Which model would you retain if the test RMSE was not your only choice criterion? Specifically, which model would you retain if you were also concerned about the model's interpretability and implementation, i.e., its everyday use by clinicians and administrators