ESTRO 38 Abstract book

S1055 ESTRO 38

EP-1937 Distributed AUC algorithm: a privacy- preserving approach to measure the performance of Cox models C. Masciocchi 1 , A. Damiani 1 , N.D. Capocchiano 1 , J. Van Soest 2 , J. Lenkowicz 1 , E. Meldolesi 1 , G. Chiloiro 1 , M.A. Gambacorta 1 , A. Dekker 2 , V. Valentini 1 1 Policlinico Universitario Agostino Gemelli-, Divisione di Radioterapia Oncologica - Gemelli ART, Roma, Italy ; 2 GROW School for Oncology and Developmental Biology, Department of Radiation Oncology Maastro, Maastricht, The Netherlands Purpose or Objective Recent years have brought both a notable rise in the ability to efficiently harvest vast amounts of information, and a concurrent effort in preserving and actually enforcing the privacy of patients and their related data, as evidenced by the European GDPR. In these conditions, the Distributed Learning Ecosystem has shown great potential in allowing researchers to pool the huge amounts of sensitive data need to develop and validate prediction models in a privacy preserving way and with an eye towards personalized medicine. The aim of this abstract is to propose a privacy-preserving strategy for measuring the performance of Cox Proportional Hazard (PH) model. Material and Methods A privacy-preserving AUC strategy has been proposed, developed and tested. The algorithm is mainly composed of 4 distinct steps: - Each site calculates the vector of the linear predictor based on the sites’ local data and sends to the results to the master. - The master merges the received lists of linear predictors and sorts them by calculating the threshold values. This vector is then sent back to each site. - All sites then calculates the confusion matrix for each value of the threshold and sends the results back to the master; - The master sums up the confusion matrices for each value of the threshold, and subsequently calculates the true positives and false negatives to plot the ROC points. The area under the ROC curve is finally computed by using trapezoidal approximation. The proposed method has been tested on a centralized and a distributed infrastructure, with real rectal cancer clinical data, with the data split into 2 random and independent datasets. Age, gender, clinical TNM stage, overall survival status and time were considered in the analysis. The resulting ROC and AUC from both the centralized and distributed infrastructure have then been analyzed and compared. Results A total number of 945 rectal cancer patients (pts) were selected to develop and validate the proposed model. The whole dataset was split into two distinct sites: site A with 473 pts and site B with 472 pts. A description of the available data is summarized in figure 1. Once set up and launched, 9 iterations were needed for all sites to converge on a result and compute a distributed Cox model and an additional 2 iterations for the algorithm to calculate the ROC and AUC, for a total of 11 iterations over the course of 172 seconds. The results of the computation are shows in figure 2 with an additional comparison between the distributed and centralized approaches.

Conclusion The distributed AUC algorithm we developed fills an important void in the current distributed learning environment, as up until now there was, to the best of our knowledge, no practical way of measuring the performance of a privacy preserving and distributed Cox model. This additional step is fundamental to evaluate the model performance of final distributed models. EP-1938 A high precision irradiation system for in vivo RBE measurements with ion beams J. Besuglow 1,2,3 , G. Echner 4 , A. Mairani 3,5,6 , M. Alber 1,2,3 , E. Bahn 1,2,3,7 1 Department of Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany ; 2 NCT, National Center for Tumor Diseases, Heidelberg, Germany ; 3 HIRO, Heidelberg Institute of Radiation Oncology, Heidelberg, Germany ; 4 Division of Medical Physics in Radiation Oncology, German Cancer Research Center DKFZ, Heidelberg, Germany ; 5 Heidelberg Ion- Beam Therapy Center HIT, Department of Radiation Oncology- Heidelberg University Hospital, Heidelberg, Germany ; 6 National Centre of Oncological Hadrontherapy CNAO, Medical Physics, Pavia, Italy ; 7 Clinical Cooperation Unit Radiation Oncology, German Cancer Research Center DKFZ, Heidelberg, Germany Purpose or Objective Research platforms for experimental small animal irradiation can foster to the acute need of robust in vivo data for current challenges in radiotherapy, such as the determination of the RBE of high-LET radiation in the brain. Due to the strong dependence of dose response on the volume of the small irradiation fields employed in these experiments, highly identical dose distributions are required for inter-comparison of different ion types, which presents a major challenge for experimental design. We devised a multi-component system that meets the requirements of a precise comparative RBE measurement within the pristine Bragg peak of ion beams. Via a passive collimation system, we obtain range adjustment to sub- millimetre precision and sharp lateral beam collimation, combined with precise and rapid animal positioning. Material and Methods The model requires irradiation of the frontal lobe of a mouse brain with a pristine Bragg peak while sparing sensitive adjacent regions. The dose volume is a 3.5 mm x 7 mm wide cuboid of 4 mm depth for proton, helium, carbon and oxygen ion irradiation. This is achieved via passive collimation with a 3D printed bolus that can be