Extension to DeepHit

[Anivader]

Hello,

This is great work!

I was wondering if SurvSHAP can be applied to Deep Learning survival Networks such as DeepHit that can produce survival function estimates without being constrained by the "Proportional Hazards (PH)" assumption.

Please let me know. I am keen to try it out. Ani

[krzyzinskim]

Hi, thanks for your interest!

SurvSHAP(t) can definitely be used for deep learning models, there are no theoretical counter-arguments for that. In fact, it is a good idea because the lack of PH assumption makes DL models more flexible, but thus also more complex and complicated - it is useful to know what is going on inside such a black-box model.

As for the practical side of this problem - currently this Python package is adapted to work with models from the scikit-survival package, where no DL models are available. Other models can, of course, be explained, but their prediction interface must then be adapted to return predictions in the form of sksurv.functions.StepFunction.