Add a Comment
Save my information
The term “black box” is misleading. AI predictions are not made based on inscrutable mechanisms unknowable in principle.
That users of these models do not fully understand their mathematical/statistical underpinnings does not make them more dubious than statistically complex randomized controlled trials using multiple regression, survival curves, NNTs, and other statistical tools that most users of these reports understand in only superficial ways.
When the recommended interval between mammogram screenings is changed in a guideline, the follower of the new guideline is, for the most part, insofar as he/she is concerned in practice, relying on a “black box”.
Of course, it is important to have some understanding of the variables and the weights they are given in the model.
These are not value-free.
If the model is used to determine action, e.g., treat/do not treat, does it take into account cost, tolerability, etc.?