notebook for imputation methods using PhysioNet2012

[sueoglu]

fixes theislab/ehrapy#1061

• Adds imputation_nb.ipynb, a new tutorial that walks through ehrapy's six imputation methods (explicit_impute, simple_impute, locf_impute, knn_impute, miss_forest_impute, mice_forest_impute) using the PhysioNet Challenge 2012 ICU dataset.
• All methods are benchmarked on a 200-patient subset by masking 10% of observed values and comparing RMSE/MAE, mean/std deviation from ground truth, and runtime/memory cost (last one in progress)
• Closes out with a decision guide (which method to pick based on variance sensitivity, speed needs, temporal structure, or cluster structure) and applies mice_forest_impute (we could also use locf) to the full dataset with before/after sanity checks.

side note: will see how to fix the rendering problem especially for the raincloud plots since its a little weird atm

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:20Z
----------------------------------------------------------------

no tem_data layer requried anymore

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:21Z
----------------------------------------------------------------

Not "manageable", but better "quick for demonstration purposes"

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:22Z
----------------------------------------------------------------

We use this masking strategy because real-world NaNs in the dataset have no ground truth, so direct accuracy evaluation on them is impossible.

This sentence can be completely deleted

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:23Z
----------------------------------------------------------------

you can save the need for 2 lines of code if you instead do

edata_explicit = p.pp.explicit_impute(edata=edata_explicit, layer="tem_data", replacement=0, copy=True)

this reads a bit shorter. also for the others

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:24Z
----------------------------------------------------------------

Avoid calling this three dimensions; three dimensions is already used above to describe the data shape.

Call it better "three aspects" or so

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:25Z
----------------------------------------------------------------

You can add the reduce miss-forest already on this figure, just among the others - the notebook is quite comprehensive, and adding this here makes the information more condensed :)

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:25Z
----------------------------------------------------------------

again, extra line for copy can be avoided with the copy=True argument

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

eroell commented on 2026-08-04T17:39:26Z
----------------------------------------------------------------

Can we have a closing word afterwards? A small wrapup paragraph, just 2-3 sentences roughly, mentioning e.g. the textbook you refer to for a deeper discussion