Counterfactual Explanations

A valid counterfactual must identify the smallest possible change to input features that alters the prediction. This is typically formulated as an optimization problem minimizing the distance between the original instance and the counterfactual in feature space.

• Uses distance metrics like L1 (Manhattan) or L2 (Euclidean) norm
• Balances sparsity (changing few features) against closeness (small magnitude changes)
• For biomarkers, this might mean identifying a single lab value threshold rather than altering an entire panel

Unlike feature attribution methods that only highlight important variables, counterfactuals provide prescriptive guidance—they tell a clinician or patient exactly what must change to achieve a desired outcome.

• Distinguishes between mutable features (lifestyle, medication dosage) and immutable features (age, genetic markers)
• Ensures recommendations are feasible within clinical constraints
• Example: 'If your HbA1c decreased by 1.2 percentage points, the model would reclassify you as low-risk for diabetic retinopathy'

A single prediction can have multiple valid counterfactuals. Generating a diverse set of explanations gives users options to choose the most feasible path.

• Uses determinantal point processes (DPP) or latent space sampling to enforce diversity
• Each counterfactual represents a distinct intervention strategy
• Critical in medicine where one treatment pathway may be contraindicated for a specific patient

Counterfactuals must respect the data manifold—they cannot suggest feature combinations that are biologically impossible or statistically unrealistic.

• Incorporates density constraints using autoencoders or generative models to reject out-of-distribution counterfactuals
• Prevents suggestions like 'reduce age by 10 years' or impossible lab value combinations
• Maintains clinical credibility for FDA regulatory review

Standard counterfactuals assume features are independent, which is rarely true in biology. Causal counterfactuals use structural causal models (SCMs) to respect known causal relationships.

• Accounts for downstream effects: changing a parent variable propagates changes to its children
• Prevents suggestions that violate established physiological pathways
• Integrates with Mendelian randomization frameworks for target validation

Counterfactual quality is assessed through multiple quantitative dimensions to ensure clinical utility.

• Proximity: Average distance between original and counterfactual instances
• Sparsity: Number of features changed
• Validity: Percentage of counterfactuals that actually flip the prediction
• Plausibility: Distance to nearest training data point
• Diversity: Pairwise distance between generated counterfactuals

Unlike counterfactuals that ask 'what if' about inputs, causal attribution identifies features that are direct causes of an outcome using a structural causal model. This method distinguishes true disease drivers from correlated biomarkers, making it essential for target validation. While counterfactuals show minimal changes to flip a prediction, causal attribution reveals whether those changes would actually alter the biological outcome.

A model-agnostic framework that produces prediction sets with rigorous, finite-sample coverage guarantees. Instead of a single counterfactual, conformal prediction generates a set of plausible outcomes with statistical validity. For diagnostic systems, this means quantifying the uncertainty around a counterfactual recommendation—ensuring clinicians know how reliable the 'what-if' scenario truly is before acting on it.

An axiomatic attribution method that assigns importance scores by integrating gradients along a path from a baseline input to the actual input. This satisfies the completeness property, ensuring attributions sum to the prediction difference. In biomarker identification, Integrated Gradients can highlight which molecular features would need to change—and by how much—to shift a diagnosis, directly informing counterfactual generation.

The ability of a model to maintain accurate predictions when presented with intentionally perturbed inputs. Counterfactual explanations are closely related to adversarial examples—both involve minimal input changes that alter predictions. However, counterfactuals seek actionable, realistic changes for clinical decision support, while adversarial robustness research focuses on defending against malicious perturbations that could compromise diagnostic safety.

Quantitative measures that assess how accurately an explanation reflects the model's true reasoning. Key metrics include:

• Comprehensiveness: Does removing attributed features change the prediction?
• Sufficiency: Do attributed features alone produce the same prediction?
• Monotonicity: Do higher attributions correspond to larger effects? These metrics validate whether counterfactual explanations genuinely capture the model's decision boundary.