A model that's accurate only 50% of time is far from helpful in terms of public health: it's high enough so that people could trust it and low enough to cause harm by misdiagnosing stuff.
If you want it to be used by the public it doesn't matter if it's more accurate on some things if it's very bad at other things and the user has no idea in which situation we are.
As a senior developer I routinely use LLMs to write boilerplate code, but that doesn't mean that the layman can get something working by using an LLM. And it's exactly the same for other professions.