Background & Summary

Head and neck cancers (HNC) affect over 58,000 Americans annually, with a growing proportion attributed to human papillomavirus (HPV) infection1. HPV-associated HNC are notably diagnosed in younger populations and are associated with higher survival rates in comparison to HPV-negative HNC2. Radiation therapy (RT) remains the mainstay of treatment for HPV-positive HNC, but the combination of RT with extended survival has led to an increased incidence of RT-induced late toxicities in normal tissues. One such complication is osteoradionecrosis of the jaw (ORNJ), a severe sequela following RT with an incidence ranging from 4 to 15%3. The mechanism of ORNJ is believed to be first instigated by compromised vascularity through hypoxic, hypovascular, and hypocellular tissue (Marx’s 3 H’s)4 followed by progressive loss in cortical bone integrity, ultimately impairing oral function and quality of life5,6. Due to the favorable RT response and prognosis of HPV-associated HNC and the subsequent number of patients transitioning to survivorship, there is a need to better understand the timing and progressive risk of ORNJ in relation to radiation treatment of HNC in order to optimize prevention efforts of this often debilitating condition.

Previous cross sectional statistical analyses7,8,9,10, including Normal Tissue Complication Probability (NTCP) models of ORNJ11,12, have identified clinical and dosimetric risk factors associated with this sequela. Additionally, some studies have explored statistical correlations on longitudinal ORNJ data13,14. In recent work, we developed a fully parametric multivariable Weibull Accelerated Failure Time (WAFT) model to predict patient-specific ORNJ risk over time based on longitudinal data.

This data descriptor presents the underlying dataset used for the development of the ORNJ WAFT model15. The dataset is comprised of a large, longitudinal cohort of HNC patients and includes detailed demographic, clinical, and dosimetric variables, along with structured follow-up data and time-to-ORNJ events. The availability of this dataset offers a valuable resource for modeling ORNJ and supports the development of predictive tools for personalized survivorship care in HNC.

Methods

IRB protocol

After the University of Texas MD Anderson Cancer Center Institutional Review Board approval, data were extracted from a philanthropically funded observational cohort at the University of Texas MD Anderson Cancer Center (Stiefel Oropharynx Cancer Cohort, PA14-0947). A waiver of informed consent was approved through the MD Anderson RCR030800 protocol, allowing for retrospective analysis. All patients included were consented RT cases. We implemented formal reporting guidance as per Enhancing the QUAlity and Transparancy Of health Research Network guidance, using the RECORD Statement16, attached as a supplement.

Patient population

1129 HNC patients from an internal MD Anderson Cancer Center cohort were treated with curative intent RT from 2005 to 2022. Patients were closely followed via clinical and radiological assessments every 3, 6, 12, 18, and 24 months, and then approximately annually following the conclusion of RT. As this cohort derives from a single institution, generalizability to institutions with different patient demographics and treatment practices may be limited. This dataset has been externally validated on an independent cohort in the parent study15. The patient data was stored in and accessed via the Epic Electronic Health Record System.

Demographic data

All demographic, clinical, and dosimetric variables are summarized in Table 1. The patients’ demographic data included: gender (male or female), age (in years), smoking status (current, former, never), and smoking pack-years. Smoking pack-years were calculated by the product of tobacco packs smoked per day and number of years smoked (Table 2).

Table 1 Demographic, clinical, and dosimetric data for included patient cohort showing variable, its respective coding (qualitative or quantitative) and units (when applicable).
Full size table
Table 2 Distribution of demographic and clinical data stratified by ORNJ status (control group and ORNJ group).
Full size table

Clinical data

The patients’ clinical data included: overall survival, ORNJ status (binary, yes or 1 vs. no or 0), time to event, ORNJ grade, pre-RT dental extractions, T stage, N stage, chemotherapy (induction vs. induction and concurrent vs. concurrent vs. no chemotherapy), post-operative RT vs. definitive RT, HPV/p16 + Ve status (yes vs. no or unknown), tumor site group (oropharynx vs. oral cavity vs. nasopharynx/nasal cavity/paranasal sinuses vs. larynx/hypopharynx vs. major salivary glands vs. other), and mandible volume (in cubic centimeters, cc). 916 patients (81%) were coded with an HPV/p16 + Ve Status of ‘Unknown.’ While this is reflective of practical limitations17, it may bias future analyses, as HPV/p16 status has been shown to impact survival rates and quality of life2,18. Multiple imputation serves as a potential strategy to derive missing HPV/p16 statuses19,20. Patients with missing data were not included in the original analysis15. Overall survival time, binarily coded as 0 for no survival and 1 for survival, represents the time in months between time of RT start date and time of death or time to last follow-up. As this dataset covers a wide range of years preceding ORNJ grading consensus21, ORNJ status was binarily coded to account for any variability across staging systems22. 0 indicated no ORNJ detected and 1 indicated an active ORNJ diagnosis (of any grade) at time of last follow-up. To reflect current clinical standards, ORNJ grade was also specified using a numeric value of 0–4 following the Tsai staging system23. For patients with active ORNJ, time to event is calculated in months from the RT start date to time of ORNJ diagnosis. For patients without an active ORNJ diagnosis, the time to event was censored to be the time in months from RT start date to either time of death or last follow-up. Pre-RT dental extractions were binarily coded—0 indicated negative and 1 indicated positive for pre-RT dental extractions. T stage and N stage indicate the cancer stage, following the standard TNM staging system by the American Joint Committee on Cancer (AJCC, 7th/8th edition) and the International Union Against Cancer. Chemotherapy and post-operative RT vs. definitive RT indicate if RT was combined with another treatment; ‘concurrent chemotherapy’ indicates chemotherapy occurred simultaneously with RT while ‘induction chemotherapy’ indicates chemotherapy was completed before RT. Likewise, ‘post-op RT’ indicates RT was completed following surgery while ‘definitive’ indicates RT was completed without surgery. HPV/p16 + Ve indicates positive expression of HPV/p16 via ‘yes’, ‘no’, or ‘unknown.’ Mandible volume was reported (in cc) from delineated mandible contours; mandible bone was auto-segmented with a previously validated multiatlas-based auto-segmentation using commercial software ADMIRE (research version 1.1; Elekta AB, Stockholm, Sweden).

Dosimetric data

The patients’ dosimetric data included the following dose-volume metrics: volume of the mandible receiving at least a specified dose (V5-V80 Gy in 5 Gy increments), and dose received by a specified volume of mandible (D0.5%, D1%, D2%, D3%, D5-D95% in 5% increments, D97%, D98%, D99%, D99.5%). These metrics were calculated directly from the radiation dose distribution DICOM files utilizing a Python-based software developed from core standards and software24,25,26,27,28,29,30, notably pydicom and RT Dose Module Attributes as specified in DICOM PS3.3, and tested in-house.

Data Record

The complete comma-separated value (CSV) file containing demographic, clinical, and dosimetric data for the aforementioned patient population is publicly available on figshare31. This CSV file provides the unique opportunity for analysis of a large HNC cohort with detailed treatment-related information related to prevalence and timing of ORNJ.

The authors acknowledge the dichotomy of open science while maintaining patient confidentiality; this is particularly important with cohorts of long-term survivors. As such, patient identification was anonymized through a randomly assigned subject ID independent from their medical record number (MRN). The dataset contains no other patient identifiers (Figs. 1, 2).

Fig. 1
figure 1

Right-censored Kaplan-Meier curves denoted time-to-ORNJ diagnosis, stratified by (a) D25% with a 50 Gy threshold, (b) dental extractions, with 0 and 1 indicating the absence or presence of a pre-RT dental extraction, respectively, and (c) gender.

Full size image
Fig. 2
figure 2

Diagram showing the data collection workflow and input into the final CSV file. Patients underwent RT at MD Anderson Cancer Center (left), in which dosimetric data was generated and acquired from a treatment planning system (top middle). Patient demographic and clinical data were also acquired from initial and follow-up visits (bottom middle). These data were then inputted into the CSV file included within this data descriptor (right).

Full size image

Technical Validation

Patient demographic and clinical data was stored and accessed via manual extraction by post-doctoral fellows with radiation oncology training from the University of Texas MD Anderson Cancer Center’s Epic Electronic Health Record System server and imported into REDCap electronic data capture tools hosted at the University of Texas MD Anderson Cancer Center32,33. The dataset was curated by multiple observers over time using a standardized template and variable dictionary. When discrepancies were suspected, records were double-checked against the original sources and corrected if inconsistencies were identified. Although formal inter-rater reliability statistics were not calculated, this approach provided additional quality assurance during the curation process to minimize misclassification and confirmation biases34.

Dosimetric data was obtained from clinical radiotherapy treatment plans using the RayStation treatment planning system (RaySearch Laboratories AB, Stockholm, Sweden). These data were first exported in standardized DICOM-RT (Digital Imaging and Communications in Medicine – Radiation Therapy) format and then analyzed to calculate dose-volume metrics to be used in the model.

1471 patients were examined for eligibility for this analysis. 342 patients were excluded due to clinical reasons such as prior irradiations; others were excluded for incomplete or missing data. The final cohort included a dataset of 1129 HNC from MD Anderson Cancer Center.

Usage Notes

The WAFT-based time-to-ORNJ online calculator graphical user interface (GUI) is available at https://uic-evl.github.io/OsteoradionecrosisVis/.