2 ms·

A diagnostician which crosses all specialties would be really nice. I imagine this is an area where an easily replicated computing tool might be the best advan

A diagnostician which crosses all specialties would be really nice. I imagine this is an area where an easily replicated computing tool might be the best advancement in the field. Initial symptoms and the low hanging easy tests (as well as any others 'that may as well also be run' for any labs or test types indicated given a moderate extra effort) would drive a root cause diagnosis and recommendation for a specialist who can best treat any of the possibly multiple issues.

tsimionescu 6y ago
The reason doctors become specialized is that they can't learn multiple, and definitely not all, specialities to a good enough degree to be able to diagnose and treat patients across them. There's probably a better chance to have a diagnosis team that meets and handles this type of thing, but it would likely significantly increase the cost of medical care if every patient had to go through a team of specialists. And as for computer aid in this area, it would be awesome, but I fear we are still far away from it, especially since diagnosis is an interactive, continuous process, and is not separate from treatment. The patient presents with some symptoms a history and a current course of treatment, you order some tests and may even prescribe some medication for a preliminary diagnosis, then you wait and adjust as new information comes along. I'm not sure current approaches in machine learning deal very well with this type of process.
- jfengel 6y ago
  I find it remarkable that conventional expert systems didn't succeed at that. It seems like a straightforward application. If you have symptoms X, Y, and Z, then it could be P, Q, or R, so run tests T1 and T2. Obviously I'm not a doctor and can only assume that this is like somebody telling me "Debugging is easy, you just set a breakpoint..." But I'd be curious to know more about why this approach failed.
  - tsimionescu 6y ago
    I think the problem is that the expert system's fact base would have to be essentially the world's most complex medical encyclopedia, except that encyclopedias are addressed to thinking humans, whereas this would have to have enough detail so that it can be "read" by an extremely simplistic algorithm. And not only that, but it would have to be in constant update as new medical knowledge is discovered, new medications (with new interactions) come out etc. I also wonder how well the inference algorithms would actually perform with the scale of facts you would need to add.
- colechristensen 6y ago
  For well understood you don't need complicated machine learning you just need a searchable dictionary of test outcomes and flow diagnosis flow charts, these things already exist.
cycomanic 6y ago
What you describe is essentially a general practitioner. If you look at what they do (and how their training is organised), they essentially follow decision trees to either a clear diagnosis or "refer to specialist". The difficulty for using computers for this tasks is that several of the things going into the tree are based on observations of the patients behaviour, look, speech etc., which is difficult to categorise for a machine.

tsimionescu 6y ago

The reason doctors become specialized is that they can't learn multiple, and definitely not all, specialities to a good enough degree to be able to diagnose and treat patients across them. There's probably a better chance to have a diagnosis team that meets and handles this type of thing, but it would likely significantly increase the cost of medical care if every patient had to go through a team of specialists. And as for computer aid in this area, it would be awesome, but I fear we are still far away from it, especially since diagnosis is an interactive, continuous process, and is not separate from treatment. The patient presents with some symptoms a history and a current course of treatment, you order some tests and may even prescribe some medication for a preliminary diagnosis, then you wait and adjust as new information comes along. I'm not sure current approaches in machine learning deal very well with this type of process.

jfengel 6y ago
I find it remarkable that conventional expert systems didn't succeed at that. It seems like a straightforward application. If you have symptoms X, Y, and Z, then it could be P, Q, or R, so run tests T1 and T2. Obviously I'm not a doctor and can only assume that this is like somebody telling me "Debugging is easy, you just set a breakpoint..." But I'd be curious to know more about why this approach failed.
- tsimionescu 6y ago
  I think the problem is that the expert system's fact base would have to be essentially the world's most complex medical encyclopedia, except that encyclopedias are addressed to thinking humans, whereas this would have to have enough detail so that it can be "read" by an extremely simplistic algorithm. And not only that, but it would have to be in constant update as new medical knowledge is discovered, new medications (with new interactions) come out etc. I also wonder how well the inference algorithms would actually perform with the scale of facts you would need to add.
colechristensen 6y ago
For well understood you don't need complicated machine learning you just need a searchable dictionary of test outcomes and flow diagnosis flow charts, these things already exist.

cycomanic 6y ago

What you describe is essentially a general practitioner. If you look at what they do (and how their training is organised), they essentially follow decision trees to either a clear diagnosis or "refer to specialist". The difficulty for using computers for this tasks is that several of the things going into the tree are based on observations of the patients behaviour, look, speech etc., which is difficult to categorise for a machine.