i was debating this with some doctors who say that it would be impossible to program a computer to make diagnoses as well as they can. i find this pretty ridiculous. whatever thought process/string of questions they would use to analyze the situation are the same that the computer would be programmed to use. the computer would then analyze all available information, ask questions, analyze the answers and assign probabilities. in fact, it seems like this would be way simpler than some of the things computers have already been programmed for. what do u think?

no, your friends are probably right. There is a reason we haven't been able to make a good bot for NLHE

evolution wins this one

though I think doctors with computers would be better.

Main problem is that people are bad at giving answers. I remember reading that computers do beat doctors at diagnosis provided doctors or nurses are putting the symptoms/patient history in. Not sure how true that is, but it seems highly likely. Once you have the symptoms, diagnosis is database matching with some database based probabilities and math. No question computers can do that better than doctors.

But yeah...the problem is that people are stupid. Have you ever used a Windows Help troubleshooter? How well did it work? Did you ever get an old relative to use it?

There are some excellent bots for NLHE. But they're not public, and they don't use the traditional overly broad computer science models - the good bots are very specific. They have a lot more hand tweaking and focus on tediously writing out complex algorithms and developing models for all situations rather than a generalized approach.

I call BS and HU4ROLLZ vs alleged bot

This is not true. Not one person in that thread said it was impossible. Just not possible at the moment.

As for the actual question at hand, you're basically implying that we can program a computer to do anything. I can describe any task as you just did, but that doesn't make it easier for the computer. As I said in the other thread, computers have been trying to tackle radiology and pathology for years now. They're made progress, but they aren't at the human level yet.

Watson?

OP, diagnosing patients involves parsing language, which is an incredibly difficult task in AI. What would be a simple task in the vein you are thinking probably involves tasks post-diagnosis, like checking conflicts for new prescriptions, or determining proper dosages based on medical history and current best practices.

Thinking about it further, has anyone tried an Akinator 20 question style algorithm to diagnose patients? A few thousand med students trying to stump a Dr. Akinator could train it quickly.

In one of Atul Gawande's books I think it was complications, he talked about how a computer program beat a top cardiologist at reading ECGs. But I doubt a computer will ever be used to inform a patient about their life threatening condition.

doctors + computers is the way to go here imo

So what sort of magic are these doctors performing? Is their conclusion not rationally derived?

Expert systems are difficult to make but I see no reason why one couldn't give good diagnoses.

Not sure how a computer tests for a hernia.

Computers can be great tools for doctors to use though, and they can do much more with them.

I look forward to the day that when I am sick and need to see my doctor, I log onto his web site, schedule an appointment, and input all of my symptoms and relevant information.

Then when I show up for the appointment, he has all the basic info already including a computer generated short list of possible explanations for my symptoms. Then he asks me the questions he needs to ask and proceeds with any appropriate physical examination.

The primary benefit of this for me would be not having to deal with the doctor's almost certainly annoying staff in setting up the appointment, but it should also improve efficiency for the doctor.

Someday certainly computers will be able to diagnose better than a doctor, there is no way that technology would just stop before that point.

That being said, I have no qualms about going to medical school right now because I think this is a long way off and as it starts to be developed its going to require a massive amount of input and monitoring from technology oriented physicians so my engineering degree might come in handy.

One thing that I think will be tricky to overcome is the fact that some much of diagnosis is based on the severity of symptoms. For example, while there are pain scales, I hear patients responses to them are generally pretty useless and that physicians more often make a judgment based on the patients appearance and demeanor that would be hard to quantify if the computer was functioning completely alone.

Wasn't there a chapter in Blink about this? (not exactly but it was a similar type of question)

Basically someone did statistical analysis to determine which factors were correlated with someone having a heart attack vs. some other chest pain, and a very simple 3 question examination was more accurate (resulted in both less type I and type II errors) than doctors using their own judgment were able to.

I dont think this is a hypothetical, I think this study has been done, and the computers won. If you were sufficiently motivated to create such a program with any sophistication I think you would win hands down.

The things you would be missing with your computer program is all the subtle pattern-recognition that clinicians pick up and aid them in diagnosis, that they couldnt even put into words and probably dont know what they are doing. This is a big advantage for the doctors over the computers.

However, the thing you avoid with the computer program is the huge amount of bias (availability heuristic, confirmation bias, etc.) that plague most doctors. If you walk into a hospital with the classic symptoms of some extremely bizarre, rare disease, it is far more likely that you do NOT in fact have that disease, even though you look like the literal textbook example, but most doctors are terrible at Bayesian analysis and would call it Syndrome X. The computers would presumably never make this error. The computers would accurately diagnose the common condition more often, and miss the subtle or rare conditions more often.

I think the computers would win by smallish but significant margin. The reason its smallish is because they would both get the right diagnosis the majority of times.