By being trained on how that wide range of NDs communicate, what their symptoms are, how medical professionals diagnose them, etc.

That's the problem. The standard for NDs in terms of how they communicate can be literally anything that isn't typical of an NT. Same with symptoms, and even medical professionals can often fuck up diagnoses.

NDs tend to recognize other NDs; if we can do it, an AI sure as hell can.

There are plenty of NDs that are very good at masking. To the point where no one would be able to tell just by looking at them.

And an AI doesn't have the same datasets you do. You can look at their body language, listen to their voice, etc. Any privacy respecting AI will have to go from written language alone. And have fun adapting your model for other languages!