I think the main issue is that most mental illnesses are vaguely defined clusters of problems - if you represented a diagnosis as a vector where each index corresponded to severity of a particular symptom, the diagnosis space would not properly aligned with the actual disease basis. This introduces ambiguity - a single disease tends to fall into multiple clusters when your representational domain is misaligned with the actual data axes.
The solution is ML. Neural networks with appropriate architectures effectively perform a change of basis, mapping the data basis to an output coordinate system. When properly trained they can automagically orient their internal representation with the true data bases.
I agree with your first paragraph. However, a NN can't explain its reasoning. You start with a set of symptoms, and you end with a vector, and they're two representations of exactly the same thing. The problem is that NNs are representational, not explanatory, so they won't help a doctor get to a root cause any more than just thinking about the collection of symptoms and what they're associated with.
Yes: but what was implicit in mine is that that output vector is no more or less informative than the current method of associating strong symptoms with specific diagnoses which have known drug dosages. Same idea, different representation.
The difference is whether the process can be automated, and that is where the problem of the black box comes in. With the current system, we know why the drug was prescribed. With an ML system, we would have no clue.
The solution is ML. Neural networks with appropriate architectures effectively perform a change of basis, mapping the data basis to an output coordinate system. When properly trained they can automagically orient their internal representation with the true data bases.