AIF-C01 Practice Question: A healthcare organization uses an ML model to…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Partial dependence plots

  Why it's wrong here

  Partial dependence plots show global, average feature effects across the dataset, not per-patient explanations, so they cannot justify an individual readmission prediction to a clinician. They are tempting because they visualise how a feature influences predictions overall, which suits model-level interpretation rather than the local, case-specific reasoning this scenario demands.

• ✗

  Amazon SageMaker Autopilot

  Why it's wrong here

  SageMaker Autopilot automates model selection and hyperparameter tuning; it produces no per-prediction attribution for clinicians. It is tempting because it is a managed SageMaker capability that surfaces feature-importance reports during training, and it would be correct when the goal is automated model building rather than explaining individual outputs.

• ✗

  Global feature importance from a random forest

  Why it's wrong here

  Global feature importance aggregates one ranking across the whole training set, so it cannot attribute any single patient's readmission score. It is tempting because it is model-agnostic and cheap to compute; it is correct when reporting overall model behaviour rather than per-patient justification.

• ✓

  LIME (Local Interpretable Model-agnostic Explanations)

  Why this is correct

  LIME fits because it perturbs individual instances and fits a local surrogate model, producing explanations for single predictions without needing access to model internals. This model-agnostic, per-patient approach satisfies the regulatory need to explain individual readmission predictions efficiently to clinicians.