3
$\begingroup$

I’m currently working in a survival analysis setting with the goal of risk prediction.

In earlier work, I used static (baseline) covariates together with Random Forest–based survival models, and I relied on SHAP for model-agnostic explainability of the predictions.

I am now trying to explicitly exploit longitudinal / time-series data, and I have started working with joint models for longitudinal and survival data. In particular, I am using the JMbayes2 package (mixed-effects models for the longitudinal part + Cox model for the event process).

Since I am relatively new to this framework, I am struggling to understand how to perform explainability of the predictions in a model-agnostic way, similar to what SHAP provides for machine-learning models.

Even though joint models are parametric and fully specified, which would suggest that interpretation should be relatively straightforward, I have not been able to find papers or examples that implement SHAP-like or model-agnostic explainability for joint models, especially at the level of individual predictions or dynamic risk estimates.

Do you know of any references, papers, or methodological discussions addressing explainability or local interpretation for joint models? Any pointers or suggestions would be greatly appreciated.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

SHAP has little to do with explainability and doesn't work well if predictors are collinear. This approach will not even reveal how much of relationships are nonlinear or interactive. Take a look at relative explained variation and partial effect plots.

Random survival forests tragically overfit the data, resulting in the worst calibration curves I've seen, and are also uninterpretable. We once had to spend 6 months tweaking a survival forest to even approach getting a calibration curve that is near the line of identity.

But to your main, interesting, question, interpreting joint models is a great area for research and I can't think of a practical paper that shows the way. If it is logical to reframe the problem as a longitudinal ordinal model then interpretation is easy, e.g., how does a predictor relate to mean time in a set of states, to transition probabilities, or to state occupancy probabilities. This way of modeling works best when there is a single continuous outcome and bad events that override that outcome. Per time period you form an ordinal outcome having $n + m$ levels where $n$ is the number of distinct values in the continuous marker and $m$ is the number of types of ordered outcome events.

Improve this answer
$\endgroup$

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.