ESTRO 2021 Abstract Book

S764

ESTRO 2021

23 Hôpital universitaire Robert-Debré, Pediatric Surgery, Paris, France; 24 Princess Máxima Center for pediatric oncology, Pediatric Oncology, Utrecht, The Netherlands; 25 Alder Hey Children’s Hospital, Department of Paediatric Oncology, Liverpool, United Kingdom; 26 Centre Hospitalier Universitaire de Reims, Service d'oncologie pédiatrique, Reims, France; 27 Hôpital La Timone, Pediatric Oncology, Marseille, France; 28 Centre Hospitalier Universitaire Necker, Pediatric Surgery, Paris, France; 29 Aarhus University hospital, Oncology Department and Danish Center for Particle Therapy, Aahrus, Denmark; 30 Sapienza University of Rome, Pediatric Oncology Unit, , Rome, Italy; 31 Centre Hospitalier regional et Universitaire Clocheville, Pediatric Oncology, Tours, France; 32 Groupe Hospitalier Pellegrin Hôpital des Enfants, Pediatric Oncology, Bordeaux, France; 33 Universitäts-Kinderspital beider Basel (UKBB), , Pediatric Oncology, Basels, Switzerland; 34 Aarhus Universitetshospital , Hematology and oncology, Aahrus, Denmark; 35 University Hospital in Bydgoszcz, Poland, Department of Pediatric Hematology and Oncology , , Bydgoszcz, Poland; 36 Gustave Roussy Cancer Campus, Department of Pediatric and Adolescent Oncology, Villejuif, France; 37 Bicêtre Hospital. Assistance Publique- Hôpitaux de Paris, Department of pediatric surgery, le Kremlin Bicêtre, France Purpose or Objective Childhood cancer is a rare disease. Disparities in survival and long-term side-effects encourage the establishment of networks to increase access to complex radiotherapy procedures, such as brachytherapy (BT). We report our 50-year experience of BT for childhood cancers treatment. Materials and Methods We examined the outcome of all children referred to our center from national or international networks between 1971 and 2020 and treated according to a multimodal approach including BT. Treatment characteristics were reported and patient outcome examined with focus on local control, survival, and probability of severe complication. Results A total of 305 patients were treated, median age at diagnosis 2.2 years (1.4 months–17.2 years). 99 (32.4%) were treated within the last five years. 172 (56.4%) were referred from national centers and 133 (43.6%) were international patients, from 31 countries (mainly Europe). Genito-urinary tumors were the most frequent sites, with 56.4% bladder/prostate rhabdomyosarcoma and 28.5% gynecological tumors; other sites were head and neck tumors (6.9%), perineum (4.9%) and limbs (3.2%). Tumor size at diagnosis was ≥5 cm in 46.2%. Prior to BT, all rhabdomyosarcoma (RMS) patients had received primary chemotherapy, median number of 6 cycles (range: 3–18 cycles). In addition to BT, local treatment comprised a partial resection of primary tumor in 207 (67.9%) and 39 had additional external radiotherapy. Catheters placement was performed as perioperative procedure in 180 (59.0%). 225 (73.8%) patients had interstitial BT only, 68 (22.3%) had intracavitary BT only, and 12 (3.9%) had a combined technique. Median BT dose was 60 Gy (range: 10–80 Gy), delivered through low dose rate (59.0%) or pulse-dose rate BT (41.0%). Median follow-up was 58 months (range: 1 month – 48 years). It was 93 months for national patients and 37 months for international patients (p<.0001). 79 patients (26%) had a follow-up time > 10 years. At last follow-up, 16.4% patients had long-term severe complications. Five- year estimated local control (LC) was 91.2% (CI95: 87.8–94.7%), disease-free survival (DFS) was 84.7% (CI95: 80.4-89.2) and overall survival (OS) was 93.2% (CI95: 90.1-96.5) . In multivariate analysis, alveolar RMS histology was poor prognostic factor for DFS (HR: 5.47; 95%CI: 2.65-11.30, p<0.01) and OS (HR: 4.46; 95%CI: 1.80-11.06, p<0.01). The only factor associated with LC was referral at time of relapse (HR: 3.03; 95%CI: 1.34- 6.83, p<0.01). Two patients (0.7%) developed second malignancy. Conclusion A multinational collaboration is possible, offering patients an access to highly specialized treatments such as brachytherapy. Despite paths for improvement (especially for long-term follow-up in international patients), this cooperation model could serve as a basis for generating international reference networks for high-tech radiation to increase treatment care opportunities and cure rates. PD-0924 Machine learning (ML) for predicting patient-reported symptoms during breast and prostate RT J. Kononen 1 , J. Ekström 2 , S. Mentu 2 , V. Kataja 2 , L. Lang 2 , H. Virtanen 2 , M. Metso-Lintula 2 , T. Joensuu 3 , C. von Briel 4 1 Docrates Cancer Center, Personalized Oncology, Helsinki, Finland; 2 Kaiku Health, Medical Science, Helsinki, Finland; 3 Docrates Cancer Center, Radiotherapy, Helsinki, Finland; 4 Hirslanden Medical Center, Institut für Radiotherapie Aarau, Aarau, Switzerland Purpose or Objective Electronic collection of patient-reported symptoms (ePROs) during and after RT treatment as part of routine clinical care can provide meaningful insights on patient experience and increase the feeling of safety in patients. The objective of this study was to explore to what extent ePROs collected from breast and prostate cancer patients could be used to predict the occurrence of RT-related symptoms by utilizing ML. Materials and Methods The patient population consisted of 409 breast and 363 prostate cancer patients who had reported their symptoms (based on CTCAE) through the Kaiku Health platform as part of routine care. The breast cancer ePRO dataset consisted of 2106 and the prostate cancer dataset of 3362 patient filled symptom questionnaires. The symptom prediction was performed using an approach based on the XGBoost ML library. Using supervised learning, the models were trained to predict occurrence of symptoms during the next week based on the three previous ePRO reports, as well as the age of the patient and the time from the start of treatment. Three occurrence severities were predicted if it was feasible with respect to the data (grade 0, 1 or 2-3), otherwise two were used (grade 0 or 1-3). The prediction performance was evaluated on a test dataset, excluded from model training, consisting of 30% of the data by using Matthew’s correlation coefficient (MCC), which considers both true and false positives and negatives. MCC returns values between -1 and 1 where 0 indicates random guessing and 1 perfect classification.