ESTRO 37 Abstract book

S1097

ESTRO 37

EP-2009 Metamodelling of late rectal bleeding in patients undergoing radiotherapy for prostate cancer. A. Cicchetti 1 , T. Rancati 1 , F. Palorini 1 , B. Avuzzi 2 , C. Stucchi 3 , V. Vavassori 4 , G. Fellin 5 , P. Gabriele 6 , C. Degli Esposti 7 , M. Ebert 8 , A. Kennedy 9 , D.J. Joseph 9 , J.W. Denham 10 , C. Cozzarini 11 , C. Fiorino 12 , R. Valdagni 13 1 Fondazione IRCCS Istituto Nazionale dei Tumori, Prostate cancer program, Milan, Italy 2 Fondazione IRCCS Istituto Nazionale dei Tumori, Radiation Oncology 1, Milan, Italy 3 Fondazione IRCCS Istituto Nazionale dei Tumori, Medical Physics Department, Milan, Italy 4 Cliniche Gavazzeni Humanitas, Radiation Oncology, Milan, Italy 5 Ospedale Santa Chiara, Radiation Oncology, Trento, Italy 6 Istituto di Candiolo- Fondazione del Piemonte per l'Oncologia IRCCS, Radiation Oncology, Torino, Italy 7 Ospedale Bellaria, Radiation Oncology, Bologna, Italy 8 University of Western Australia, School of Physics, Perth, Australia 9 Sir Charles Gairdner Hospital, Radiation Oncology, Perth, Australia 10 University of Newcastle, School of Medicine and Public Health, New South Wales, Australia 11 San Raffaele Scientific Institute, Radiation Oncology, Milano, Italy 12 San Raffaele Scientific Institute, Medical Physics, Milano, Italy 13 Università degli Studi di Milano, Oncology and Hemato- Oncology, Milano, Italy Purpose or Objective Predictive models for late rectal bleeding (LRB) after radiotherapy (RT) for prostate cancer (PC) were established by several studies, with good performance on development cohorts. Nevertheless, they were often found to be unsatisfactory in their generalization to independent validation. Aim of this study was to consider a modelling procedure that goes beyond maximum likelihood model fitting of data. Thus, to create a metamodel for grade 2-3 (G23) and grade 3 (G3) LRB starting from literature evidence and validate it on a population of PC patients. Material and Methods A model is a theoretical description of phenomena in the real world. More specifically, metamodelling allows to highlight specific features of the model itself. The common characteristic of the predictive models for LRB was the global harmony of radiobiological parameters (volume effect parameter, n, steepness of the curve, k, and dose parameter associated to 50% of complication probability, D50) and the recurrence of specific clinical factors. Thus, models including rectal Equivalent Uniform Dose (EUD) with/without patient-related dose-modifying factors were retrieved by literature search. Dosimetric coefficients were resolved by weighted mean of published values, using their standard deviation as weight. Identified clinical features, expressed by dose modifying factors (DMF), were differently weighted taking into account the prevalence of the features and the size of the study since DMF were usually reported in literature without confidential intervals. Finally, both factors were inserted in a modified logit-EUD model. The resulting metamodel was validated on a pooled population of radically treated patients (3DCRT & IMRT) enrolled in three large international cohorts. Perfor- mance was assessed through calibration. Results Literature search identified rectal EUD, previous abdominal surgery, hormone therapy and use of cardiovascular drugs as relevant features: associated coefficients are presented in Table. Validation cohort included 1591 patients with complete dosimetric, clinical and toxicity data: 240 (15%) had LRBG23 and 98 (6.2%)

calculated in each voxel of the reference anatomy, thereby controlling for tumour size, and age (Figure 1). Voxels in which the hazard ratio was statistically significant were labelled using a permutation test method.

Results A statistically significant region, in which excess dose is associated with poor survival, was identified in the base of the heart (Figure 2). This region is similar to previous findings using IBDM on a similar cohort but has shifted slightly inferior compared to previous analyses. The hazard ratio in this region is between 1.012 and 1.017 Gy - 1 .

Conclusion IBDM using a Cox regression per voxel is feasible and identified a region in the heart in which excess dose is related to poorer survival, in agreement with recent literature. This technique is more selective than standard IBDM techniques, in which confounding variables are accounted for after the identification of a region and is therefore well suited for the presented heart case where increasing tumour size increases dose to the region of interest while being associated with a reduction of survival. This technique may lead to more easily testable hypotheses and accelerate clinical benefit. Further work will include additional variables, such as tumour stage and performance status and explore interaction of variables.