15 (11.8)
Glucagonoma	3 (2.4)
Insulinoma	99 (78.0)
Pancreatic Polypeptide	1 (0.8)
VIPoma	8 (6.3)
Non-functional	777
Sclerosing Variant - no. (%)
No	853 (94.4)
Yes	51 (5.6)
Cystic Variant - no. (%)
No	813 (89.9)
Yes	91 (10.1)
Vascular Invasion - no. (%)
No	667 (75.4)
Yes	218 (24.6)
No. Missing	19
Perineural Invasion - no. (%)
No	601 (67.5)
Yes	290 (32.5)
No. Missing	13
Margins - no. (%)
Negative	814 (90.7)
Positive	83 (9.3)
No. Missing	7
Margins (Excluding Enucleations) - no. (%)
Negative	768 (93.1)
Positive	57 (6.9)
No. Missing	79

Cm centimeter, d days, max maximum, min minimum, no number, SD standard deviation, LFU last follow-up, ACTCH adrenocorticotropic hormone-producing tumor

Methods of Determining Ki-67 Labeling Index

One hundred of the 534 in which the pathologist’s assessment was used to determine the Ki-67 labeling index were freshly stained for Ki-67 (to avoid assessing the labeling of the original slides which may have faded over time), and at least 500 cells in a hot spot were manually recounted. As has been reported in other tumor types, a strong positive linear relationship (r=0.78) was observed when the Ki-67 labeling index determined by manual counting of these 100 PanNETs was compared to the original pathologist’s count [57].

Prognosticators in the Entire Cohort

OS for the entire cohort is shown in Fig. 1A, and a sub-analysis of OS by stage in Supplemental Fig. 1. Vascular invasion (HR 3.00, 95% CI: [2.31, 3.91], p<0.0001), tumor size ≥ 2 cm (HR 2.88, 95% CI: [2.10, 3.93], p<0.0001), perineural invasion (HR 2.42, 95% CI: [1.88, 3.12], p<0.0001), positive margins (HR 2.18, 95% CI: [1.56, 3.04], p<0.001 for the overall cohort and HR 3.10, 95% CI: [2.18, 4.43], P≤0.0001 excluding enucleations), age ≥60 years (HR 2.08, 95% CI: [1.62, 2.68], p <0.0001), and male sex (HR 1.46, 95% CI: [1.14, 1.89], p=0.003) were all associated with worse OS (Table 2). Insulinoma (HR 0.34, 95% CI: [0.19, 0.62], p<0.001), sclerosing variant (HR 0.47, 95% CI: [0.22, 1.00], p=0.05), and cystic variant (HR 0.61, 95% CI: [0.38, 0.99], p=0.05) were associated with improved OS. Although DAXX and ATRX mutation status has been shown previously to correlate with prognosis, DAXX and ATRX mutations or immunolabeling were not specifically examined in these tumors. Grade, T stage, N stage and M stage, as defined by the World Health organization (WHO) and Union for International Cancer Control (UICC), were also all statistically significant classifiers of outcome [7, 43, 58].

Fig. 1.

A Kaplan-Meier curves for survival for the entire cohort. B Changes in the hazard ratio for death with increases in tumor size, estimated with restricted cubic splines. The plotted results represent the hazard ratio for different values of the continuous predictor (tumor size), visualizing any non-linear effects captured by the restricted cubic spline. C Kaplan-Meier curves for overall survival, separately according to grade. D Changes in the hazard ratio for death with increases in Ki-67, estimated with restricted cubic splines. The plotted results represent the hazard ratio for different values of the continuous predictor (Ki-67), visualizing any non-linear effects captured by the restricted cubic spline. E Predicted probability [95% CI] of being diagnosed with stage I, II, III, or IV disease (y-axis) by year of surgery grouped into quintiles (x-axis). Probabilities and corresponding odds ratios for the average relative change in the odds of a stage I, II, or III diagnosis compared to stage IV with every increasing quintile of year of surgery, adjusted for age and sex, are estimated from a multinomial regression model

Table 2.

Overall Survival: Estimates of survival after surgery for the whole cohort and according to patient subgroups defined at the time of surgery, from univariate Cox proportional hazards models

	N	Follow-Up, PY	Median [95% CI]	10-year OS	20-year OS	HR	95% CI	P
Whole Cohort	875	7882	19.7 [16.8, 23.9]	74 [70, 77]	49 [44, 55]
Age, years
< 60	493	4882	26.5 [23.0, 40.2+]	81 [77, 85]	60 [53, 67]	1.0 (ref)	–
60+	382	2999	14.3 [12.1, 16.0]	64 [58, 70]	32 [24, 43]	2.08	[1.62, 2.68]	< 0.0001
Sex
Female	409	3894	26.5 [21.4, 40.2+]	76 [71, 81]	60 [53, 68]	1.0 (ref)	–
Male	466	3987	15.9 [14.6, 18.9]	72 [67, 77]	38 [30, 48]	1.46	[1.14, 1.89]	0.003
Tumor Size
< 2 cm	326	3222	27.1 [22.5, 40.2+]	88 [84, 92]	66 [57, 77]	1.0 (ref)	–
2 cm +	549	4659	14.9 [12.6, 18.1]	65 [61, 70]	40 [34, 47]	2.88	[2.10, 3.93]	< 0.0001
T Stage
T1	326	3222	27.1 [22.5, 40.2+]	88 [84, 92]	66 [57, 77]	1.0 (ref)	–
T2	319	2793	18.1 [15.5, 40.2+]	72 [66, 78]	48 [40, 58]	2.27	[1.61, 3.21]	< 0.0001
T3	208	1730	12.1 [10.2, 15.1]	58 [51, 66]	31 [23, 43]	3.68	[2.60, 5.20]	< 0.0001
T4	22	134	9.88 [6.90, 40.2+]	46 [26, 81]	0 [5, 81]	5.81	[3.01, 11.2]	< 0.0001
N Stage
N0	536	4988	23.0 [21.2, 40.2+]	81 [77, 85]	58 [50, 67]	1.0 (ref)	–
Nx or Unknown	78	869	31.2 [26.5, 40.2+]	90 [82, 99]	78 [66, 93]	0.54	[0.29, 1.01]	0.05
N1	261	2023	10.5 [8.94, 12.4]	54 [48, 62]	25 [18, 34]	3	[2.32, 3.86]	< 0.0001
M Stage
M0	767	7235	22.5 [19.7, 27.1]	80 [76, 83]	54 [48, 61]	1.0 (ref)	–
M1	108	646	6.61 [5.09, 8.94]	35 [26, 47]	15 [8, 30]	4.71	[3.53, 6.28]	< 0.0001
Grade
G1	448	4383	23.0 [19.7, 32.2+]	81 [77, 86]	55 [48, 64]	1.0 (ref)	–
G2	321	2325	12.9 [12.1, 18.07]	64 [58, 71]	35 [25, 48]	2.07	[1.57, 2.72]	< 0.0001
G3	24	96	3.04 [1.67, 32.2+]	23 [8, 61]	23 [8, 61]	9.31	[5.42, 16.0]	< 0.0001
Grade, Proposed
G1	448	4383	23.0 [19.7, 32.2+]	81 [77, 86]	55 [48, 64]	1.0 (ref)	–
G2a	270	2060	16.8 [12.4, 21.4]	68 [61, 76]	38 [28, 52]	1.78	[1.33, 2.39]	0.0001
G2b	51	264	7.88 [5.93, 32.2+]	44 [29, 66]	11 [2, 63]	5.09	[3.21, 8.06]	< 0.0001
G3	24	96	3.04 [1.67, 32.2+]	23 [8, 61]	23 [8, 61]	9.7	[5.64, 16.7]	< 0.0001
Insulinoma
No	781	6923	17.9 [15.5, 21.5]	72 [68, 76]	45 [39, 52]	1.0 (ref)	–
Yes	94	957	27.1 [23.0, 40.2+]	92 [85, 99]	80 [69, 94]	0.34	[0.19, 0.62]	0.0003
Sclerosing Variant
No	824	7423	18.9 [16.7, 23.0]	73 [69, 77]	48 [43, 54]	1.0 (ref)	–
Yes	51	458	31.2 [14.6, 40.2+]	91 [83, 100]	72 [50, 100]	0.47	[0.22, 1.00]	0.05
Cystic Variant
No	786	7018	18.1 [15.6, 23.9]	72 [69, 76]	48 [42, 54]	1.0 (ref)	–
Yes	89	863	NR	86 [78, 94]	64 [48, 85]	0.61	[0.38, 0.99]	0.05
Vascular Invasion
No	646	6288	23.0 [19.7, 40.2+]	80 [76, 83]	56 [50, 63]	1.0 (ref)	–
Yes	211	1316	10.3 [8.26, 14.3]	51 [43, 60]	19 [10, 35]	3.00	[2.31, 3.91]	< 0.0001
Perineural Invasion
No	577	5526	23.9 [21.5, 40.2+]	80 [76, 84]	58 [51, 65]	1.0 (ref)	–
Yes	286	2150	12.2 [10.4, 15.5]	61 [54, 68]	26 [17, 40]	2.42	[1.88, 3.12]	< 0.0001
Margins
Negative	787	7130	21.4 [18.1, 27.1]	76 [72, 80]	52 [46, 59]	1.0 (ref)	–
Positive	82	661	11.0 [8.48, 15.5]	52 [41, 66]	27 [17, 45]	2.18	[1.56, 3.04]	< 0.0001
Margins*
Negative	744	6755	19.7 [17.2, 23.9]	75 [72, 79]	50 [43, 57]	1.0 (ref)	–
Positive	56	390	8.68 [5.59, 12.1]	39 [26, 57]	15 [6, 35]	3.10	[2.18, 4.43]	< 0.0001

*Excludes patients who had an enucleation procedure. Abbreviations: Median survival is shown in 10- and 20-year OS are the respective survival probabilities from the Kaplan-Meier estimate for each subgroup, HR=Hazard Ratio, estimated from univariate Cox proportional hazards regression models, NR=Median survival not reached Median OS is shown in years, OS=Overall survival; Calculated from the date of surgery to last follow-up or death, PY=Person-years

Non-linear models for the association between tumor size and survival revealed a stronger positive association between 0 and 4 cm that reduces but remains positive without an additional inflection point for tumors >4 cm (Fig. 1B).

Subdividing Grade 2 PanNETs

We found that patients with G1 PanNETs had 10-year OS of 81%, 95% CI: [77%, 86%], G2a PanNETs 68%, 95% CI: [61%, 76%], G2b PanNETs 44%, 95% CI: [29%, 66%], and G3 PanNETs 23%, 95% CI: [8%, 61%] (Fig. 1C and Table 2). When compared to patients with G2b disease, patients with G2a had a better overall survival (HR = 0.35, 95% CI: [0.22–0.56], p < 0.001) while patients with G3 disease did worse (HR = 1.91, 95% CI: [1.01–3.59], p = 0.05).

Plotting Ki-67 versus the HR (Fig. 1D) revealed a slight change to the slope of the curve at around a Ki-67 of 10%, supporting the introduction of a 10% threshold in the grading system.

Prognosticators in M0, N0M0 and T1N0M0 Cohorts

As the most significant surgical decisions present in patients free of metastases at diagnosis, we estimated OS in lower stage subgroups, including patients who were M0 at surgery (Supplemental Table 1), N0M0 at surgery (Supplemental Table 2) and T1N0M0 at surgery (Table 3). In particular, 251 patients had tumors < 2 cm and N0M0 disease. These patients lived a median of 27 years after surgery, and their ten-year overall survival was 88%, 95% CI: [83%, 93%]. Both positive margins (excluding enucleations, HR 19.5, 95% CI: [5.49, 66.8], p <0.0001) and vascular invasion, HR 5.7, 95% CI: [1.65, 19.8], p=0.006) were associated with poor outcome in this group.

Table 3.

Overall Survival in Patients with T1N0M0 Disease: Estimates of survival after surgery for patients with M0, N0, and T1 disease, overall and according to patient subgroups defined at the time of surgery, from univariate Cox proportional hazards models

	N	Follow-Up, PY	Median [95% CI]	10-year OS	20-year OS	HR	95% CI	P
Whole Cohort	251	2443	27.0 [22.5, 36.8+]	88 [83, 93]	65 [52, 80]
Age, years
< 60	143	1504	27.1 [23.0, 36.8+]	93 [89, 98]	69 [54, 90]	1.0 (ref)	–
60+	108	938	22.5 [15.6, 36.8+]	79 [70, 90]	58 [41, 82]	2.63	[1.34, 5.15]	0.005
Sex
Female	129	1254	NR	87 [80, 94]	64 [47, 87]	1.0 (ref)	–
Male	122	1189	22.5 [17.0, 36.8+]	89 [82, 96]	65 [50, 86]	1.05	[0.55, 2.00]	0.89
Grade
G1	175	1682	23.0 [22.5, 27.1+]	88 [83, 94]	66 [51, 85]	1.0 (ref)	–
G2	48	367	17.0 [14.4, 27.1+]	83 [69, 100]	33 [8, 100]	1.38	[0.56, 3.44]	0.49
Grade, Proposed
G1	175	1682	23.0 [22.5, 27.1+]	88 [83, 94]	66 [51, 85]	1.0 (ref)	–
G2a	46	360	17.0 [14.4, 27.1+]	89 [75, 100]	35 [8, 100]	0.94	[0.32, 2.76]	0.92
G2b	2	6	3.41 [0.89, 27.1+]	0 [13, 100]	0 [13, 100]	35.7	[7.56, 168.1]	< 0.0001
Insulinoma
No	216	2021	22.5 [18.1, 36.8+]	86 [80, 92]	55 [39, 77]	1.0 (ref)	–
Yes	35	422	27.1 [23.0, 36.8+]	100 [100, 100]	94 [83, 100]	0.22	[0.05, 0.93]	0.04
Sclerosing Variant
No	214	2120	23.0 [22.5, 36.8+]	87 [82, 93]	64 [51, 80]	1.0 (ref)	–
Yes	37	322	NR	91 [82, 100]	80 [60, 100]	0.92	[0.32, 2.61]	0.88
Cystic Variant
No	219	2099	23.0 [22.5, 36.8+]	89 [83, 94]	64 [51, 82]	1.0 (ref)	–
Yes	32	343	NR	82 [68, 98]	69 [52, 93]	1.40	[0.61, 3.20]	0.42
Vascular Invasion
No	231	2240	23.0 [22.7, 31.5+]	88 [83, 94]	65 [51, 82]	1.0 (ref)	–
Yes	12	64	NR	55 [26, 100]	55 [26, 100]	5.70	[1.65, 19.8]	0.006
Perineural Invasion
No	200	1896	23.0 [22.5, 31.5+]	86 [80, 92]	64 [50, 83]	1.0 (ref)	–
Yes	44	417	NR	91 [83, 100]	57 [33, 97]	1.13	[0.49, 2.60]	0.77
Margins
Negative	242	2367	23.0 [22.5, 36.8+]	89 [84, 94]	65 [52, 81]	1.0 (ref)	–
Positive	8	44	9.22 [5.59, 36.8+]	33 [7, 100]	33 [7, 100]	6.62	[1.98, 22.1]	0.002
Margins*
Negative	233	2295	23.0 [22.5, 36.8+]	89 [84, 94]	64 [51, 80]	1.0 (ref)	–
Positive	4	19	5.59 [0.89, 36.8+]	0 [8, 100]	0 [8, 100]	19.2	[5.49, 66.8]	< 0.0001

*Excludes patients who had an enucleation procedure. Abbreviations: Median survival is in years 10-and 20-year OS are the respective survival probabilities from the Kaplan-Meier estimate for each subgroup, HR=Hazard Ratio, estimated from univariate Cox proportional hazards regression models, NR=Median survival not reached, Median OS is shown in years, OS=Overall survival; Calculated from the date of surgery to last follow-up or death, PY=Person-years

The differences in OS between G2a and G2b persisted for patients with M0 or N0M0 disease, but they were not statistically significant for these disease subgroups for comparing G3 to G2b. There were too few (N=2) patients with G2b disease among the T1N0M0 cohort to make any comparisons.

Multivariate Analyses

Multivariate analyses of OS were performed on four cohorts (all patients, patients with M0 disease, patients with N0M0 disease, and patients with T1N0M0 disease) (Supplemental Table 3) using the LASSO method. For each cohort, the following candidate variables were offered to the model selection: age, sex, year of surgery, tumor size, Ki67 (measured continuously), presence of insulinoma, sclerosing variant, cystic variant, perineural invasion or positive margins, T stage, N stage, and M stage. For the whole cohort, M stage (M1 vs M0, HR 2.46, 95% CI: [1.85, 3.30], p<0.001), perineural invasion (HR 1.62, 95% CI: [1.24, 2.09], p=0.002), N stage (N1 vs N0, HR 1.56, 95% CI: [1.19, 2.02], p=0.005), age (per 5 years) (HR 1.26, 95% CI: [1.20, 1.32], p <0.001), Ki-67 (HR 1.06, 95% CI: [1.04, 1.09], p <0.001), tumor size (per 1 cm) (HR 1.04, 95% CI: [1.01, 1.12], p=0.023), and year of surgery (per 1 year) (HR 0.96, 95% CI: [0.93, 0.97], p<0.001) were selected by the LASSO. Sex, vascular invasion and positive margins were also retained but not statistically significant. For the T1N0M0 cohort, age (per 5 years) (HR 1.38, 95% CI: [1.18,1.59], p<0.001), year of surgery (HR 0.97, 95% CI: [0.91, 1.25], p = 0.556) and Ki-67 labeling (HR 1.15, 95% CI: [1.08,1.23], p<0.001) were retained.

Cause-Specific Survival

Information on cause-specific survival for patients with M0 disease and M0N0 disease are presented in Supplemental Tables 4 and 5 respectively. With the exception of patient age (HR 1.29, 95% CI: [0.79, 2.11], p=0.31) and patient sex (HR 1.30, 95% CI: [0.80,2.10], p=0.30), all variables identified as statistically associated with OS remained significant. These results suggest that the patient age and sex findings observed with OS were due to the long follow-up obtained in this study, and not tumor-specific biological drivers.

Cumulative Incidence of Relapse

Next, we estimated the cumulative probabilities of disease relapse at 2, 5, and 10 years after surgery for the M0, M0N0 and T1N0M0 cohorts (Table 4 and Supplemental Tables 6 and 7). In the M0 cohort, tumor size (HR 6.19, 95% CI: [3.90, 9.85], p<0.001), vascular invasion (HR 4.41, 95% CI: [3.18, 6.12], p<0.001), and perineural invasion (HR 3.17, 95% CI: [2.31, 4.35], p<0.001) remained poor prognosticators, while insulinoma (HR 0.26, 95% CI: [0.11, 0.63], p=0.003), sclerosing variant (HR 0.27, 95% CI: [0.09, 0.84], p=0.02), and cystic variant (HR 0.29, 95% CI: [0.13, 0.65], p=0.003) were associated with improved outcome. T stage and N stage were also significantly associated with relapse.

Table 4.

Cumulative Incidence of Relapse: Estimates of cumulative probability of relapse at 2, 5 and 10 years after surgery, accounting for death before relapse as a competing event, for the M0 cohort and according to patient subgroups defined at the time of surgery

	N	N Relapses	N Deaths	CIR, 2y	CIR, 5y	CIR, 10y	HR (95% CI)	P
Whole Cohort	723	154	91	0.07 (0.05, 0.08)	0.17 (0.14, 0.2)	0.23 (0.19, 0.26)
Age, years
< 60	409	88	35	0.06 (0.04, 0.08)	0.17 (0.13, 0.21)	0.23 (0.19, 0.28)	1.0 (ref)
60+	314	66	56	0.07 (0.04, 0.1)	0.18 (0.13, 0.22)	0.22 (0.17, 0.27)	1.01 (0.74, 1.39)	0.93
Sex
Female	349	73	36	0.06 (0.03, 0.09)	0.17 (0.12, 0.21)	0.22 (0.17, 0.27)	1.0 (ref)
Male	374	81	55	0.07 (0.04, 0.1)	0.18 (0.14, 0.22)	0.24 (0.19, 0.28)	1.06 (0.78, 1.46)	0.7
Tumor Size
< 2 cm	308	20	37	0.01 (0, 0.02)	0.05 (0.02, 0.07)	0.07 (0.04, 0.1)	1.0 (ref)
2 cm +	415	134	54	0.11 (0.08, 0.14)	0.27 (0.22, 0.32)	0.35 (0.3, 0.4)	6.19 (3.90, 9.85)	< 0.001
T Stage
T1	308	20	37	0.01 (0, 0.02)	0.05 (0.02, 0.07)	0.07 (0.04, 0.1)	1.0 (ref)
T2	265	69	35	0.08 (0.05, 0.12)	0.22 (0.16, 0.27)	0.29 (0.23, 0.35)	4.87 (2.98, 7.95)	< 0.001
T3	139	60	17	0.16 (0.09, 0.22)	0.34 (0.26, 0.43)	0.45 (0.36, 0.54)	8.66 (5.24, 14.31)	< 0.001
T4	11	5	2	0.1 (0, 0.3)	0.51 (0.17, 0.84)	0.51 (0.17, 0.84)	9.97 (3.73, 26.66)	< 0.001
N Stage
N0	481	71	57	0.03 (0.01, 0.05)	0.11 (0.08, 0.14)	0.16 (0.13, 0.2)	1.0 (ref)
Nx or Unknown	70	2	8	0 (0, 0)	0 (0, 0)	0.02 (0, 0.06)	0.18 (0.05, 0.71)	0.01
N1	172	81	26	0.19 (0.13, 0.25)	0.41 (0.33, 0.48)	0.49 (0.4, 0.57)	4.1 (2.99, 5.64)	< 0.001
Grade
G1	397	44	51	0.03 (0.01, 0.05)	0.08 (0.05, 0.11)	0.11 (0.08, 0.14)	1.0 (ref)
G2	246	86	26	0.1 (0.06, 0.14)	0.31 (0.24, 0.37)	0.44 (0.36, 0.51)	4.5 (3.13, 6.48)	< 0.001
G3	16	11	4	0.25 (0.03, 0.47)	0.69 (0.44, 0.94)	0.69 (0.44, 0.94)	11 (5.24, 23.07)	< 0.001
Grade, Proposed
G1	397	44	51	0.03 (0.01, 0.05)	0.08 (0.05, 0.11)	0.11 (0.08, 0.14)	1.0 (ref)
G2a	213	65	24	0.08 (0.04, 0.12)	0.25 (0.19, 0.32)	0.39 (0.31, 0.47)	3.81 (2.6, 5.58)	< 0.001
G2b	33	21	2	0.23 (0.08, 0.38)	0.62 (0.43, 0.8)	0.75 (0.53, 0.97)	10.82 (6.37, 18.36)	< 0.001
G3	16	11	4	0.25 (0.03, 0.47)	0.69 (0.44, 0.94)	0.69 (0.44, 0.94)	11.07 (5.25, 23.35)	< 0.001
Insulinoma
No	642	149	83	0.07 (0.05, 0.09)	0.18 (0.15, 0.22)	0.24 (0.21, 0.28)	1.0 (ref)
Yes	81	5	8	0.01 (0, 0.04)	0.07 (0, 0.13)	0.09 (0.01, 0.16)	0.26 (0.11, 0.63)	0.003
Sclerosing Variant
No	677	151	86	0.07 (0.05, 0.09)	0.18 (0.15, 0.21)	0.24 (0.2, 0.27)	1.0 (ref)
Yes	46	3	5	0 (0, 0)	0.05 (0, 0.11)	0.08 (0, 0.18)	0.27 (0.09, 0.84)	0.02
Cystic Variant
No	642	148	81	0.07 (0.05, 0.09)	0.19 (0.16, 0.22)	0.25 (0.21, 0.29)	1.0 (ref)
Yes	81	6	10	0.01 (0, 0.04)	0.04 (0, 0.09)	0.06 (0, 0.12)	0.29 (0.13, 0.65)	0.003
Vascular Invasion
No	558	86	69	0.03 (0.02, 0.05)	0.1 (0.08, 0.13)	0.16 (0.12, 0.19)	1.0 (ref)
Yes	151	65	20	0.18 (0.12, 0.25)	0.45 (0.36, 0.54)	0.53 (0.43, 0.62)	4.41 (3.18, 6.12)	< 0.001
Perineural Invasion
No	500	73	61	0.03 (0.02, 0.05)	0.11 (0.08, 0.14)	0.16 (0.12, 0.19)	1.0 (ref)
Yes	211	80	29	0.14 (0.09, 0.19)	0.33 (0.26, 0.4)	0.4 (0.33, 0.48)	3.17 (2.31, 4.35)	< 0.001
Margins
Negative	662	137	78	0.06 (0.04, 0.08)	0.17 (0.14, 0.2)	0.22 (0.19, 0.26)	1.0 (ref)
Positive	55	16	12	0.13 (0.04, 0.23)	0.26 (0.14, 0.38)	0.31 (0.18, 0.45)	1.47 (0.85, 2.54)	0.17
Margins*
Negative	619	137	76	0.06 (0.04, 0.08)	0.18 (0.14, 0.21)	0.23 (0.2, 0.27)	1.0 (ref)
Positive	30	15	7	0.21 (0.06, 0.36)	0.42 (0.24, 0.61)	0.51 (0.31, 0.71)	2.78 (1.58, 4.88)	< 0.001

*Excludes patients who had an enucleation procedure. Abbreviations: CIR Cumulative Incidence of Relapse; Calculated from the date of surgery to last follow-up, date of relapse or death, HR Hazard Ratio, estimated from univariate competing risks regression models

As with OS, the cumulative incidence of relapse was associated with refined tumor grade. In the M0 cohort, patients with G1 PanNETs had 5-year cumulative incidence of relapse of 8%, 95% CI: [5%,11%], those with G2a PanNETs 25%, 95% CI: [19%, 32%], those with G2b PanNETs 62%, 95% CI: [43%, 80%], and those with G3 PanNETs 69%, 95% CI: [44%, 94%] (Table 4). These results again support the inclusion of a 10% threshold in the grading system.

For the 251 patients with T1N0M0 disease, 10 were missing information on recurrence, leaving 241 patients. Of these 241, 15 recurred and 29 died without relapse. The cumulative probability of relapse by 10 years was 6%, 95% CI: [3%, 10%] (Supplemental Table 7). Only three of the 251 patients died of their disease, and the disease-specific survival at 20 years was 96%, 95% CI: [92%, 100%].

Year and Age of Surgery

A comparison of the distribution of patients diagnosed with stage I, II, III, or IV disease by year of surgery grouped into quintiles revealed a down shift in stage at diagnosis over time (Fig. 1E). Over the study period, patients were more likely to be diagnosed with stage I (Odds Ratio (OR) 1.61, 95% CI: 1.34, 1.93, p <0.001), stage II (OR 1.30, 95% CI: 1.08, 1.55, p =0.005), or stage III disease (OR 1.2, 95% CI: 1.00, 1.45, p=0.054) compared to stage IV disease.

Hypothesizing that a down shift in stage over time would be associated with an earlier age at diagnosis, we next examined trends in age at surgery by year of surgery (Supplemental Fig. 2). Surprisingly, and perhaps reflecting regional referral practices, we observed a slight increase in the mean age at diagnosis over time, with mean age increasing 1.08 years every five years (95% CI: 0.51, 1.65, p<0.001).

Discussion

The incidence of PanNETs is increasing [1–7]. Fortunately, new therapeutic options beyond surgery are available. As clinical trials are designed to define the impact of novel therapies and as clinicians struggle to determine the best therapy for their individual patients, it is important that prognosticators are identified and that their impact on patient survival is quantified accurately.

In this single-institution series of 904 patients with surgically resected PanNETs and extensive follow-up, we provide support for the proposed separation of G2 PanNETs (Ki-67 of 3-≤ 20%) into two grades (G2a for Ki-67 3- <10%, and G2b for Ki-67 of 10%- ≤20%) [41, 44, 45, 59–63]. Here, the proposed grading system clearly stratified the patients for both cumulative incidence of relapse and for OS (Fig. 1C). Furthermore, analysis of the non-linear relationship between Ki-67 and OS (Fig. 1D) revealed a subtle inflection point near a Ki-67 labeling index of 10%. While the separation of patients is clear, we should emphasize that the relationship between Ki-67 and recurrence and OS is generally fairly linear, and all cut-offs are therefore somewhat arbitrary. Nonetheless, as has been suggested, the division of G2 tumors into two grades (G2a and G2b) may have therapeutic implications [44, 45, 59–62, 64].

We also confirm the major prognosticators and refine the magnitude of risk associated with each. The highest HRs for recurrence and OS were associated with stage at surgery (T, N and M stages), vascular invasion, perineural invasion, and positive margins. Separately analyzing the influence of margin status in a cohort excluding enucleations, as enucleations are primarily performed on low-risk tumors, demonstrates an even greater impact of margin status on prognosis.

Here we show that the serotonin-positive duct-centric sclerosing variant of PanNET comprises 6% of surgically resected PanNETs, and, as has been reported, that patients with this tumor type have an excellent prognosis with a five-year cumulative incidence of relapse of 5% (95% CI: 0%, 11%) and a 10-year OS of 91% (95% CI: 83%, 100%) [33–35, 48–50]. This finding has clinical implications as these tumors often involve and stricture the main pancreatic duct causing ductal dilatation and can therefore be detected on imaging as ductal dilatation abruptly beginning at a small enhancing mass lesion on computed tomography [33–35, 48–50]. We confirm that insulinomas and cystic PanNETs are associated with a lower risk of recurrence and improved OS [7, 16–20]. Further, we confirm that patients with T1N0M0 disease have excellent outcomes after surgery, with a disease-specific survival at 20 years of 96% (95% CI: 92%, 100%). We also confirm, in a finding that will improve risk stratification, that grade, vascular invasion and positive margins are important prognosticators of recurrence and OS in the critical group of patients with T1N0M0 disease [42].

Finally, we examined changes in stage at diagnosis and patient age at diagnosis over time (Fig. 1E and Supplemental Fig. 1). The trend towards a lower stage at diagnosis supports the hypothesis that a growing number of asymptomatic patients are being diagnosed incidentally on imaging performed for another indication [1, 46]. However, we did not observe a trend towards younger age at the time of diagnosis over time. In interpreting these trends, one should note that the patients included in this study, who underwent surgery at a high-volume tertiary care center, may not be representative of the entire population of patients with a PanNET.

A weakness of the study is the use of two methods to determine the Ki-67 labeling indexes of tumors. It should be noted, however, that we found a strong positive linear relationship between the two methods, and both methods have been used in previous studies [23, 51, 65, 66]. Finally, OS was used in our initial analyses as it provided, compared to disease-specific survival, the most complete dataset and the exact date of death was clear, while accuracy of the date of recurrence depends on the frequency of surveillance. Because the follow-up on many of the patients was so long, as demonstrated in the cause-specific survival analyses, caution should be taken in interpreting the risk associated with age and sex in OS, as these could be attributable to gender- and age-based differences in mortality in the general population.

Accurate risk stratification of PanNETs will help predict the clinical behavior of these tumors and increase the precision with which future treatment efficacy is assessed. The prognostic significance of grade (Ki-67 index) will help inform the design and interpretation of clinical trials, while other prognostic factors reported here such as histologic subtype or invasive behavior may inform future basic science inquiry into these increasingly common pancreatic tumors.

Author Contributions

ALK, EDY, ALB, JH and RHH conceived of the presented idea and wrote the initial draft of the manuscript. ALB prepared Fig. 1 content and performed the statistical analysis and verified the analytical methods. ALK, EDY, PW, ZG, UAE, IR, YS, LD, VM, GK, AK and RHH reviewed and annotated the cases and histologic material. RAB, WRB, JLC, KL, CS and JH contributed and reviewed cases and encouraged the investigation of the Ki67 analysis in small tumors. All authors discussed the results and contributed to the final manuscript.

Data Availability

No datasets were generated or analysed during the current study.

Declarations

The authors declare that they do not have any financial or non-financial interests that are directly or indirectly related to the work submitted for publication.

Competing interests

The authors declare no competing interests.