Mahmoud Bentriou ^1* Veronique Letort ¹ Stefania Chounta ^1,2,3,4 Brice Fresneau ^2,3,4,5 Duyen Do ^2,3,4 Nadia Haddy ^2,3,4 Ibrahima Diallo ^6,7 Neige Journy ^2,3,4 Monia Zidane ^2,3,4 Thibaud Charrier ^2,3,4 Naila Aba ^2,3,4 Claire Ducos ^2,3,4 Vincent S Zossou ^2,3,4 Florent De Vathaire ^2,3,4 Rodrigue S Allodji ^2,3,4,8 Sarah Lemler ¹

• ¹ Université Paris-Saclay, CentraleSupélec, Mathématiques et Informatique pour la Complexité et les Systèmes, Gif-sur-Yvette, France
• ² INSERM U1018 Centre de Recherche en Épidémiologie et Santé des Populations (CESP), Villejuif, Île-de-France, France
• ³ INSERM, CESP-U1018, Radiation Epidemiology Team, F-94805, Villejuif, France
• ⁴ Gustave Roussy, Department of Clinical Research, Radiation Epidemiology Team, F-94805, Villejuif, France
• ⁵ Gustave Roussy, Department of Pediatric oncology, Villejuif, France
• ⁶ Gustave Roussy, Department of Radiation Oncology, Villejuif, France
• ⁷ Université Paris-Saclay, INSERM, Gustave Roussy, Radiothérapie Moléculaire et Innovation Thérapeutique, Villejuif, France
• ⁸ Polytechnic School of Abomey-Calavi (EPAC), University of Abomey-Calavi, 01 P.O. Box 2009, Cotonou, Benin

Background: Cardiac disease (CD) is a primary long-term diagnosed pathology among childhood cancer survivors. Dosiomics (radiomics extracted from the dose distribution) have received attention in the past few years to assess better the induced risk of radiotherapy (RT) than standard dosimetric features such as dose-volume indicators. Hence, using the spatial information contained in the dosiomics features with machine learning methods may improve the prediction of CD. Methods: We considered the 7670 5-year survivors of the French Childhood Cancer Survivors Study (FCCSS). Dose-volume and dosiomics features are extracted from the radiation dose distribution of 3943 patients treated \vero{with} RT. Survival analysis is performed considering several groups of features and several models (Cox Proportional Hazard with Lasso penalty, Cox with Bootstrap Lasso selection, Random Survival Forests (RSF)). We establish the performance of dosiomics compared to baseline models by estimating C-index and Integrated Brier Score (IBS) metrics with 5-fold stratified cross-validation and compare their time-dependent error curves. Results: An RSF model adjusted on the first-order dosiomics predictors extracted from the whole heart performed best regarding the C-index ($0.792 \pm 0.049$), and an RSF model adjusted on the first-order dosiomics predictors extracted from the heart's subparts performed best regarding the IBS ($0.069 \pm 0.05$). However, the difference is not statistically significant with the standard models (C-index of Cox PH adjusted on dose-volume indicators: $0.791 \pm 0.044$; IBS of Cox PH adjusted on the mean dose to the heart: $0.074 \pm 0.056$). Conclusion: In this study, dosiomics models have slightly better performance metrics but they do not outperform the standard models significantly. Quantiles of the dose distribution may contain enough information to estimate the risk of late radio-induced high-grade CD in childhood cancer survivors.