Study Finds Machine Learning Alone Insufficient to Replace Physicians' Expertise
-
Pay attention to learning confounds - the ML model was designed to learn from data, but physicians never got the chance to learn from outcomes over time.
-
Don't take "biases" at face value - heuristics like availability and representativeness are generally useful for physicians.
-
Look out for smuggled expertise - the ML model relied on judgments and expertise embedded in the electronic health records.
-
Don't be cowed by big data - though the model had over 16K variables, predictions plateaued at around 20 variables.
-
Make use of ML to supplement expertise - the authors suggest using ML to help train physicians rather than wholly replace human judgment.