Causal Attribution

The formal mathematical framework underpinning causal attribution. An SCM defines a directed acyclic graph (DAG) representing causal relationships between variables, where each node is a function of its direct causes and an independent noise term.

• Nodes: Represent observed variables (e.g., pixels, clinical biomarkers)
• Edges: Represent direct causal influence, not just correlation
• Interventions: The do() operator simulates setting a variable to a fixed value, severing its incoming edges
• Counterfactuals: Answers "what would the prediction have been if feature X had a different value?"

In medical imaging, an SCM might model how tumor size causally influences texture features, which in turn influence the classifier, distinguishing this causal chain from spurious correlations with scanner type.

Standard feature attribution methods like Integrated Gradients or SHAP are observational—they analyze the model using only the existing data distribution. Causal attribution is interventional—it actively perturbs or intervenes on input features to observe the effect on the output.

• Observational: "How much did this pixel contribute to the prediction?" (correlational)
• Interventional: "If I change this pixel while holding all causal parents constant, how does the prediction change?" (causal)

This distinction is critical in medical imaging. An observational method might highlight a radiologist's markup pen as important; a causal method would correctly identify that the pen mark is correlated with but not a cause of the pathology, preventing the model from relying on confounding artifacts.

A key metric in causal attribution that quantifies the expected change in a model's output when an input feature is intervened upon. ACE is computed by:

1. Selecting a feature (e.g., a specific region in a chest X-ray)
2. Applying an intervention—setting the feature to a baseline value (e.g., zero, blurred, or a counterfactual value)
3. Measuring the difference in the model's output probability before and after intervention
4. Averaging this difference over a population of images

Unlike gradient-based saliency, ACE directly measures the causal effect of removing or altering information. A high ACE for a lung nodule region confirms that the model's "malignant" classification is causally dependent on that region, not on background pixels.

A confounder is a variable that influences both the input feature and the model's output, creating a spurious statistical association that is not causal. In medical imaging, the scanner manufacturer is a classic confounder—it affects image texture (input) and may correlate with patient demographics (output).

Causal attribution methods address confounding through:

• Backdoor adjustment: Blocking non-causal paths by conditioning on confounders
• Front-door adjustment: Using mediators when confounders are unobserved
• Instrumental variables: Leveraging variables that affect the input but not the output directly

Deconfounding ensures that attribution maps reflect genuine anatomical pathology rather than site-specific imaging protocols, making models robust across hospitals.

A rigorous technique for testing whether a proposed causal attribution is faithful to the model's true reasoning. The process:

• Hypothesize a causal graph that describes which input features the model uses
• Scrub (randomize) all features that are not causally connected to the output according to the hypothesis
• Measure if the model's behavior remains unchanged
• If the model's output is preserved after scrubbing supposedly irrelevant features, the causal hypothesis is validated

In practice, causal scrubbing can reveal that a pneumonia classifier is relying on the "portable" X-ray marker rather than lung opacities. When the marker region is scrubbed and performance drops, the spurious causal pathway is exposed.

An emerging paradigm that trains models to learn disentangled, causally structured representations of data, rather than relying on post-hoc attribution. The model's latent space is explicitly factorized into independent causal mechanisms.

• Invariant Risk Minimization (IRM): Trains models to rely only on features whose relationship to the label is stable across environments
• Causal Disentanglement: Separates latent factors into those that cause the output and those that are merely correlated
• Benefits: Inherently interpretable, robust to distribution shift, and resistant to confounding

For medical imaging, a causally disentangled model would automatically separate anatomical structure from imaging protocol artifacts, making its decisions transparent by design rather than requiring external explanation tools.

The general class of methods that assign a relevance or importance score to each input feature of a model. While standard feature attribution identifies correlated pixels, causal attribution seeks to identify the true causes of a prediction.

• Quantifies contribution to a specific output
• Foundation for saliency maps and heatmaps
• Does not inherently distinguish correlation from causation

An explanation that describes the minimal change to an input instance's features that would alter the model's prediction to a predefined, alternative outcome. This directly answers 'what if' questions central to causal reasoning.

• Example: 'If this tissue density were lower, the classification would be benign'
• Closely aligned with interventional causal logic
• Critical for clinical decision support

A quantitative metric that evaluates the accuracy of an explanation by measuring how well the attributed importance scores correlate with the actual change in model output when the corresponding features are perturbed or removed.

• Validates if an explanation reflects true model behavior
• Essential for comparing causal vs. correlational methods
• Low faithfulness indicates an interpretability illusion

An attribution method that assigns importance scores by integrating the gradients of the model's output along a straight-line path from a baseline to the actual input. It satisfies the completeness axiom, ensuring the sum of attributions equals the prediction difference.

• A foundational, non-causal attribution technique
• Often used as a baseline for evaluating causal methods
• Implemented in the Captum library

A scientific experiment to understand a model's behavior by systematically removing or disabling specific components—such as layers, neurons, or input features—and measuring the resulting impact on performance.

• A direct, interventional approach to probing causality
• Can be applied to both model architecture and input space
• Computationally intensive but highly interpretable

The specific requirements for model transparency mandated by health authorities like the FDA or under regulations like the EU MDR. Causal attribution methods are increasingly favored as they provide a higher standard of evidence for auditability.

• Moves beyond 'where' a model looked to 'why' it decided
• Supports the creation of a robust SaMD Audit Trail
• Essential for clinical validation and post-market surveillance