"Artificial Intelligence - The Revolution Has Not Happened Yet"
Public.icon
When my wife was pregnant 14 years ago, we had an ultrasound. A geneticist was present and pointed out some white spots around the fetus's heart. "Those are markers for Down syndrome," she said, "and your risk has gone up to 1 in 20." We were told that an amniocentesis could determine if the fetus actually had the genetic mutation that causes Down syndrome, but the procedure carried a risk of about 1 in 300 of the fetus dying during the operation. As a statistician, I decided to investigate where these numbers came from. My research revealed that a statistical analysis had been conducted in the UK 10 years ago, which established these white spots reflecting calcium deposits as a predictor for Down syndrome. I also noticed that the imaging machine used in the test we received had several hundred more pixels per square inch than the machine used in the UK study. I went back to the geneticist to tell her that I believed these white spots were likely false positives. "Oh, that's why the number of Down syndrome diagnoses started to increase a few years ago. That's when the new machines came in."
We did not undergo amniocentesis, and my wife gave birth to a healthy baby girl a few months later, but the episode troubled me. Especially after calculating on the back of a notebook showing that many people around the world received the same diagnosis on the day I handled it, many people chose amniocentesis, and some babies died unnecessarily. The problem revealed by this episode was not just a personal medical issue for me, but one about medical systems that measure variables and outcomes in various places and times, conduct statistical analyses, and use the results in other situations. This problem was related not only to data analysis itself, but also to what researchers of "provenance" call, where the data came from, what inferences were drawn from the data, and how those inferences were related to the current situation. Trained humans may be able to solve everything on a case-by-case basis, but the problem was to design a global healthcare system that could do this without the need for such detailed human supervision.