I was just reading a post on LinkedIn regarding the purpose of artificial intelligence. As the author appropriately states, the fact that more recent artificially intelligent [AI] systems are winning at more complicated games [specifically, Go] does not inherently mean that such AI will have a practical use.
At most, one can argue that this particular demonstration of AI is a proof of concept. The game Go is a far more complicated game, in terms of possible moves and gameboard arrangements, than chess. As such, the computer algorithm that made it possible for chess games to win against the top players in the world were not applicable to the game Go.
In order to win at Go against the world’s top players, the algorithm used a much more human approach to the game. Basically, the algorithm scanned the board and looked at the pattern of pieces. Based on this overall pattern a best next move was selected, based upon millions and millions of recorded moves in the computers’ databanks.
One could reduce this down to the simple description of “pattern matching.” Human beings do a great deal of pattern matching as part of their day-to-day functions. We look at a parking space and we intuit whether it is big enough for our car to fit into. We don’t exit our car with a measuring tape and do an exact calculation in order to determine if there is enough room to park. We pattern match to previous episodes of parking and our own 3-D perception of our car.
This system is not perfect. We might think that we can fit into a space that is in fact too small. Contrarily, we may pass on a space that was sufficiently big. But overall, our natural tendency to pattern match works very well. Our ability to identify a person’s face, even if that person has grown a beard, has aged, has a new scar, and other changes, is quite remarkable. And once again, this is not based on a detailed mathematical analysis of the features of the face, but rather on a gestalt impression of the pattern of that face.
The author of the article that I referred to above, notes that home-based healthcare still suffers from a serious lack of information. And with a lack of information, even the smartest algorithms ultimately fail. To be fair, a human doctor would also fail to diagnose a new disease if critical information was lacking.
Let’s say, in a magical world, we could get patients to painlessly and quickly weigh themselves every day. The scale would be connected to a wireless hub, which would transmit the information to the patient’s online record. Acute changes in body weight are in fact a very good indicator of certain problems. A patient with heart failure can gain a significant amount of water weight if their heart failure worsens over a short period of time. The same is true for patients with kidney failure who are on the border of total kidney shutdown. And more so, the same is true for patients with severe liver disease who collect fluid in their abdomen as their liver continues to fail.
Therefore, there is no question that a machine learning system could quickly identify a deterioration in a patient’s health by tracking noninvasive measures. If you add to this single measurement of weight, the patient’s pulse, blood pressure and even noninvasive ECG, the amount of information being collected is tremendous. Suddenly, a smart analysis system could identify a patient with a problematic change in heart rhythm which is causing a worsening of their heart failure and requires immediate care. Conversely, the same data could reassure a patient that there is no immediate evidence of an acute problem, and the patient could book an appointment to see the family physician within a couple of days.
Complaints such as chest pain are definitely far more complicated. Chest pain can be due to anything from a certain type of skin problem, inflammation of the layer of tissue around the lungs or the heart, a pneumonia, a blood clot in the lungs, a problem with the food tube [esophagus] that brings the food from the mouth to the stomach, a stomach ulcer, a back problem, a problem with the ribs, and yes, the heart itself.
And for each of the anatomical locations for the source of pain, there is a whole range of types of problems that could be causing pain. Anything from an inflammation to a full-scale infection, from a tumor to a foreign body to bleeding, all could all be a cause of pain emanating from a particular part of our body. Therefore, it is difficult to diagnose certain kinds of problems over the phone, even with a full home-based examination system backed up by IBM Watson. The day will come when we have home monitoring systems that literally combine the capabilities of ultrasound, CT, MRI and more, that will do a full home-based analysis of our bodies and then transmit that information, via an AI medical analysis system, to a health professional. But this can still be listed as science fiction and is not expected any time in the near future.
In the movie “Contact” [1997,starring Jodie Foster and Matthew McConaughey], a scientist makes contact with an alien race. The scientist is only given a few moments to ask questions, but begs for more time. The alien responds that this interaction was the “first step”. What is to follow are “small moves, small moves”.
When reviewers speak of the weaknesses in AI systems and the lack of applicability of certain technologies to medicine, I find that more often than not, the true problem is lack of vision. I just read a paper yesterday speaking to the failure of a computer system to derive a list of possible diagnoses, for a given patient, based on entered data. As I read the article, I unfairly smiled. The foundation of this diagnostic system was a “rules based” system that is already considered to be outdated.
So, although this was a recently written paper, today, the focus would absolutely be on building a deep learning system that would learn from previous cases that were similar to the present case. Based on millions of previous similar cases (versus a set of rules pre-programmed by the developers), an AI learning system would far more accurately derive a set of possible alternate diagnoses, including of course the most likely diagnosis. The system would also generate a list of tests that would most quickly and accurately diagnose the problem.
With the knowledge that we have today, trying to solve such a problem in any other way almost seems like using a horse-drawn carriage to cross town. Despite the speed at which technology is changing, it still does feel as if each individual discovery is just a baby step along the way to a final system that truly can diagnose a patient better than any doctor.
As usual, I steer away from any predictions as to how long it will take until such systems are literally baked into the standard software that all doctors use. Money, ego, momentum, politics, shortsightedness — all of these play a part in keeping technology from advancing as quickly as possible. Even so, I find it hard to believe that such technologies will take longer than 20 to 30 years to become standard. That means that a medical student starting his or her first-year studies today will face a fundamental and radical change in medical practice right in the middle of their career.
As I’ve said before, if medical schools do not begin to properly prepare medical students for such a future, such schools will quickly become dinosaurs. Students will attend classes, take notes and write exams. But they will learn what they really need to know from alternative sources online. I truly hope that the educational leaders of today understand that the educational leaders of tomorrow will have to be of a totally different brand of teacher and doctor.
Thanks for listening.