PFI is shown here in the wrapper-method context of Raschka’s lecture.
Original age column
| Row | Age aᵢ | Low oxygen | Observed outcome | Original f(xᵢ) | Permuted f(xᵢ perm) |
|---|
Schematic symbols, not patients: aᵢ = age; oᵢ = low oxygen; yᵢ = target. f(·) denotes a model prediction. No row-level probabilities or outcomes are displayed.
One column at a time
Full study test set · Brier score · lower is better
The original error is the reference. The second error follows the age permutation.
The model, records and outcomes stay fixed. Only the age values move between rows.
Building the ranking…
Molnar · the same experiment
Extra error in the metric’s own units.
0 means no change.
Error as a multiple of the original.
1 means no change.
Above 1: more error. Below 1: less error.
Difference: the extra Brier error. Zero means no change. The ranking stays the same because every feature shares the same original error.
The whiskers show variation across column shuffles, conditional on this fitted model and evaluation set. They are not population confidence intervals. The Brier score averages squared differences between predicted death probabilities and observed outcomes.
Cross-slide aggregate evidence ↗
This is permutation importance for predictive performance. It does not establish a causal effect of a variable on death. Ratios below 1 are possible when shuffling reduces the error; 1 is a reference, not a lower bound. The top-ranked features shown here all increase mean error.
Technical sheet ↗ · Aggregate evidence · PFI trace ↗ · Molnar: difference and ratio ↗
The motion replays the recorded Python results. Each full-test shuffle uses the same fitted model, outcomes and evaluation rows. The compact table is a symbolic mechanism diagram, without patient values.