Towards improved prescription metrics in novel radiotherapy techniques: a machine learning study.

FLASH radiotherapy (RT), microbeam RT (MRT) and minibeam RT (MBRT) are novel RT techniques that have been shown to reduce normal tissue complication probabilities, by modulating the dose distributions through different parameters in space and time. This study aims to investigate the importance of these parameters for predicting biological outcomes using a machine learning (ML) approach and to compare the findings with previous correlation analyses in the context of the current understanding of these techniques.A ML algorithm was trained for predicting normal tissue toxicity, tumor control and increased lifespan (ILS) quantitative metrics on published datasets of preclinical MRT, MBRT and FLASH RT data. The influence of different variables on the performance of the model over unseen data was quantified, and their importance on its predictive power was ranked.An accuracy of 70% or superior was achieved for the prediction of most metrics, reduced for normal tissue toxicity to 60% in MBRT and 40% in FLASH RT. In MRT, valley dose was found as the most influencing physical parameter for normal tissue sparing, while in MBRT the peak dose was highlighted as one of the most influential parameters. Valley dose showed the greatest impact over ILS in a conjoint study of both techniques. In FLASH RT, the total dose, along with the tissue characteristics, were identified as the most influencing variables for tumor control and normal tissue toxicity. The importance of dose rate increased when considering therapeutic index.These results agree with previous studies that highlight how dose heterogeneity prevents normal tissue damage in MBRT and MRT and the need of prescribing under critical tissue specific valley and peak dose values respectively for optimal sparing and tumor control. The described findings are also consistent with FLASH RT tumor control being driven by the same mechanisms as in conventional RT.