Challenging a clinical dogma with multimodal machine learning: a retrospective analysis of transplant mismatched donor selection