Plan What a Prediction Model Does When an Input Is Missing

Describe the point-of-use input set

List every predictor required when the model is used, its measurement timing and what happens if it is unavailable. Distinguish a test not ordered, a failed measurement, a delayed result and an unknown historical value. These situations can carry different information and should not all be represented by an arbitrary numeric code.

The scikit-learn imputation guide describes several ways to handle missing predictors. None is universally appropriate. The chosen approach must fit the prediction setting and be evaluated as part of the full pipeline, including the information available at the intended prediction time.

Freeze preprocessing with the model

If a model uses learned imputation values, estimate them from training data and save them with the fitted model. During validation, apply the stored transformation rather than recalculating it from the test outcomes or silently replacing it with a new local rule. Place imputation inside resampling loops when evaluating development.

Check whether missingness indicators or model-native missing-value handling are supported and whether their interpretation remains plausible in the new setting. A change in ordering practice can alter the meaning of a missing test even when the underlying clinical risk has not changed.

Audit a hypothetical median replacement

Suppose a fictional training set has observed predictor values 2, 4 and 8, giving a median of 4. A fitted median-imputation step stores 4. A missing value in the later evaluation set is replaced by that stored value under the frozen pipeline.

If the evaluation set happens to have observed values 20 and 30, replacing its missing value by 25 instead would use a different transformation. That may be a deliberately proposed adaptation, but it is no longer the unchanged pipeline being validated. Neither replacement can be justified solely because it allows the software to produce a prediction.

Test operational and statistical consequences

Create synthetic records with each required input missing, multiple missing inputs, impossible values and delayed measurements. Verify whether the system returns a prediction, requests information or declines to score. Ensure those behaviors match the documented intended use and do not expose a raw software error to the user.

Report performance and missingness patterns in the evaluation population, including the fraction of people who cannot receive a score under the planned workflow. Complete-case performance alone may describe a selected subgroup. Preserve preprocessing parameters, missing-value conventions and any adaptation decisions in the model version so future users can reproduce the same prediction rule and recognize when a new validation is needed.

Source verification: 6 November 2026. Examples are hypothetical.