Closed JMathiszig-Lee closed 3 years ago
This is a nice idea. One advantage of using a GAM is that we can show the explicit contribution of each feature to the mortality risk estimate, and we could exploit this to build intuition about the predicted risk distributions. Agree that age is an intuitive feature to adjust by if we are going to present an adjusted distribution for comparison.
One conceptual difficulty is that even the 'patient-specific' predicted risk distributions from the novel model are really just adjusted population average risk distributions like the above (they are just more finely adjusted). There are lots of potentially-important things about the individual patient that the model doesn't know. This makes the comparison between the novel model predictions and an 'average risk distribution' trickier to explain properly.
There are some other things to work out re. obtaining the average risk distribution for age. Simplest thing to do is use the patient's age and the population median values for the other features. Better still to use the median feature values for that age group. Given our emphasis on uncertainty, perhaps we should instead obtain a population average risk by sampling from a distribution over each feature? We have already fit univariate parametric distributions for each feature and are using them in simulate_initial_df
, however we haven't fit bivariate distributions which might be necessary for proper input to the interaction features.
Lots to think about here - shall we pick it back up after finishing manuscript 1?
When deploying the model should we simply display our distribution or is more valuable that there is a comparison also displayed.
E.g age looks likely to be the most significant individual variable, should the patient specific distribution be displayed alongside one for other people of the same age?