> Each patient's case is unique and cases with the highest levels of difficulty are often treated by the most experienced people. These cases, of course, are likely to have worse outcomes than simple cases which may be treated by less experienced (worse?) physicians.

Every individual is unique and yet we are able to segment folks using multiple variables and draw predictions for a bunch of things across businesses - it's just like saying "every shopper online is unique", which is true, yet we still find ways to influence them of target them in somewhat meaningful ways.

You should be able to do the same with patients, by clustering them in larger groups that make sense (for cancer many trials refer to specific mutations nowadays to define whether a patient is expected to get efficacy from a specific drug).

The only question in the end is whether or not you have enough sample size for each doctor. Maybe not. Then it would make sense to at least draw comparisons between groups of specialists across multiple hospitals, at least.

Even if metrics are imperfect to begin with, there is no excuse not to start at least, and improve over time.