A Supplementary Material for Neural Additive Models A.1 NAMs on MIMIC-II: Mortality Prediction in ICUs

NAMs (like all GAMs), are not causal models. Although the shape plots can be informative, and can help uncover problems with the data that might need correction before deploying the models, the plots do not tell us why the model learned what it did, or what the impact of intervention ( e.g.,