ESTRO 2024 - Abstract Book

S1345

Clinical - Head & neck

ESTRO 2024

dysfunctions, dental caries, periodontal disease, and osteoradionecrosis. Cancer Med. 2017 Dec;6(12):2918-2931. doi: 10.1002/cam4.1221. Epub 2017 Oct 25. PMID: 29071801; PMCID: PMC5727249.

3. Marks LB, Yorke ED, Jackson A, Ten Haken RK, Constine LS, Eisbruch A, Bentzen SM, Nam J, Deasy JO. Use of normal tissue complication probability models in the clinic. Int J Radiat Oncol Biol Phys. 2010 Mar1;76(3Suppl):S10-9. doi:10.1016/j.ijrobp.2009.07.1754. PMID: 20171502; PMCID: PMC4041542.

1821

Digital Poster

Support vector data description-based control chart to monitor organ-at-risk dose

Sruthi Sivabhaskar 1 , Jacob Buatti 1 , Kristen Duke 1 , Michelle De Oliveira 1 , Neil Kirby 1 , Nikos Papanikolaou 1 , Arthur Yeh 2 , Arkajyoti Roy 3 1 The University of Texas Health Science Center at San Antonio, Radiation Oncology, San Antonio, USA. 2 Bowling Green State University, Applied Statistics and Operations Research, Bowling Green, USA. 3 The University of Texas, Management Science and Statistics, San Antonio, USA

Purpose/Objective:

The purpose of this study is to develop a nonparametric control chart based on the support vector data description (SVDD) 1 to monitor dose to organs-at-risk (OARs) after adjusting for the presence of interpatient variations in anatomy and dose prescriptions. 2 Specifically, six OAR dose-volume histogram (DVH) points, D 2 , D 20 , D 40 , D 60 , D 80 , and D 98 , were monitored by the control chart. SVDD, as a nonparametric machine learning technique designed for one-class classification problems, does not assume that the underlying data follows a normal distribution. This is particularly advantageous given that OAR DVH points often deviate from a normal distribution.

Material/Methods:

Data from 80 head-and-neck patients previously treated at our institution were used to develop the SVDD-based control chart. The brainstem was selected as the OAR to demonstrate the application of the control chart in detecting patients receiving extreme brainstem dose. The Gaussian kernel was employed to transform the patient data into a higher-dimensional space. Bootstrap sampling technique was used to generate 5000 random samples with replacement, each containing 80 observations. Kernel distances for each observation in a bootstrap sample were computed and ordered from least to greatest. The 76th kernel distance, corresponding to a significance level of 0.05, was taken as the control limit for that bootstrap sample. The bootstrap estimate of the control limit was obtained by averaging the 5000 control limits determined from each sample. This approach allows us to assess the variation in control limits across different random samples of observations. An observation is categorized as in control (IC) if the kernel distance is less than or equal to the bootstrap estimate of the control limit, while an observation is identified as out-of-control (OC) if the kernel distance is greater than the bootstrap estimate of the control limit.