The aim of this study was to implement a clinically deliverable VMAT planning technique dedicated to advanced breast cancer, and to predict failed QA using a machine learning (ML) model to optimize the QA workload. Methods: For three planning methods (2A: 2-partial arc, 2AS: 2-partial arc with splitting, 4A: 4-partial arc), dosimetric results were compared with patient-specific QA performed with the electronic portal imaging device of the linac. A dataset was built with the pass/fail status of the plans and complexity metrics. It was divided into training and testing sets. An ML metamodel combining predictions from six base classifiers was trained on the training set to predict plans as 'pass' or 'fail'. The predictive performances were evaluated using the unseen data of the testing set. Results: The dosimetric comparison highlighted that 4A was the highest dosimetric performant method but also the poorest performant in the QA process. 2AS spared the best heart, but provided the highest dose to the contralateral breast and lowest node coverage. 2A provides a dosimetric compromise between organ at risk sparing and PTV coverage with satisfactory QA results. The metamodel had a median predictive sensitivity of 73% and a median specificity of 91%. Conclusions: The 2A method was selected to calculate clinically deliverable VMAT plans; however, the 2AS method was maintained when the heart was of particular importance and breath-hold techniques were not applicable. The metamodel provides promising predictive performance, and it is intended to be improved as a larger dataset becomes available.