Another thing to keep in mind is that a chart is not some purely dispassionate recording of standardized metrics. Even before you get to the diagnosis, you are already looking at pre-processed information, and not raw data.
Although there are some parameters recorded for most patients regardless of the issues at hand (such as vital signs, or maybe listening to the patient's heart/lungs), other history and data is selected -- much of the history, review of systems, and physical exam is performed as a way of supporting or excluding items in a differential that the physician is already thinking about. The list of things that I could observe and write down about a human in a chart is practically unlimited, including a long and useless lists of things that patient doesn't have -- and if I choose to specifically mention what findings are absent, it is because I am making a case for one condition or another..
The other thing to consider is that part of the evidence that is used to support or refute a diagnosis is time. Often I ask a patient to follow-up over the next few days, weeks or months as a way to see if things change (or don't change) the way I expect it too. When a patient follows up with me, I may be making a "second opinion" each time, and the ability to follow someone over time is a valuable tool that gets discarded when someone bounces from doctor-to-doctor searching for answers, without ever looping back to previously visited doctors for follow-up. Sometimes there is a list of rare "zebra" diagnosis that I usually don't write down, because I would usually get laughed at by colleagues for jumping at shadows (but I may keep them in mind if things start not making sense in the future). In any case, a doctor who is asked for a second opinion gets the luxury of having both the first doctor's records (hopefully), as well as a data point occurring later in time.
Anyway, can the thought process for making a final analysis of the case a doctor has made from examining a patient be improved? Certainly, but keep in mind Garbage-In, Garbage-Out, if you feed the AI medical records from someone with excellent vs poor clinical diagnosis skills, the quality of your results will vary greatly. Presumably the best results would be obtained when applied to a doctor with good intuition and observation skills, but poorly organized decision making.