Gravar-mail: Balancing accuracy and interpretability of machine learning approaches for radiation treatment outcomes modeling