ESTRO 2024 - Abstract Book

S5139

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

2907

Digital Poster

Spatial outcome modelling for M.D. Anderson Dysphagia Inventory subscales and bootstrapping

Eliana M Vasquez Osorio 1,2 , Chloe Pratt 1 , Marcel van Herk 1,2 , Alan McWilliam 1,2 , James Price 1,3

1 The University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom. 2 The Christie NHS Foundation Trust, Christie Medical Physics and Engineering, Manchester, United Kingdom. 3 The Christie NHS Foundation Trust, Clinical Oncology, Manchester, United Kingdom

Purpose/Objective:

The MD Anderson Dysphagia Inventory (MDADI) is a 20-item questionnaire designed for evaluating the impact of dysphagia on quality of life for patients with head and neck cancer (HNC) [1] . Its widespread adoption in clinical research means that the MDADI is now arguably the principal patient-reported outcome measure for dysphagia [2] . MDADI includes four subscales: global, emotional, functional, and physical [1] , but often only the value integrating all subscales (composite, calculated including/excluding the global item [2] ) is recorded in electronic patient records. Using composite scores, such as the MDADI, poses extra challenges in outcome modelling as there may be different drivers for individual components of the composite outcome. In this study, we assessed the impact of MDADI subscales in spatial outcome modelling using image-based data mining (IBDM), a voxel-based analysis technique.

Material/Methods:

IBDM was applied to data from 50 patients with oropharyngeal cancer treated with radical radiotherapy (66 Gy/30 fractions). MDADI scores, including composite score, emotional, functional and physical subscales were recorded for each patient at least 2 years post radiotherapy. Due to the low number of patients with subscales recorded, we only performed the first two steps in the IBDM approach thereby identifying regions where the dose is significantly associated with the studied outcomes, i.e., MDADI composite score and each subscale independently. First , we performed spatial normalisation following a registration pipeline previously proposed [4] . In short, we registered each patient’s planning CT to three open-data reference anatomies using a combination of thin-plate splines deformation (for neck alignment) and B-spline non-rigid registration (NiftyReg). To reduce target laterality bias, all patients were mirrored in the left-right axis and mapped twice to each reference patient. These registrations were used to map all patient’s planning dose matrices (in equivalent dose at 2Gy/fraction, α/β=3Gy) to the three reference anatomies. Second , we performed voxel-wise analysis, where Spearman correlations were estimated on each voxel in the anatomy and significant regions were identified after permutation testing (n=1000 and p<0.05). We further applied bootstrapping to robustly identify the significant regions in this small cohort. Bootstrapping is particularly useful when dealing with small sample sizes for deriving robust statistical estimates. For this, we randomly resampled our data 100 times with replacement and repeated the process described above to identify significant regions. Next, we selected all voxels that were part of identified regions at least 50% of the times.