[Houshalter]

That issue is only true for legal reasons. And even then it's entirely speculation what an actual court would decide.

If you actually cared about your patients, then you would use whatever method has the highest accuracy. False predictions mean injury or death. Using a suboptimal method means people die.

The best of both worlds is to use the whatever model gets the best predictions. Then train another model which is understandable on the output of the first one. I.e. generate random data, see what predictions the good model makes. Then the understandable model has infinite data to train with and doesn't need to worry about overfitting.

But still, the utility of being able to understand the model is limited. It's just a big set of parameters, without any reasoning or explanation of why the parameters are what they are.

[closed]

Your account seems to assume that people will faithfully adhere to the suggestions of any model. In reality, not only have statisticians had a hard time accepting more prediction-centered approaches [1, see comments at end], but these approaches may need to win over practitioners and lay-people in the field in which they are being applied (e.g. how much do doctors value prediction over interpretable parameters).

I like machine learning, and prediction centered approaches--but there are many factors (such as adherence both by doctors and their patients) that are important, here. In a sense, the model needs to take into account "model type" into its predictions, which could lead to a model that predicts disease treatments well, but believes it should not be used!

[1]: http://projecteuclid.org/euclid.ss/1009213726