|
Abstract:
|
With the emerging incorporation of the machine learning methods in healthcare prediction problems, there has been recent skepticism articulated with respect to their added value in comparison with the mainstay logistic regression. In fact, more appropriately, the discussion should be framed in the context of linear machine learning methods vs their nonlinear counterparts. The key question is whether linear methods are enough for healthcare prediction. In our work, we discuss the underpinnings of the linear prediction models and their merits such as widespread use and interpretability. While non-linear models are often considered better for “pure” predictions, these model predictions can directly be used for interpretation through metrics such as variable importance, partial dependence plots/accumulated local estimates, and causal estimands. We also discuss the consequence of no free lunch theorem for prediction and present analysis of a real-life data set from metformin adherence that demonstrates this point. Furthermore, we show how recently proposed approaches to variable importance can be leveraged to provide important insights for a given prediction problem.
|
