This exists. Current implementations work badly enough even in the trivial case that they are completely useless. Making a better implementation would involve technology that we don't have yet. Also, https://news.ycombinator.com/item?id=27428671
Automated diagnosis and clinical recommendation engines are seeing a lot of research, yes. As a clinician unfortunately, I can already tell you that all that will lead pretty much nowhere. The first problem to solve is data collection/information retrieval, and the solution is not "see, with this clean data my model achieves 101% accuracy, that's better than a human" but will be a systemic effort to enforce policies about data collection and storage. Only then can we hope to use this data consistently.
I did research in physics - something much simpler than the human body so please take this as a disclaimer :)
More than an AI kind of system, I was thinking about something more predictable, intended to supplement the MD memory.
I imagine that if you input the symptoms you see, some basic measurements (temperature, blood pressure, O2 saturation, ..) + some input from the general news (flu season, allergy season, ...) you end up with the typical diagnosis.
But then, there could be other pathologies th eMD does not think about immediately but when displayed would ring a bell.
My wife for instance had several mis-diagnoses after our second child, only to find with a doctor that she may have a neurological issue. Then came the neurologist. Then came MS.
In her case time was not that of an issue but maybe in other cases that would help.
Soless AI and more "decision tree" kind of software.
I hear you. Like I said, all attempts using decision trees have up to now failed miserably. The main problem is that for any given set of symptoms, the ~5 most likely diagnoses will be very obvious to most docs, but there will be a gigantically long distribution tail of less likely diagnoses not even worth testing for. This is especially problematic for rare diagnoses, and I'm sorry to say that your wife unfortunately fell into that category. Then come all the other UX issues, such as what search terms to use (different users will use different names), etc.
I know a decision tree seems attractive to someone coming from science. But in clinical practice, it's totally useless.
As an input you would provide the typical information doctors look for otr get from the user (cough, the elbow hurts, ...) and then suggest
It would then suggest a set of tests to got to XX% where the diagnosis is "good enough".It would help to avoid missing a broken ankle when this is flu season and the MD assumes that this is "like everyone".
(the examples are extreme for some lightness)