Central Precocious Puberty and Optic Pathway Glioma in Children with Neurofibromatosis 1: Associations with Tumor Location, Vision, and Treatment

Abstract

Neurofibromatosis 1-associated optic pathway gliomas (OPGs) may co-occur with central precocious puberty (CPP). Using large language model-based data analysis from 2 NF centers, we compared OPGs in children with NF1, with and without CPP. CPP was associated with hypothalamic involvement, poorer vision, and a higher likelihood of OPG treatment.

Neurofibromatosis type 1 (NF1) is a complex, multisystem neurogenetic disorder characterized by a broad spectrum of clinical manifestations and comorbidities, including benign and malignant tumors involving the central and peripheral nervous systems.¹ As such, children are at increased risk of developing a low-grade astrocytic neoplasm of the optic pathway, termed an optic pathway glioma (OPG). OPGs occur in approximately 15%-20% of affected children, which can lead to decreased visual acuity, visual field defects, or proptosis.^2,3 In addition, the proximity of the optic chiasm to the hypothalamus has been postulated as one cause for the development of central precocious puberty (CPP), growth hormone deficiency or excess, ACTH deficiency, hypogonadotropic hypogonadism, and thyrotropin deficiency.^4,5 Among these, CPP is the most common neuroendocrine feature⁶, defined as the development of secondary sexual characteristics at a younger age than expected due to early maturation of the hypothalamic-pituitary-gonadal axis.^4,7,8 Although prior studies have reported an association between CPP and OPG^9–13, it is not clear whether the co-occurrence of these 2 features of NF1 results is associated with differences in OPG clinical features or patient outcome. For this reason, we leveraged artificial intelligence (AI)-based methods, such as large language models, to compare NF1-OPGs arising in children with CPP relative to those without CPP across 2 established high-volume NF centers. We aimed to determine whether the presence of CPP in children with NF1-OPG is associated with differences in tumor characteristics, visual outcomes, and treatment patterns compared with those without CPP.

Methods

Cohort Selection

This study was approved under a waiver of consent by the Institutional Review Boards at Washington University and Lurie Children’s Hospital/Northwestern University. NF1 patients diagnosed and seen at Washington University (WU) between June 1, 2018, and May 1, 2025, and at Lurie Children’s Hospital (LCH) between January 1, 1994 and April 30, 2023, were identified. Their associated data were extracted from each institution’s Epic electronic health record (EHR) database using a common structured query language, with slight modifications to account for differences between Epic instances. NF1 patients were identified using the International Classification of Diseases (ICD)-10 code Q85.0.

At WU, OPG patients were identified and verified from the NF1 cohort using a combination of a structured diagnosis table and unstructured clinical notes. Patients with OPG were identified from the structured EHR diagnosis table using a set of OPG-related ICD-10 diagnosis codes (C72.3, C72.30, C72.31, C72.32). To analyze clinical notes, a natural language processing (NLP)-based phenotype extraction model was harnessed to generate a list of predicted patients harboring OPGs with a high degree of accuracy.¹⁴ A preliminary list of predicted OPG patients was created by taking the union of the patient lists derived from ICD-10 codes and clinical notes. Two subject matter experts (SMEs; D.H.G. and J.K.) reviewed the list to confirm the diagnosis of OPG by manual chart review. Those confirmed from chart review became the final NF1-OPG cohort. Subjects with central precocious puberty (CPP) were then identified from this final WU OPG cohort using a similar methodology to the creation of the preliminary OPG list. CPP patients were identified using 2 methods: (1) using a set of CPP related ICD-10 codes [E30.1, E22.8] and (2) using the NLP-based phenotype extraction model. A union of these 2 lists was used as the final CPP list and confirmed by 2 subject matter experts (D.H.G., J.K.).

At LCH, children with NF1-OPG were identified using the structured diagnosis table only. The same OPG-related ICD-10 codes used at WU were applied to identify children with NF1-OPG from the structured diagnosis table, and 2 LCH SMEs (J.G. and C.E.P.) subsequently reviewed these cases to confirm the OPG diagnosis. Those confirmed by the SMEs became the final OPG cohort. From the final LCH OPG cohort, children with CPP were identified through manual annotation by the LCH SMEs (R.L., C.E.P.).

Feature Selection

The following demographic and clinical features were extracted from the EHR data: (a) patient age at the time of NF1 diagnosis, (b) patient age at the time of OPG diagnosis and treatment, (c) age at the time of CPP diagnosis, (d) patient sex, (e) location of the OPG on MRI, (f) visual acuity, and (g) chemotherapy regimen.

At WU, features (a)-(d) were extracted from the structured diagnosis table. The remaining features, (e)-(g), were extracted from the unstructured clinical progress notes using large language model (LLM) pipelines. At LCH, feature (d) was extracted from the structured EHR demographics table while the remainder of the features was extracted through manual annotation by a SME.

At WU, the location of the OPG on MRI was extracted from narrative reports from relevant brain MRI scans selected by SMEs. The radiology reports were reviewed by SMEs to confirm tumor location (J.K., D.H.G., C.E.P., and R.L.). The tumor location used for data analysis was obtained from the patients’ MRI at the time of OPG diagnosis. OPGs on MRI were classified using the modified Dodge classification (MDC).¹⁵ OPGs involving the optic nerves were classified as MDC1, those involving the optic chiasm as MDC2, and tumors involving the optic tracts and radiations as MDC3. Tumors involving multiple regions were assigned the highest (most posterior) extent of tumor involvement. Treatment- and vision-related features were extracted from endocrinology, oncology, and neurology clinical progress notes.

To perform feature extraction from MRI reports and specialty progress notes at WU, patient clinical notes were fed into 2 different large language model (LLM)-based pipelines. The LLM-based pipelines were built using Microsoft’s Azure OpenAI Service, which provides application programming interface access to OpenAI’s language models through a HIPAA-compliant subscription within Washington University’s Azure Tenant. A GPT-4 Omni (GPT-4o) endpoint was deployed. The first pipeline uses a customized prompt that instructs GPT-4o to extract the location of the OPG and the MDC subclassification from the MRI reports. The second pipeline analyzes the endocrinology, oncology, and neurology progress notes with a prompt that instructs GPT-4o to extract OPG-related treatment regimens and visual acuity at time of treatment.

At LCH, all feature extraction was performed with manual annotation by a local SME. Annotations, including those for diagnosis of CPP, were collected using a REDCap survey. The survey dataset from REDCap was deidentified prior to transfer to WU for analysis.

Statistical Analysis

Statistical analyses were conducted using Python and Microsoft Excel. Descriptive statistics were utilized to summarize demographic and clinical characteristics. A combination of chi-square test of independence, proportions Z-test, and Mann–Whitney U tests were used to compare demographics and disease characteristics between the 2 centers. An independent samples t-test was used to compare the age distribution between groups. A chi-square test of Iindependence was applied to assess differences in tumor characteristics between groups. All statistical tests were performed using a significance level set at P < .05.

Results

Two hundred eight children with NF1 and a diagnosis of OPG were included in this study, and 101 (57%) were males (Table I). The mean age of NF1 diagnosis in this cohort was 7.0 years (range of 0.2 to 17.4 years). Fifteen (7%) of these patients were reported as African American or Black, 1 (<1%) as American Indian or Alaska Native, 3 (1%) as Asian, 164 (79%) as White, and 25 (12%) patients with an unknown or undocumented ancestry.

Table I.

Study participant demographics

Feature	Total (n = 208)	CPP (n = 39)	No CPP (n = 169)
Sex	101 male (49%)	25 male (64%)	76 male (45%)
Age at NF1 diagnosis (y)*	Mean 7.0 Range 0.2-17.4	Mean 4.2 Range 0.75-12.0	Mean 4.9 Range 0.2-10.0
Ancestry	15 (7%) African American or Black 1 (<1%) American Indian or Alaska Native 3 (1%) Asian 164 (79%) White 25 (12%) Unknown	2 (5%) African American or Black 31 (79%) White 6 (15%) Unknown	13 (8%) African American or Black 1 (<1%) American Indian or Alaska Native 3 (2%) Asian 133 (79%) White 19 (11%) Unknown
NF1 inheritance	82 (39%) with a known gene mutation	13 (33%) with a known gene mutation	69 (41%) with a known gene mutation

CPP, central precocious puberty; NF1, neurofibromatosis type 1.

^*CPP vs No CPP, P = .85.

Thirty-nine (19%) patients in this cohort had a diagnosis of CPP. Twenty-five (64%) of these patients were male (Table I) with a mean age at diagnosis of 5.7 years old. The mean age at diagnosis of NF1 in those with CPP was 4.2 years compared with 4.9 years for those without CPP (P = .85).

Demographic variables (sex and race) and disease characteristics (CPP prevalence, age at CPP diagnosis, hypothalamic involvement, and MDC group) were compared between centers. Sex and race differed slightly between centers (P = .034 and P = .012, respectively). No significant differences were observed for CPP prevalence, age at CPP diagnosis, or hypothalamic involvement (P = .70, .82, and 0.49, respectively). In contrast, MDC group distribution differed between centers (P = .025).

The mean age at OPG diagnosis in children with CPP was 5.3 years compared to 6.6 years for those without CPP (P = .10; Table II). Of children who developed CPP, 5 (13%) of their OPGs were classified as MDC1 compared with 81 (48%) without CPP. Fourteen (36%) children with CPP had an OPG classified as MDC2 compared with 54 (32%) without CPP. Twenty (51%) children with CPP had an OPG classified as MDC3 compared with 34 (20%) without CPP (x²(2, n = 208) = 21.451, P < .0001). In total, 31 (79%) children with CPP had OPGs involving the hypothalamus compared with 50 (30%) patients without CPP (x²(1, n = 208) = 32.929, P < .0001).

Table II.

OPG characteristics

Feature	CPP (n = 39)	No CPP (n = 169)
Age at OPG diagnosis (y)	Mean 5.3, Range 1.0-15.0	Mean 6.6, Range 0.9-18.0
Tumor location*	MDC1 = 5 (13%) MDC2 = 14 (36%) MDC3 = 20 (51%)	MDC1 = 81 (48%) MDC2 = 54 (32%) MDC3 = 34 (20%)
Hypothalamic involvement†	31 (79%)	50 (30%)

CPP, central precocious puberty; MDC, modified Dodge criteria; OPG, optic pathway glioma.

^*MDC (MDC1, optic nerves; MDC2, chiasm; MDC3, optic tracts and radiations) Tumors involving multiple regions were assigned the most posterior extent of tumor involvement. CPP v. No CPP; P < .0001.

^†Hypothalamic involvement: CPP v. No CPP; P < .0001.

Of children with CPP, 17 (44%) had decreased visual acuity (>0.2logMAR above age-based norms) compared with 45 (27%) of those without CPP, x²(1, n = 208) = 4.191, P = .041 (Table III). In addition, 20 (51%) of those with CPP underwent chemotherapy for their OPG, compared with 52 (31%) of patients without CPP (x²(1, n = 208) = 5.752, P = .016).

Table III.

Patient characteristics

Feature	CPP (n = 39)	No CPP (n = 169)
Decreased VA (Y/N)*	17 (44%) Y 22 (56%) N	45 (27%) Y 124 (73%) N
Tumor treatment (Y/N)†	20 (51%) Y 19 (49%) N	52 (31%) Y 117 (69%) N

CPP, central precocious puberty; N, no; VA, visual acuity; Y, yes.

^*Decreased VA: CPP v. No CPP; P = .041.

^†Tumor treatment: CPP v. No CPP; P = .016.

Our team had previously developed and published a methodology for extracting NF1-subphenotypes from clinical notes using rule-based NLP and LLMs.¹⁴ This method was adapted for the current study, with new OPG-related entities integrated into the pipeline. All extracted entities were reviewed and validated by subject matter experts [C.E.P., J.G., D.H.G., J.K.]. Central precocious puberty was identified using the NLP model, while other OPG–related entities were extracted using a large language model. The entity-level performance, as measured by F1 score, was as follows: 0.96 for CPP, 0.75 for midline diencephalic compression group, 0.87 for hypothalamic involvement, 0.76 for reduced visual acuity, and 0.86 for OPG treatment.

Discussion

In this study, we used an AI-based LLM approach to investigate the impact of co-existing CPP on the clinical features and progression of OPGs in children with NF1. Our findings demonstrate a significant association between tumor location and the presence of CPP, with OPGs located more distally along the optic pathway being more common in children diagnosed with CPP. Interestingly, while a greater proportion of individuals with CPP had hypothalamic involvement compared with those without CPP, hypothalamic involvement was not universally present, as reported in prior studies.^9–13

In addition, we found that children with OPG and co-xexisting CPP had worse visual acuity outcomes and a higher likelihood of undergoing chemotherapy relative to individuals with OPG and no CPP. This relationship may be explained by more distal tumor involvement, as distal optic pathway gliomas (corresponding to higher MDC) more typically exhibit progressive vision loss and are more likely to require therapy.^2,16 Prior studies have similarly reported that children with NF1 harboring chiasmatic and hypothalamic tumors are more frequently treated with chemotherapy.^3,17,18

Although chemotherapy exposure has been implicated as a potential contributor to CPP development, existing evidence suggests that this effect is rare and unlikely to account for the increased CPP prevalence observed in our cohort.¹⁸ Instead, our findings support the hypothesis that tumor-related hypothalamic disruption, rather than treatment effects, primarily underlies CPP pathogenesis in NF1-associated OPG. It should be noted that the relationship among OPG, chemotherapy, and CPP is complex. Although chemotherapy is not a risk factor for CPP, and endocrinopathies, such as CPP, are not indications for chemotherapy, OPG treatment with chemotherapy does not eliminate the risk of development of CPP.

This study has several limitations. This study is retrospective, limiting the ability to infer causality between CPP and OPG characteristics. In addition, as this was a multicenter study, there may be institutional differences in documentation, diagnostic practices, and the availability of longitudinal data. Furthermore, 79% of the study population identified as White, limiting generalizability to more diverse NF1 populations. Of the patients in our cohort diagnosed with CPP, 11 patients were diagnosed with CPP prior to imaging and the diagnosis of OPG. This may influence the identification of hypothalamic involvement among the identified OPGs, complicating the interpretation of these results. The data extracted using NLP and LLMs were manually reviewed by a single reviewer. Therefore, no inter-reliability analysis was conducted to validate the annotations, which may introduce potential bias.

Taken together, these results highlight the importance of tumor location as a predictor of CPP in NF1-related OPG. Recognition of this association has important clinical implications for surveillance and early intervention, as children with NF1 who harbor most distal OPGs may benefit from closer monitoring for precocious puberty and visual decline. In addition, these results emphasize the utility of AI phenotyping in the early recognition of NF1 patients at risk for visual and endocrine complications.

Glossary

ACTH

Adrenocorticotropic hormone

CPP

Central precocious puberty

EHR

Electronic health record

ICD

International Classification of Diseases

HIPAA

Health Insurance Portability and Accountability Act of 1996

LCH

Lurie Children’s Hospital

LLM

Large language model

MDC

Modified Dodge classification

MRI

Magnetic resonance imaging

NF1

Neurofibromatosis type 1

OPG

Optic pathway glioma

SME

Subject matter expert

VA

Visual acuity

WU

Washington University