Back in March, I wrote about Edward Frenkel and his explanation of the purported backdoor in the Dual_EC_DRBG random number generator. I also mentioned his book Love and Math: The Heart of Hidden Reality. Now Zygmunt Zając over at FastML has a nice post on one of the stories from the book.
It’s about how Frenkel hand-constructed a decision tree to determine the treatment plan for kidney transplant patients. What’s interesting is that Frenkel has no medical training or domain knowledge. Rather, he worked with a physician to capture the physicians ad hoc decisions and turn them into an algorithm.
The story is interesting to folks like us because of the way he solved it. It’s a method that developers might try when attempting to capture a human decision process into a program. At first he asked the physician a series of question trying to understand the various parameters and how they related to a final decision. That approach was singularly unsuccessful so Frankel took another approach.
He randomly selected some patient records and had the physician ask him questions, which he answered by consulting the patient’s file. By analyzing the questions, the order in which they were asked, and what the follow up questions were, he was able to construct a decision tree that accurately captured the physician’s subjective thought process. After two and a half dozen cases, Frankel’s algorithm was almost as good as the physician at making diagnoses. By the end of the process his diagnoses were 95% accurate.
In the end, the process was good enough to patent. What I like is how simple and direct the process was and yet it produced outstanding results. I really recommend this post: it’s short but clearly outlines an approach to problem solving that you may find useful.