ESTRO 2020 Abstract Book

S829 ESTRO 2020

of 574 patients with LAPC was included in this study. Radiation Therapy (RT) doses ranged from 24.0-61.2 Gy delivered in fractions of 1.25-8 Gy. GEM doses ranged from 0.2-3.6 g/m 2 , administered at individual doses of 40-600 mg/m 2 once or twice a week intravenously over 30 or 60 minutes, during the course of RT. Total tumour response was described using the Poisson model where the total biological effective dose (BED) of RCT was calculated as the sum of the biological effective dose of RT (BEDrt) and the equivalent biological effective dose of CH (BEDch). Considering only the cases irradiated with 50.4 Gy in 1.8 Gy fractions (BEDrt of 54.6 Gy), the variation in tumour response with the dose of GEM was assessed and modelled. Other dose prescriptions were used to assess the robustness of the models with BEDrt. Tumour response to RT was described by the Linear- Quadratic-Time model using the parameters: g of 3.1, a/β of 6.4, Tpot of 16 days and Tk of 21 days. Three models describing tumour response to CH were evaluated: 1) constant model: tumour response to CH was constant independent of the delivered dose of GEM; 2) linear model: response increased linearly with the dose of CH; 3) Hill model: common model used to describe cell survival to CH having as parameters: Emax (maximum effect), ED50 (dose causing a 50% maximum inhibition effect) and H (is the Hill exponent describing the slope of the curve). Model parameters were determined using a non-linear least square function. Results For the constant model and a BEDrt of 54.6 Gy resulted in an equivalent BEDch, of 64.7 Gy. This value did not change significantly for other BEDrt values. Overall tumour response to RCT increased with the dose of GEM up to a maximum of 50% for the total dose of 2.4 g/m 2 (400mg/m 2 /week). Thus, with the linear model an increase in BEDch of 5.1 Gy/g/m 2 was determined to which a baseline value of 57.4 Gy needs to be added. The estimated parameters of the Hill model are Emax = 0.2%, ED50 = 7.6x10 -5 g/m 2 and H = -0.48. Within the range of GEM doses up to 2.4 g/m 2 , both the linear model and the single Hill model showed a better association with clinical outcome (sum square error ~80) than the constant model (sum square error 149). Conclusion The linear model is the simplest approach to describe the variation of tumour response with the dose of GEM during RCT. The single Hill model, requiring the determination of three parameters, has shown to also be a good descriptor of the impact of RCT on tumour response. However, both these models fail considerably for high dose values of GEM as those used in CH as monotherapy. PO-1535 Machine Learning and Oversampling techniques to predict urinary toxicity after prostate cancer RT E. Mylona 1 , F. Filias 2 , M. Ibrahim 3 , S. Supiot 4 , N. Magne 5 , G. Crehange 6 , M. Hatt 3 , O. Acosta 1 , R. De Crevoisier 1 1 Université de Rennes 1, LTSI, Rennes CEDEX, France ; 2 University of Patras, Physics Departement, Patras, Greece ; 3 University of Brest, LaTIM, Brest, France ; 4 Centre Georges François Leclerc, Department of Radiation Oncology, Dijon, France ; 5 Lucien Neuwirth Cancer Institute, Radiotherapy Department, St Priest en Jarez, France ; 6 Institut de Cancérologie de l'Ouest, Medical Physics Department, Saint Herblain, France Purpose or Objective Machine learning methods have the potential to improve the prediction capabilities of radiation-induced toxicity models. However, there is currently no consensus on the best performing algorithms, particularly in the presence of low number of toxicity events. Oversampling techniques have been developed to tackle the lack of informative data. The goal of this work was to thoroughly compare several machine-learning strategies combined with

different minority class oversampling techniques and to investigate whether they can provide improved predictive power of urinary side-effects, following prostate cancer radiotherapy. Material and Methods A dataset of 254 prostate cancer patients treated with IMRT/IGRT was used and different urinary symptoms were analyzed. Depending on the symptom, datasets presented different levels of class imbalance. A large number of parameters, including clinical-/treatment-related and dosimetric extracted from the bladder's dose-volume histograms (DVH), were analyzed in a nested 5-fold cross- validation pipeline, using different popular classification methods. The performance of these classifiers (using the area under the ROC curve (AUC) and the F-measure) was first evaluated on the original dataset and compared with the logistic regression model (considered as reference). Then, we computed and compared the classifiers’ performance on datasets that had been balanced with different oversampling techniques. Results Table 1 shows the improvement brought by different oversampling techniques averaged over all the classifiers (cross-comparison). The alternative hypothesis was that the oversampling techniques 1 performed better than the oversampling techniques 2. Oversampling with Synthetic Minority Oversampling Technique (SMOTE) followed by Edited Nearest Neighbour algorithm (ENN) outperformed other oversampling techniques. Using SMOTE+ENN, increased the prediction performance (AUC) by 22% to 89%, compared to the original dataset without oversampling. The performance of the classifiers was symptom- dependent. For late retention, for example, the performance of the logistic regression model on the original dataset was AUC=0.57 whereas the best performance was obtained with PLS-DA classifier (AUC=0.65). On the oversampled datasets, the Regularized Discriminant Analysis (RDA) and the partial least square discriminant analysis (PLS-DA) scored higher, with average AUCs ranging from 0.67 to 0.71 and from 0.61 to 0.70, respectively. Table 2 shows the average performance of each classifier and for each dataset on the hold-out test sets, for predicting late retention.

Conclusion Synthetic oversampling approaches significantly increased the prediction performance of machine learning methods,