ESTRO 38 Abstract book

S512 ESTRO 38

of realistic systematic deviations in delivered dose existing throughout the treatment course.

PO-0948 Predicting lung function post-RT in lung cancer using multivariate and principal component analysis Y. Dong 1 , M. Tawhai 2 , C. Veiga 3 , T. Doel 4 , D. Landau 5 , J. McClelland 3 , K. Burrowes 1 1 University of Auckland, Chemical and Material Engineering, Auckland, New Zealand ; 2 University of Auckland, Auckland Bioengineering Institute, Auckland, New Zealand ; 3 University College London, Department of Medical Physics & Biomedical Engineering, London, United Kingdom ; 4 University College London, Wellcome EPSRC Centre for International and Surgical Sciences, London, United Kingdom ; 5 University College London, Department of Oncology, London, United Kingdom Purpose or Objective Radiation therapy to treat lung cancer presents a trade- off between eradicating cancer cells and minimising radiation-induced lung damage (RILD). It is challenging to predict which patients will suffer from RILD due to the large variation in individualised response to RT, due to factors such as age, smoking status and baseline lung function. The ability to accurately predict this risk and the functional implications of RILD for an individual patient is lacking. Previous studies have typically used univariate analyses to predict lung function post-RT, with limited multivariate statistical analysis applied in this area. In this study, we apply statistical methods including principle component analysis (PCA) and multiple linear regression on a well-defined clinical dataset of patients with lung cancer before and after RT to develop a computational tool to improve RT planning and treatment in patients with lung cancer. Material and Methods Forty-four patients were selected from a Phase 1/2 trial of isotoxic dose-escalated RT and concurrent chemotherapy in patients with stage II/III non-small cell lung cancer, known as the IDEAL-CRT trial. All patients were treated with 63-73 Gy in 30 fractions over 6 weeks (daily). CT and pulmonary function data (FEV 1 , DLCO and FVC) were collected 12 months post-RT. We applied multivariate analysis, using PCA and multiple stepwise regression to develop a model to predict lung function post-RT. The dataset contained 14 variables, shown in Fig. 1. We used the PCAmix method which combines standard PCA for numerical data and multiple correspondence analysis for categorical data in the R software environment. PCAmix calculated the principal component scores (PCs) which were used as predictors to build a multiple linear regression model applying a stepwise method. Our results ( PCAmix predicted ) were compared with predictions using general multiple linear regression ( MLR predicted ) and measured values of lung function.

Based on treating patients on different machines within a clinic the Fox Chase case, shows the greatest variation in TCP with 5 th and 95 th percentiles of 71.0% and 80.7% (range 9.7%). The RT01 prostate case and head and neck case had 5 th and 95 th percentiles of 52.5% and 58.9% (range 6.4%), and 60.1% and 66.8% (range 6.7%) respectively. Figure 1 shows variation in predicted outcome for a cohort of patients due only to machine assignment.

Conclusion This analysis highlights the importance of accurate dosimetry, not only at initial calibration but also with QA. Current recommended action levels of 2% may require revision to reduce this potential difference in clinical outcome. Changes of this magnitude are readily detected; however this information is seldom used routinely. Clinical trials may benefit from reducing uncertainty in delivered dose to provide more robust assessments of response between trial arms, particularly in the context of reduced variation possible with modern treatment techniques. Refs [1] Thomas et al. 2017 Phys. Imaging Radiat. Oncol. 3 21– 7 [2] Bolt et al. 2017 Phys. Imaging Radiat. Oncol. 4 39-43 [3] Hanks et al. 2002 Radiat. Onc. Biol. Phys. 54(2) 427- 435

Fig. 1: Demonstration of the 14 variables used in our multivariate analysis. Results The r-squared value of the PCAmix regression model was 81%, compared to 55% when using the MLR model. The FEV 1 % predicted at 12 months post-RT measured/predicted were compared to assess how well