ESTRO 2021 Abstract Book

S78

ESTRO 2021

PH-0109 Reactive Feeding Tube Insertion Relates to Dose to Tongue Subunits and Brainstem E.M. Vasquez Osorio 1,2 , A. Abravan 1,2 , A. Green 1 , A. Greeen 2 , V. Kota 3 , S. Gaito 3 , A. McPartlin 3 , M. van Herk 1,2 1 The University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom; 2 The Christie NHS Foundation Trust, Radiotherapy Related Research, Manchester, United Kingdom; 3 The Christie NHS Foundation Trust, Department of Clinical Oncology, Manchester, United Kingdom Purpose or Objective Outcomes of head and neck (H&N) cancer treatments are compromised by malnutrition. Causes for this are complex and are not completely understood. During radiotherapy, malnutrition can be mitigated using enteral nutrition via a feeding tube insertion (FTi) when required. Here, we explore the relation between dose to healthy anatomy and reactive FTi. Materials and Methods Voxel-wise image-based data mining (IBDM) was applied to data from 210 H&N cancer patients treated in 2016-2017 with definitive (Chemo)RT. We spatially normalised all patient’s planning dose matrices (in equivalent dose at 2 Gy/fraction, α/β=10) to a reference patient anatomy using NiftyReg [1]. To reduce target laterality bias, patients were mirrored in the lateral axis and also mapped to the reference patient. The reference patient contained contours of typical OARs, such as the parotid glands and brainstem, and of functional swallowing subunits [2] derived from 8 atlas patients. Regions where the mean dose for groups with/without FTi differed significantly were identified using permutation test (1000 permutations). Mean/max doses were collected for OARs overlapping the identified region. Random forest classifiers (RFC) were used to reduce number of highly correlated dosimetric variables. 1000 RFCs with 100 trees each were trained using random 80%-20% data splits. The 5 most important variables were identified based on the number of times they were selected as the top factor in any RFC. Dose and clinical variables were then evaluated in a multivariable logistic regression model. Backward stepwise model selection was used to create the final model. Results A highly significant region, t-value=-5.5, was found where dose correlated with reactive FTi (Figure 1). This region included most functional subunits and salivary glands, with overlaps up to 96% (Table 1). Interestingly, 8% of the brainstem was also included in the region of interest. RFCs selected mean dose to the floor of the mouth, intrinsic tongue muscle (ITM), brainstem, and max/mean dose to the genioglossus (GGS) muscle. Following backward selection, sex, age, chemotherapy (yes/no), weight before treatment, mean ITM, mean GGS and mean brainstem doses remained (Table 2). The area under the curve for the final model was 0.73, compared to 0.69 for a model with the selected clinical variables only.