ESTRO 2024 - Abstract Book

S1368

Clinical - Head & neck

ESTRO 2024

The PREDMORN multi-institutional consortium was set up in 2022 among six centres to develop PREDiction models for Mandibular OsteoRadioNecrosis in HNC with the largest and most diverse dataset ever published. In this work, we present the study's first phase results, which focused on identifying clinical, demographic and DVH-based dosimetric predictors of ORN.

Material/Methods:

Data from 1225 patients, 397 ORN cases and 828 matched controls, was retrospectively collected by six institutions as per the established study protocol [1].

The endpoint was dichotomised as no ORN vs. any stage of ORN. Categorical clinical variables were encoded as binary variables, and continuous clinical variables (age, mandible volume and follow-up time) were max-min scaled. Clinical variables with >25% of missing data were excluded. Iterative imputation was implemented for all other variables to substitute the missing data. Data collection and patient population varied between centres, and thus, a centre-specific data imputation approach was considered more appropriate. Table 1 summarises the clinical and demographic characteristics and the missing data percentages of the study dataset (before imputation) for the ORN and control groups. The dosimetric information was provided as the dose-volume metrics Dmax, Dmean, Dmin, D2%, D5%-95% (in 5% steps), D98% and V5Gy-V70Gy (in 5 Gy steps) extracted from the uncorrected DVH data for the mandible. Fraction size was included as a variable to account for the different fractionation schemes. Multivariable logistic regression with stepwise forward feature selection (SFLR) was performed on the dosimetric data and clinical variables. The SFLR model was trained on the study dataset using a 5-fold cross-validation approach with the area under the receiver operating characteristic curve (ROC AUC) as the scoring metric. The datasets were then transformed to include only the selected features, and a logistic regression model was fit on the transformed dataset.

Table 1. Clinical and demographic characteristics of the original dataset (before imputation). Percentage of missing data for the variables included is also shown.

Missing data

ORN

Control

p-value

Count

397

828

Gender: Male vs. Female

91 (22.9%) / 306 (77.1%)

206 (24.9%) / 622 (75.1%)

0 (0.0%)

0.498

Age

(years )

0 (0.0%)

60.6 (28.0-84.8)

60.7 (24.0-91.0)

0.892

mean (range)

Follow-up

time

(years) mean (range)

3 (0.2%)

4.9 (0.1-14.2)

4.5 (0.1-12.2)

0.011

Primary tumour site:

Oropharynx

247 (62.2%)

531 (64.1%)

0.672

Oral cavity

120 (30.2%)

239 (28.7%)

0.557