The AI Problem Hidden Inside Medical Data

IBM tried to use AI to predict whether cancer patients would survive. One sentence in the data meant 2 completely opposite things, and the reason why will surprise you.

A funny example from about a decade ago, because IBM wanted to see if they could use— before GenAI, but they could build a neural net around cancer data. And some of the best information input data they had, the output data was, did the patient survive or not? But input data was the doctor's notes. So could they anonymize the whole thing, take the doctor's notes, read them, OCR them, and pull something out of that? And I'm going to give you the example that they gave me because it's really funny. A doctor would see, Betty walks in and Betty says whatever she says, and he writes in his notes, Betty says she's not smoking. Okay, to some doctors that literally means to them Betty is not smoking. To other doctors it means Betty says she's not smoking, but I can smell it on her breath, but I'm not going to put that in the notes. That is such a huge 180-degree difference that without knowing the specific doctor, the specific circumstance, because the notes were only meant for him, there is no way to understand what that data actually meant. So it's 180 degrees out of phase, and it may be one of the most important drivers of, say, breast cancer. Are they really smoking? Are they not smoking?

Are they lying to me, which means they're smoking a lot and they don't even want to tell their doctor?

And so the docs would all write, Betty says she's not smoking, and to half of them it meant she's smoking.