ESTRO 37 Abstract book

S294

ESTRO 37

modify patients’ congenital risk of GU toxicity and good prediction can be achieved by combining contributions of genome-wide SNPs to the risk using machine learning methodology. Material and Methods We studied 324 prostate cancer patients who received brachytherapy with or without external beam radiotherapy, were genotyped for 606895 SNPs, and followed up for minimum 3 years according to patient- reported International Prostate Symptom Score (IPSS) guidelines. A late toxicity event for each symptom was defined as an increase in IPSS grade ≥ 3 from good (grade 0/1) baseline. We studied four (frequency, nocturia, urgency and weak stream) GU symptoms with an event rate higher than 10%. We considered 14 previously reported clinical GU risk factors including RT dose to a tumor. The pre-conditioned random forest regression (PRFR) model (Oh et al., 2017, Scientific Reports) was trained using the SNPs as predictors with association p- value < 0.001 and clinical variables with p-value < 0.05. Model training and validation was respectively performed in randomly split training (2/3) and validation (1/3) datasets. Using a database of biological knowledge, we searched for enriched gene ontology biological processes and a group of connected proteins from the SNPs identified by PRFR as high importance. Results No clinical variable was significantly associated with any of the endpoints after Bonferroni correction. Performance of PRFR varied across symptoms: the areas under the curve (AUC) on the hold-out validation set were: weak stream: 0.7, frequency: 0.64, nocturia: 0.55, and urgency: 0.53. The AUC was significant only at weak stream (p = 0.01) where the odds ratio between the 1/3 lower and 1/3 higher risk groups was 3.6 (fig1). At this endpoint, the proposed model outperformed alternative multivariate methods including conventional random forest and logistic regression. Out of the 617 SNPs with high importance to weak stream, we discovered groups of enriched biological processes characterized by neurogenesis and ion transport (fig2), both of which had been shown to be involved in urinary tract functions. We also discovered a network of 15 proteins with interactions, among which 7 proteins (PKC, PKG, EGFR, Schwannomin, Annexin I, ASIC2, and Neurexin) were shown to be relevant to GU functions using systematic literature survey.

Fig2: Biological processes for weak stream.

Conclusion Our results suggest, by use of an agnostic machine learning method such as PRFR and biological interpretation, that genome-wide data can be used to predict and explain GU toxicity. The model can be refined upon external validation and incorporation of accurate dose-based predictors. PV-0565 Texture Analysis of 3D dose distributions for predictive modelling of toxicity rates in radiotherapy L. Rossi 1 , R. Bijman 1 , W. Schillemans 1 , S. Aluwini 1 , M. Witte 2 , F. Pos 2 , L. Incrocci 1 , B. Heijmen 1 1 Erasmus MC Cancer Institute, Department of Radiation Oncology, Rotterdam, The Netherlands 2 Netherlands Cancer Institute Antoni van Leeuwenhoek Hospital, Department of Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective To explore the use of texture analysis (TA) features of patients’ 3D dose distributions to improve prediction modelling of treatment complication rates in prostate cancer radiotherapy, relative to more common DVH parameters. Material and Methods Late toxicity scores, dose distributions, and non- treatment related (NTR) predictors for late toxicity, such as age and baseline symptoms, of 351 patients of the hypofractionation arm of the HYPRO randomized trial (Lancet Oncol 2016;17(4):464-74) were used in this study. Texture analysis was performed for both rectum and bladder 3D dose distributions. 42 TA features were extrapolated from 2 histograms and 5 matrices: grey level frequency histogram, grey level co-occurence matrix (GLCM), grey level run length matrix (GLRLM), grey level size zone matrix (GLSZM) and neighbourhood grey tone difference matrix (NGTDM), see figure 1 for example. TA features and common DVH parameters derived from rectum and bladder dose distributions were used for predictive modelling of gastrointestinal (GI) (rectal bleeding and fecal incontinence) and genitourinary (GU) (nocturia and urinary incontinence) symptoms, respectively. Logistic Normal Tissue Complication Probability (NTCP) models were derived, using only NTR parameters, NTR + DVH, NTR + TA, and NTR + DVH + TA. Results For rectal bleeding, the area under the curve (AUC) for using only NTR parameters was 0.58, which increased to 0.68, and 0.71, when adding DVH or TA parameters respectively. For fecal incontinence, the AUC went up from 0.62 (NTR only), to 0.68 (+DVH) and 0.75 (+TA). For nocturia, adding TA features resulted in an AUC increase from 0.64 to 0.67, while no improvement was seen when including DVH parameters in the modelling. For urinary

Fig1:Risk stratification for weak stream. Error bars=1 standard error.