Perhaps the more likely scenario will be a compromise between AI taking over the world and it failing—AI functioning as another valuable tool.

Benjamin Fels is the CEO of macro-eyes, a company on the frontlines of AI’s deployment in medicine. Together with his small team of experts he is working out the challenges of making use of the present scattered medical data. As product manager, his work entails meticulous details of how the software interacts with physicians and patients alike. His daily decisions negotiate the relationship between human and machine. Our conversation makes it clear that as a society, we still have a long way to go before physicians can be fully replaced by AI/technology. 

How are you convincing doctors to give up part of their jobs to a machine?

Algorithms do things that we typically don’t understand. You need to trust that even if it’s going to make decisions which confuse you, on some basis, there’s some structural logic which you agree with. And that comes down to the interface, in terms of how you see and understand those decisions and that logic.

We want to empower all physicians to be able to observe decisions that a machine makes and understand why the machine identifies certain patients or events as similar. Then, the doctor can start to interact with higher levels of insight or abstraction. To some degree, doctora will even be able to jump to the future; even see what is likely to happen to the patient.

Where is the “intelligence” in the technology you’re working on, versus it being simply a very sophisticated search engine?

What we’re creating is an ability to find patients or medical events that are similar across hundreds of dimensions [blood type, medical history, etc.], and where any one or many of those dimensions could be missed by the human eye/by a human. What we’re really thinking about is the common essence of the patients or events we’re taking into consideration.

“Similarity”, at a mathematical and conceptual level, is a challenging problem to solve. One of the first things that we built was a way for the doctor to say, “Okay, this match you showed me is particularly good or bad”; and then the machine is able to “learn” based upon the input of the physician.

It seems like the critical element to keeping trust is to keep the doctor physicians feeling like they’re the decision maker. And indeed, they are the decision maker. I have no interest in building some kind of robot doctor, because there are several reasons why that wouldn’t even be useful. And I also believe the technology and infrastructure is not even really there yet. 

This is exactly where I’m highly critical of AI. I want to emphasise that I’m building technology for an environment in which an expert, the physician, interacts with an intelligent system, the AI; and the expert is bringing knowledge and insight that the system doesn’t have. It’s important to understand that I’m not saying data; I’m saying information, which is one step up.

We humans observe things and know things which are not captured in data and that’s because we luckily don’t yet live in a world where every particle of reality is monitored. 

Doctors always talk about gait—the way a patient walks or how they show pain—and that is one example of all-important information in the form of clues, which are hard to get from just looking at data, at least, the data we are currently able to capture.

In the end, what truly matters is whether the data reflects the ground truth. Then, it’s a question of how to bring into our intelligent machine what we humans know, see or understand and that a machine doesn’t. For me, that is the Holy Grail. A paintbrush plus a human using that paintbrush in an expert fashion can have tremendous impact in the real world.

What’s are some of the biggest limitations to AI’s progress in healthcare?

If you have bad data and you feed it to the world’s smartest machine, the machine is going to spit out something that is gibberish. That’s also where we have to get back to this interaction between humans and machines. Even getting a machine to clean data by itself can be an extremely tricky task, because the machine needs to understand what data is—what’s correct data, what’s incorrect data, what variables or values there could be and when something is an anomaly that makes sense or an anomaly that’s entered incorrectly.          

The analogy I always give is that we have a customer whose refrigerator breaks, so he calls us on the phone and says, “My refrigerator just broke but you can’t come see it. Can you fix it?” It’s hard to fix a problem without being able to see it in depth, but that’s necessary because our customers often work with data that contains protected health information. 

The most important point I want to make is that data itself has to reproduce reality. When you start to think about even very small instances of data not reproducing reality, it’s altering the truth. If decisions are made based upon data that has affected reality, we stumble into dangerous scenarios—extremely dangerous, medically speaking.

Would you say people are over-excited about AI?

There is so much hype around AI and machine learning; when that hype meets industrial and enterprise customers, and it doesn’t produce the desired results in three seconds, I am worried that they will become frustrated. They might say, “Screw this. We thought you guys could predict the future in two hours and solve all our problems. Obviously, you can’t and we aren’t interested anymore.” 

Look back to 20, 25 years ago when we saw the first boom in machine learning and AI. We basically had a drought because customers of technology said, “Yeah, we tried it, it didn’t work.” So when you overhype things and downplay the incredible difficulty of getting the technology to work, you make people think it’s plug & play.

Even software we run on our computers has a learning curve—we must understand that it doesn’t work immediately, right out of the box.