How a Russian mathematician constructed a decision tree - by hand - to solve a medical problem

Abstraction: 1980s hand-built decision tree for kidney transplant rejection diagnosis

Key points:

• Edward Frenkel (mathematician, then student) built a decision tree in 1980s Soviet Russia by interrogating doctor Sergei Arutyunyan about kidney transplant rejection decisions
• Dataset: ~270 patients, ~40 parameters each; Frenkel set aside 30 as training set and used 240 for testing
• Method: replayed patient cases to the doctor and recorded question-order patterns to reverse-engineer his decision logic
• Resulting tree uses only 4 parameters (including peripheral resistance index) with at most 4 binary questions per path
• Achieved ~95% accuracy on the held-out 240 patients; Sergei always reached correct diagnosis within 3-4 questions
• Excerpt from Frenkel's book "Love and Math"; illustrates knowledge elicitation as a precursor to modern ML

Connections: Edward Frenkel · Decision Trees · Knowledge Elicitation · Machine Learning

Source: http://fastml.com/how-a-russian-mathematician-constructed-a-decision-tree-by-hand-to-solve-a-medical-problem/