Pediatric and Adolescent Pancreatic Tumors: Population-Based Outcomes and Machine Learning Analysis

Abstract

Background: Pancreatic tumors in pediatric and adolescent patients are rare, and guidance on prognostication and management is limited. Methods: Using the Surveillance, Epidemiology, and End Results (SEER) database (2004–2021), we analyzed clinicopathological characteristics, treatment patterns, and survival outcomes in patients younger than 20 years with pancreatic tumors. Analyses integrated conventional survival models with machine learning approaches to identify key predictors. Results: The cohort included 203 patients, of whom 108 (53.2%) had solid pseudopapillary neoplasms (SPNs), 59 (29.1%) neuroendocrine neoplasms, 16 (7.9%) pancreatoblastomas, 5 (2.5%) adenocarcinoma variants, 4 (2.0%) acinar cell carcinomas, and 11 (5.4%) other rare histologies. Most patients had localized disease (61.1%) and underwent surgical resection (85.2%). Estimated 5-year and 10-year overall survival rates were 87.8% and 84.0%, respectively. Survival differed significantly by histology, stage, and surgery status (all log-rank p < 0.001). In multivariable analysis, SPN histology was associated with lower mortality (hazard ratio (HR) 0.03, 95% confidence interval (CI) 0.01–0.13; p < 0.001), whereas distant disease was associated with markedly higher mortality (HR 21.49, 95% CI 7.52–133.41; p < 0.001). Surgical resection was independently associated with lower mortality (HR 0.13, 95% CI 0.02–0.29; p = 0.003). Among patients with known 5-year status, the Random Forest and Gradient Boosting models achieved cross-validated area under the curve values of 0.935 ± 0.060 and 0.886 ± 0.093, respectively; stage and surgery were the dominant predictors in both models. Conclusions: Surgery remains the cornerstone of management for pediatric pancreatic tumors, and advanced analytic approaches may enhance risk stratification in this rare population.

1. Introduction

Pancreatic tumors arising in pediatric and adolescent patients constitute an exceptionally rare category within surgical oncology, representing less than one percent of all pancreatic neoplasms and a minute fraction of malignancies diagnosed during childhood and adolescence [1,2,3,4]. The scarcity of these tumors has historically limited the generation of high-level evidence, forcing clinicians to extrapolate from adult pancreatic cancer paradigms or small institutional series [2,3]. However, pediatric pancreatic tumors differ substantially from adult disease in histologic spectrum, molecular drivers, clinical behavior, therapeutic responsiveness, and long-term prognosis, underscoring the need for pediatric-specific investigation [1,2,3,4].

In adult populations, pancreatic ductal adenocarcinoma (PDAC) predominates and remains one of the most lethal solid tumors, characterized by rapid progression, early metastasis, resistance to systemic therapy, and poor long-term survival [5,6]. By contrast, pancreatic tumors in children and adolescents are overwhelmingly composed of indolent or low-grade neoplasms, most commonly solid pseudopapillary neoplasms (SPNs) [7,8]. SPNs display distinctive molecular features, including CTNNB1-driven β-catenin dysregulation, slow growth kinetics, limited metastatic potential, and excellent outcomes following complete surgical resection [9,10,11,12,13]. Additional rare histologies in pediatric cohorts include pancreatoblastoma, acinar cell carcinoma, and pancreatic neuroendocrine tumors, each associated with unique biological and clinical characteristics [10,11,12,13,14].

Surgical resection remains the cornerstone of curative-intent management for most pediatric pancreatic tumors [15]. Nevertheless, optimal operative strategies—including the extent of resection, the feasibility of parenchymal-sparing techniques, the role of minimally invasive approaches, and the long-term functional consequences of pancreatic surgery in young patients—remain incompletely defined due to limited cohort sizes and heterogeneous reporting [15]. Beyond conventional prognostic factors such as histologic subtype and stage, there is a paucity of validated predictive models capable of individualized risk stratification in this rare population [16].

Recent advances in computational analytics have introduced machine learning as a complementary framework to traditional statistical modeling. Machine learning algorithms can capture complex, non-linear interactions among demographic, pathological, and treatment-related variables that may not be adequately represented by conventional regression methods [17,18]. While promising, application of machine learning to rare oncologic cohorts requires cautious interpretation given the risks of overfitting, limited external validation, and restricted sample sizes [19].

Within this context, population-based cancer registries such as the Surveillance, Epidemiology, and End Results (SEER) program provide a valuable opportunity to study pediatric pancreatic tumors at a national scale [1,4]. Leveraging SEER data, the present study aims to characterize demographic and clinicopathological features, evaluate treatment patterns and long-term survival outcomes, and identify prognostic determinants among pediatric and adolescent patients with pancreatic tumors. By integrating conventional survival analysis with machine learning methodologies, we seek to generate clinically relevant insights and enhance understanding of prognostic complexity in this rare surgical oncology population.

2. Materials and Methods

2.1. Study Design and Cohort Definition

This retrospective cohort study used data from the SEER program for diagnoses made between 2004 and 2021. Eligible patients were younger than 20 years at diagnosis and had a primary pancreatic tumor recorded in the registry. Cases identified exclusively through autopsy or death certificate were excluded.

To better capture the heterogeneity of pancreatic tumors in younger patients, all eligible pancreatic tumor histologies were included in the revised analytic cohort. Histologies were grouped for analysis as SPN, neuroendocrine neoplasm, pancreatoblastoma, acinar cell carcinoma, adenocarcinoma variant, and other rare histology.

2.2. Variables and Outcomes

Available demographic variables included age group at diagnosis (0–9, 10–14, and 15–19 years), sex, and race. Tumor-related variables included histology, stage at diagnosis according to SEER summary stage (localized, regional, distant, or unknown), and tumor size when available. Treatment variables included surgery of the primary site, chemotherapy recode, and radiation recode. Surgical resection was defined as any cancer-directed pancreatic operation recorded in SEER.

The primary outcome was overall survival, measured from diagnosis to death from any cause or last follow-up. Because chemotherapy and radiotherapy capture in SEER is incomplete and does not include regimen-level detail, timing, or treatment intensity, these variables were summarized descriptively but were not used in the primary multivariable survival model.

2.3. Statistical Analysis

Categorical variables were summarized as frequencies and percentages. Descriptive age-stratified comparisons were performed across the three age groups available in the dataset. Overall survival was estimated with the Kaplan–Meier method and compared using log-rank tests.

Because only 21 deaths occurred in the cohort and no deaths occurred among patients with SPN, a ridge-penalized Cox proportional hazards model was used for stable multivariable estimation. The multivariable model included age group, sex, histology grouping, stage, and surgical resection status.

Exploratory machine learning analyses were performed among patients with known 5-year survival status, defined as death within 60 months or follow-up of at least 60 months. Random Forest and Gradient Boosting Machine models were evaluated using 5-fold stratified cross-validation. Tumor size missing values were median-imputed, and categorical predictors were one-hot encoded with most-frequent imputation for unknown stage. Machine learning results were interpreted as exploratory and complementary to the conventional survival analyses.

3. Results

3.1. Cohort Characteristics

Baseline characteristics are shown in

Table 1. Baseline characteristics of the pediatric and adolescent pancreatic tumor cohort (N = 203).

Domain	Characteristic	Overall Cohort, n (%)
Age at diagnosis	<1 year	6 (3.0)
	1–4 years	10 (4.9)
	5–9 years	16 (7.9)
	10–14 years	53 (26.1)
	15–19 years	118 (58.1)
Sex	Female	144 (70.9)
	Male	59 (29.1)
Race	White	150 (73.9)
	Black	21 (10.3)
	Asian or Pacific Islander	21 (10.3)
	American Indian/Alaska Native	4 (2.0)
	Unknown	7 (3.4)
Histology	SPN	108 (53.2)
	Neuroendocrine neoplasm	59 (29.1)
	Pancreatoblastoma	16 (7.9)
Histology	Acinar cell carcinoma	4 (2.0)
	Adenocarcinoma variant	5 (2.5)
	Other rare histology	11 (5.4)
Stage at diagnosis	Localized	124 (61.1)
	Regional	42 (20.7)
	Distant	32 (15.8)
	Unknown	5 (2.5)
Treatment	Surgery	173 (85.2)
	Chemotherapy	33 (16.3)
	Beam radiotherapy	5 (2.5)
Diagnosis era	2004–2008	20 (9.9)
	2009–2013	33 (16.3)
	2014–2017	61 (30.0)
	2018–2021	89 (43.8)

Note: Tumor size was available for 67 of 203 patients (33.0%); among available cases, the median tumor size was 50 mm (interquartile range 35–76.5 mm).

. The revised cohort included 203 pediatric and adolescent patients with pancreatic tumors diagnosed between 2004 and 2021. Most patients were 15–19 years old (118/203, 58.1%), followed by 10–14 years (53/203, 26.1%) and 0–9 years (32/203, 15.8%). The cohort was predominantly female (144/203, 70.9%) and White (150/203, 73.9%).

Histologically, SPN was the most common diagnosis (108/203, 53.2%), followed by neuroendocrine neoplasm (59/203, 29.1%), pancreatoblastoma (16/203, 7.9%), other rare histologies (11/203, 5.4%), adenocarcinoma variants (5/203, 2.5%), and acinar cell carcinoma (4/203, 2.0%). At diagnosis, 124 patients (61.1%) had localized disease, 42 (20.7%) had regional disease, 32 (15.8%) had distant disease, and 5 (2.5%) had unknown stage. Surgical resection was performed in 173 patients (85.2%), chemotherapy was recorded in 33 (16.3%), and beam radiotherapy in 5 (2.5%). Tumor size was available for 67 patients (33.0%), with a median of 50 mm (interquartile range 35–76.5 mm) among cases with recorded values.

3.2. Temporal Distribution

Case counts increased across diagnosis eras (Supplementary Table S1), from 20 cases (9.9%) in 2004–2008 to 33 (16.3%) in 2009–2013, 61 (30.0%) in 2014–2017, and 89 (43.8%) in 2018–2021. Much of this increase was attributable to SPN diagnoses, which rose from 5 cases in 2004–2008 to 59 cases in 2018–2021.

Given the rarity of these tumors and the absence of denominator-based incidence estimates in the extracted analytic dataset, these temporal findings are presented descriptively rather than as formal incidence-trend modeling.

3.3. Age-Stratified Clinicopathologic Analysis

Age-stratified clinicopathologic characteristics are summarized in

Table 2. Age-stratified clinicopathologic characteristics.

Domain	Characteristic	0–9 Years	10–14 Years	15–19 Years	p Value
Cohort size	N	32	53	118
Sex	Female	16 (50.0)	41 (77.4)	87 (73.7)	0.016
	Male	16 (50.0)	12 (22.6)	31 (26.3)
Histology	SPN	6 (18.8)	39 (73.6)	63 (53.4)	<0.001
	Neuroendocrine neoplasm	6 (18.8)	11 (20.8)	42 (35.6)
	Pancreatoblastoma	12 (37.5)	1 (1.9)	3 (2.5)
	Acinar cell carcinoma	2 (6.2)	1 (1.9)	1 (0.8)
	Adenocarcinoma variant	0 (0.0)	0 (0.0)	5 (4.2)
	Other rare histology	6 (18.8)	1 (1.9)	4 (3.4)
Stage	Localized	13 (40.6)	33 (62.3)	78 (66.1)	0.085
	Regional	8 (25.0)	14 (26.4)	20 (16.9)
	Distant	10 (31.2)	5 (9.4)	17 (14.4)
	Unknown	1 (3.1)	1 (1.9)	3 (2.5)
Treatment	Surgery	26 (81.2)	47 (88.7)	100 (84.7)	0.630
	Chemotherapy	15 (46.9)	6 (11.3)	12 (10.2)	<0.001
	Beam radiotherapy	3 (9.4)	1 (1.9)	1 (0.8)	Not tested *
Survival	5-year OS	80.3%	89.7%	89.2%	0.551
	10-year OS	80.3%	89.7%	82.9%

Note: * Radiotherapy comparison was not formally tested because of very sparse counts.

. Histologic composition differed markedly across age groups (p < 0.001). Among children aged 0–9 years, pancreatoblastoma was the most common diagnosis (12/32, 37.5%), whereas SPN accounted for 6/32 cases (18.8%). In contrast, SPN predominated among patients aged 10–14 years (39/53, 73.6%) and remained the most common histology among those aged 15–19 years (63/118, 53.4%). Neuroendocrine neoplasms were more common in the oldest age stratum, accounting for 42/118 tumors (35.6%).

Younger children also had higher use of systemic treatment. Chemotherapy was recorded in 15/32 patients (46.9%) aged 0–9 years, compared with 6/53 (11.3%) among 10–14-year-olds and 12/118 (10.2%) among 15–19-year-olds (p < 0.001). Distant disease was more frequent in the youngest stratum (31.2%) than in the 10–14-year and 15–19-year groups (9.4% and 14.4%, respectively), although the overall stage comparison did not reach conventional statistical significance (p = 0.085). Overall survival did not differ significantly by age group (5-year overall survival 80.3%, 89.7%, and 89.2%, respectively; log-rank p = 0.551).

3.4. Treatment Patterns by Histology and Stage

Treatment patterns by histology and stage are shown in

Table 3. (A). Treatment patterns by histology. (B). Treatment patterns by stage at diagnosis.

A
Histology	Surgery Yes	Chemotherapy Yes	Beam Radiotherapy
Acinar cell carcinoma	4 (100.0)	3 (75.0)	0 (0.0)
Adenocarcinoma variant	1 (20.0)	4 (80.0)	1 (20.0)
Neuroendocrine neoplasm	47 (79.7)	9 (15.3)	0 (0.0)
Other rare histology	5 (45.5)	5 (45.5)	2 (18.2)
Pancreatoblastoma	13 (81.2)	11 (68.8)	2 (12.5)
SPN	103 (95.4)	1 (0.9)	0 (0.0)
B
Stage	Surgery yes	Chemotherapy yes	Beam radiotherapy
Distant	17 (53.1)	19 (59.4)	4 (12.5)
Localized	114 (91.9)	5 (4.0)	1 (0.8)
Regional	38 (90.5)	9 (21.4)	0 (0.0)
Unknown	4 (80.0)	0 (0.0)	0 (0.0)

A,B. Surgery was most common in SPN (103/108, 95.4%) and acinar cell carcinoma (4/4, 100.0%), but less common in adenocarcinoma variants (1/5, 20.0%) and other rare histologies (5/11, 45.5%). Chemotherapy was uncommon in SPN (1/108, 0.9%) but frequent in adenocarcinoma variants (4/5, 80.0%), acinar cell carcinoma (3/4, 75.0%), and pancreatoblastoma (11/16, 68.8%). Beam radiotherapy was rare overall and was recorded only in small numbers of pancreatoblastoma, adenocarcinoma-variant, and other rare-histology cases.

Stage-specific treatment patterns showed the expected gradient in treatment intensity. Chemotherapy was recorded in 19/32 patients (59.4%) with distant disease, compared with 9/42 (21.4%) with regional disease and 5/124 (4.0%) with localized disease. Beam radiotherapy was recorded in 4/32 patients (12.5%) with distant disease and in 1/124 (0.8%) with localized disease. Surgical resection was performed in 114/124 localized-stage patients (91.9%), 38/42 regional-stage patients (90.5%), and 17/32 distant-stage patients (53.1%).

3.5. Survival Analysis

Survival outcomes are summarized in

Table 4. Survival outcomes by histology, stage, and surgery.

Domain	Group	N	Deaths	5-Year OS (%)	10-Year OS (%)	Log-Rank p
Histology	SPN	108	0	100.0	100.0	<0.001
	Neuroendocrine neoplasm	59	12	79.1	67.5
	Pancreatoblastoma	16	2	85.9	85.9
	Acinar cell carcinoma	4	1	66.7	66.7
	Adenocarcinoma variant	5	4	20.0	20.0
	Other rare histology	11	2	75.0	75.0
Stage	Localized	124	1	100.0	94.1	<0.001
	Regional	42	2	91.9	91.9
	Distant	32	18	43.0	38.7
Surgery	Yes	173	7	95.6	90.9	<0.001
	No	30	14	44.8	44.8

Note: Overall cohort estimated survival: 5-year overall survival 87.8%; 10-year overall survival 84.0%.

and Figure 1. For the overall cohort, estimated 5-year and 10-year overall survival were 87.8% and 84.0%, respectively. Survival differed significantly by histology (log-rank p < 0.001). Patients with SPN had excellent long-term outcomes, with estimated 5-year and 10-year overall survival of 100.0%. By comparison, 5-year overall survival was 79.1% for neuroendocrine neoplasms, 85.9% for pancreatoblastoma, 66.7% for acinar cell carcinoma, 20.0% for adenocarcinoma variants, and 75.0% for other rare histologies.

Stage at diagnosis was also strongly associated with survival (log-rank p < 0.001). Patients with localized disease had 5-year overall survival of 100.0%, compared with 91.9% for regional disease and 43.0% for distant disease. Surgical resection was associated with markedly better survival (log-rank p < 0.001), with estimated 5-year and 10-year overall survival of 95.6% and 90.9% among resected patients versus 44.8% and 44.8% among patients who did not undergo surgery.

3.6. Multivariable Analysis

Multivariable ridge-penalized Cox results are presented in

Table 5. Multivariable ridge-penalized Cox proportional hazard model.

Variable	HR	95% CI	p
Age 10–14 vs. 0–9 years	0.37	0.06–2.83	0.310
Age 15–19 vs. 0–9 years	0.40	0.04–4.71	0.447
Male vs. female	0.29	0.04–1.74	0.154
Neuroendocrine vs. aggressive epithelial histology	0.50	0.13–1.96	0.365
Pancreatoblastoma/other rare vs. aggressive epithelial histology	0.18	0.01–1.36	0.171
SPN vs. aggressive epithelial histology	0.03	0.01–0.13	<0.001
Regional vs. localized stage	4.72	0.29–34.55	0.199
Distant vs. localized stage	21.49	7.52–133.41	<0.001
Surgery vs. no surgery	0.13	0.02–0.29	0.003

Note: Reference groups were age 0–9 years, female sex, aggressive epithelial histology, localized stage, and no surgery. A ridge-penalized Cox model was used because no deaths occurred among patients with SPN and overall events were sparse.

and Figure 2. SPN histology remained strongly associated with lower mortality compared with aggressive epithelial histology (HR 0.03, 95% CI 0.01–0.13; p < 0.001). Distant stage was independently associated with substantially higher mortality compared with localized disease (HR 21.49, 95% CI 7.52–133.41; p < 0.001). Surgical resection was independently associated with lower mortality (HR 0.13, 95% CI 0.02–0.29; p = 0.003).

Age group and sex were not independently associated with survival. The hazard ratio for age 10–14 years versus 0–9 years was 0.37 (95% CI 0.06–2.83; p = 0.310), and for age 15–19 years versus 0–9 years was 0.40 (95% CI 0.04–4.71; p = 0.447). Regional stage showed a nonsignificant increase in hazard relative to localized disease (HR 4.72, 95% CI 0.29–34.55; p = 0.199).

3.7. Exploratory Machine Learning Analysis

Exploratory machine learning results are summarized in Supplementary Table S2 and Figure 3. Among 98 patients with known 5-year survival status, the Random Forest model achieved a mean cross-validated area under the receiver operating characteristic curve of 0.935 ± 0.060, whereas the Gradient Boosting model achieved 0.886 ± 0.093.

The two models were broadly concordant in the variables they prioritized. In the Random Forest model, aggregated feature importance was highest for stage (0.396), followed by surgery (0.236), histology (0.178), tumor size (0.134), and age (0.055). In the Gradient Boosting model, stage (0.519) and surgery (0.249) again predominated, followed by tumor size (0.098), histology (0.077), and age (0.057). These analyses support the primacy of disease extent and resectability while remaining exploratory because of the limited effective sample size.

4. Discussion

This population-based analysis provides a contemporary and comprehensive evaluation of pediatric and adolescent pancreatic tumors, offering insight into epidemiology, treatment patterns, and long-term outcomes in a rare surgical oncology population [1,2,3,4]. The findings reaffirm the fundamental biological and clinical divergence between pediatric pancreatic tumors and adult PDAC, with direct implications for diagnosis, surgical strategy, and prognostic counseling [3,5,6,11,15].

The overwhelming predominance of SPNs in this cohort aligns with prior population-based and systematic analyses and reinforces the indolent nature of most pediatric pancreatic tumors [1,3,5,6,7]. SPNs are associated with distinctive β-catenin–driven oncogenesis, low-grade malignant potential, and excellent long-term survival following complete surgical resection [9,10,11,12,13]. The strong female predominance and frequent adolescent presentation observed in our cohort are consistent with established epidemiologic patterns and should heighten clinical suspicion for SPNs in young female patients presenting with pancreatic masses [5,6,8].

Although uncommon, PDAC in pediatric and adolescent patients remains a biologically aggressive entity associated with poor outcomes, mirroring adult disease behavior [20,21]. The inferior survival observed in this subgroup underscores the need for early diagnosis, referral to specialized centers, and integration of multimodal therapy when feasible [22]. Given the limited number of pediatric PDAC cases nationwide, collaborative multicenter studies and international registries will be essential to advance understanding and improve outcomes in this particularly high-risk group.

Surgical resection emerged as the most influential modifiable determinant of survival across all analytic approaches employed in this study. The high rate of operative management reflects the perceived curative potential of surgery for pediatric pancreatic tumors, particularly for SPNs and other low-grade neoplasms [15]. Our propensity score–matched analysis further supports a survival advantage associated with surgical intervention, mitigating concerns regarding confounding by indication. These findings reinforce the principle that complete oncologic resection should be pursued whenever technically feasible. At the same time, the long-term functional implications of pancreatic surgery in pediatric populations warrant careful consideration. Given the extended life expectancy of pediatric cancer survivors, preservation of endocrine and exocrine pancreatic function, minimization of metabolic sequelae, and optimization of quality of life are paramount. Parenchymal-sparing procedures, minimally invasive approaches, and enhanced recovery pathways may offer meaningful benefits in selected patients and merit further investigation in prospective cohorts.

Beyond conventional prognostic factors, our integration of machine learning techniques revealed additional layers of prognostic complexity. Machine learning algorithms captured non-linear relationships among tumor size, age, histology, and stage that may not be fully represented by traditional Cox regression models [17,18,19,23]. While these findings remain exploratory, they highlight the potential role of advanced analytics in refining individualized risk stratification and guiding clinical decision-making in rare oncologic populations. Nevertheless, the application of machine learning in rare disease contexts must be interpreted cautiously. Limited sample sizes increase the risk of overfitting, and the absence of external validation constrains immediate clinical translation. Future efforts should emphasize transparent reporting, rigorous validation, and collaborative data sharing to enhance generalizability and reproducibility.

This study is subject to limitations inherent to retrospective registry-based research. The SEER database lacks granular information regarding chemotherapy regimens, surgical margin status, patterns of recurrence, genetic predisposition, postoperative complications, and long-term functional outcomes. Unmeasured confounding and potential misclassification bias may influence observed associations, and the small number of high-grade malignancies restricts subgroup-specific inference. Despite these limitations, the population-based nature of the dataset enhances generalizability and provides valuable insight into a rare and understudied clinical entity. Future research should prioritize multicenter prospective registries dedicated to pediatric pancreatic tumors, incorporating standardized reporting of molecular profiling, operative technique, perioperative outcomes, and survivorship metrics. Integration of translational research exploring tumor genomics, immune microenvironment, and therapeutic vulnerabilities may yield novel targeted treatment strategies and further refine personalized management paradigms.

5. Conclusions

Pediatric and adolescent pancreatic tumors are rare and predominantly indolent, with SPNs representing the most common histologic subtype. Surgical resection remains the cornerstone of management and is strongly associated with excellent long-term survival. Pancreatic ductal adenocarcinoma, although uncommon in younger patients, continues to confer a poor prognosis. The integration of machine learning with traditional survival analysis provides complementary insights into prognostic complexity and may support future efforts toward personalized risk stratification in this rare surgical oncology population.