Expected Gradients

Unlike Integrated Gradients which relies on a single baseline (often a black image or zero vector), Expected Gradients samples multiple baselines from a background dataset. This averaging process smooths out attribution noise and prevents the explanation from being anchored to an arbitrary reference point. The result is a distributional expectation that captures how features contribute relative to what the model typically sees in training data.

Expected Gradients inherits key axioms from Integrated Gradients while improving practical utility:

• Completeness: Attributions sum to the difference between the prediction and the average baseline prediction
• Sensitivity: If one input differs from baselines and changes the output, it receives non-zero attribution
• Implementation Invariance: Identical models receive identical attributions regardless of network architecture These properties make the method defensible in FDA regulatory submissions where explanation consistency is critical.

In diagnostic vision tasks, single-baseline methods often produce salt-and-pepper noise in saliency maps, obscuring clinically relevant regions. Expected Gradients suppresses this artifact by integrating over diverse reference distributions. For example, when explaining a pneumonia classifier, the method consistently highlights lung opacities rather than scattering attribution across irrelevant anatomical structures, providing radiologists with cleaner, more actionable visual explanations.

The quality of Expected Gradients explanations depends critically on the background dataset from which baselines are drawn. Best practices include:

• Using a representative sample of the training distribution
• Ensuring baselines cover diverse feature values
• Avoiding out-of-distribution references that distort attributions For clinical applications, background sets often comprise healthy control samples, allowing the method to explain what features deviate from normal physiology in a disease prediction.

Expected Gradients requires multiple forward and backward passes per explanation—one for each sampled baseline. This increases computational cost linearly with the number of baselines. Practical implementations often use:

• Batch processing of baseline samples
• Reduced integration steps per baseline (e.g., 50-200 steps)
• Pre-computed background embeddings for efficiency Despite the overhead, the improved explanation fidelity justifies the cost in high-stakes diagnostic settings where interpretability directly impacts clinical trust.

Expected Gradients is closely related to SHAP (SHapley Additive exPlanations). In fact, when using a specific kernel formulation, Expected Gradients approximates SHAP values under the gradient-based feature importance framework. Key differences:

• Expected Gradients operates on continuous integration paths
• SHAP uses coalitional game theory with feature masking
• Both satisfy similar axiomatic guarantees For deep learning models, Expected Gradients often provides faster convergence to stable attributions than permutation-based SHAP estimators.

In diagnostic radiology, Expected Gradients produces cleaner saliency maps by averaging over diverse baseline images. Unlike Integrated Gradients—which can produce noisy attributions from a single black baseline—this method:

• Samples baselines from a background dataset of normal scans
• Suppresses spurious pixel-level noise
• Highlights clinically relevant regions consistently

This is critical for chest X-ray and mammography models where spurious correlations can mislead radiologists.

The FDA's Good Machine Learning Practice (GMLP) emphasizes explainability for SaMD submissions. Expected Gradients supports regulatory approval by:

• Providing axiomatically sound attributions (satisfying sensitivity and implementation invariance)
• Demonstrating that model decisions rely on medically valid features, not artifacts
• Enabling faithfulness metrics to quantify explanation quality

This method aligns with the FDA's expectation that explanations reflect the model's true decision process.

For digital pathology biomarkers, Expected Gradients identifies which tissue regions drive a diagnosis. By sampling baselines from a library of benign slides:

• Attributions consistently highlight mitotic figures in tumor grading
• Reduces sensitivity to staining variation across laboratories
• Enables biomarker saliency for specific morphological features

This improves pathologist trust when AI suggests regions of interest on whole-slide images.

Expected Gradients addresses a key limitation of Integrated Gradients: baseline selection bias. Key differences:

• Integrated Gradients: One baseline (e.g., black image) can produce misleading attributions if the path integral crosses non-linear decision boundaries
• Expected Gradients: Averages over a distribution of baselines, yielding more robust attributions
• Computationally more expensive but produces lower-variance explanations

This trade-off is justified in high-stakes diagnostic contexts where explanation reliability is paramount.

In multi-omics data integration, Expected Gradients identifies which molecular features drive patient stratification. By sampling baselines from healthy control profiles:

• Reveals gene expression signatures consistently important across baseline samples
• Reduces false positive biomarkers from single-reference comparisons
• Integrates with pathway enrichment analysis to validate biological relevance

This approach strengthens the statistical rigor of biomarker discovery pipelines for precision medicine.

Practical deployment of Expected Gradients in diagnostic AI requires:

• Background dataset curation: Baselines must represent the data distribution (e.g., normal patient population)
• Sample size trade-off: More baseline samples improve stability but increase compute cost
• Integration with uncertainty quantification: Pairing attributions with conformal prediction sets provides confidence bounds

For real-time clinical decision support, pre-computed attributions or efficient sampling strategies are recommended.