Back to index Open roadmap

Model Interpretability and Explainable AI in Healthcare

Performance metrics tell us whether a model predicts well. Interpretability asks a different question: can researchers, clinicians, and patients understand why the model behaves as it does, where it may fail, and whether its reasoning is clinically plausible?

Audience Health researchers using ML models
Level Intermediate to advanced
Focus Global and local explanations
Outcome Safer model reporting and review
3 questions: what, why, and for whom?
2 views: population-level and patient-level
1 rule: explanations are evidence, not proof

Why this page fills a gap

The portal already includes safe splitting, leakage prevention, metrics, calibration, thresholds, bias, and fairness. A missing next step is explainability: after a health ML model looks accurate, students need to inspect which variables drive predictions and whether those drivers make clinical and epidemiological sense.

Need

Build trust carefully

Clinicians often ask for reasons, not only probabilities. Useful explanations make model review easier without promising certainty.

Audit

Find shortcuts

Feature effects can reveal leakage proxies, site effects, or variables that encode healthcare access instead of biology.

Report

Communicate limits

Interpretability outputs should be reported with uncertainty, cohort context, and a warning that association is not causation.

Choose the right explanation view

Click a method to see what it answers, when it is useful, and which healthcare caveat to remember.

Global

Feature importance

Ranks variables by their overall contribution to model performance or predictions.

Global

Partial dependence

Shows how predictions change as one feature varies while other features are averaged out.

Local

SHAP-style values

Decomposes one prediction into feature contributions above or below a baseline risk.

Local

Counterfactuals

Asks what small changes would move a prediction across a decision threshold.

Interactive local explanation demo

Adjust a simplified patient profile. The bars show a local explanation: red features increase predicted risk and green features decrease predicted risk relative to the cohort baseline.

28% Moderate predicted risk
Teaching note: These contributions are illustrative. In a real project, estimate explanations on held-out data and compare them across subgroups, sites, and time periods.

Global vs local explanations

Global explanation

Answers: "What drives the model across the study population?" Useful for model cards, manuscript reporting, and sanity checks.

  • Permutation importance
  • Mean absolute SHAP values
  • Partial dependence or accumulated local effects

Local explanation

Answers: "Why did this specific patient receive this prediction?" Useful for case review and threshold discussions.

  • Patient-level SHAP waterfall
  • Counterfactual examples
  • Nearest similar cases, if clinically appropriate

Clinical validation

Answers: "Does this explanation make sense in the care context?" Useful before deployment or publication.

  • Review with domain experts
  • Check known risk factors and implausible proxies
  • Repeat in external validation cohorts

A safe interpretability workflow

Interpretability is most useful when it is part of the evaluation plan, not a decorative figure added at the end.

1

Define the user

Decide whether the explanation is for researchers, clinicians, patients, auditors, or model developers.

2

Choose the level

Use global explanations for model understanding and local explanations for individual prediction review.

3

Check stability

Compare explanations across folds, bootstrap samples, sites, demographic groups, and time windows.

4

Look for artifacts

Flag proxies for missingness, hospital site, future information, billing behavior, or clinical workflow shortcuts.

5

Document limits

State that explanations describe model behavior. They do not prove causal effects or treatment benefit.

6

Connect to action

Pair explanations with calibration, threshold choice, and a clear clinical response plan.

Common mistakes in health ML explainability

Good practice

Compute explanations on validation or test data that reflects the intended use population.

Be careful

A feature can be important because it is a proxy for care access, site, or measurement frequency.

Avoid

Do not call a feature "causal" just because it has a large model contribution.

Be careful

Correlated features can split or swap importance, making ranked lists unstable.

Good practice

Report explanations with model type, dataset, preprocessing, and uncertainty or sensitivity checks.

Avoid

Do not use interpretability to compensate for poor discrimination, calibration, or external validation.

Key takeaway

Explainable AI is not a single chart. It is a set of questions about model behavior, clinical plausibility, subgroup robustness, and decision consequences. In healthcare, explanations should support responsible review and communication, not replace validation or clinical judgment.