Biomarker Saliency

Unlike global feature importance, biomarker saliency computes local explanations unique to each patient. For a given tumor biopsy, the model identifies which specific gene expression levels or protein concentrations drove the classification.

• Generates per-instance saliency maps for clinical review
• Enables clinicians to verify model reasoning against established pathology
• Critical for FDA submissions requiring patient-level explainability

Many saliency methods rely on Shapley values from cooperative game theory. Each biomarker is treated as a player in a coalition, with its contribution computed by evaluating model output with and without that feature across all possible subsets.

• Guarantees fair attribution with axiomatic properties
• Satisfies efficiency, symmetry, and additivity constraints
• Computationally intensive; requires approximation methods like KernelSHAP or TreeSHAP

For deep learning diagnostic models, saliency is computed by backpropagating gradients from the output prediction to the input biomarker layer. Integrated Gradients accumulates gradients along a path from a baseline to the actual input.

• Satisfies the sensitivity axiom: zero attribution for features that don't change the prediction
• Avoids saturation problems of simple gradient methods
• Commonly paired with Expected Gradients to reduce noise by averaging over multiple baselines

In multi-modal diagnostic models integrating genomics, proteomics, and imaging, saliency must be computed across heterogeneous feature spaces. Layer-wise Relevance Propagation (LRP) decomposes the prediction backward, assigning relevance scores to each modality.

• Identifies which data modality (e.g., MRI vs. liquid biopsy) dominates a decision
• Enables cross-modal validation of biomarker importance
• Supports regulatory arguments for model robustness when modalities agree

A saliency map is only useful if it faithfully represents the model's true reasoning. Faithfulness metrics measure whether perturbing high-saliency biomarkers actually changes the prediction more than perturbing low-saliency ones.

• Deletion metrics: remove top features and measure prediction degradation
• Insertion metrics: add features back and track recovery
• Sanity checks include randomization tests: saliency should change when model weights are scrambled

The FDA's Predetermined Change Control Plan (PCCP) framework and Good Machine Learning Practice (GMLP) guidelines increasingly expect saliency-based explanations for AI/ML-enabled medical devices.

• Demonstrates model reliance on clinically validated biomarkers, not spurious correlations
• Supports Software as a Medical Device (SaMD) premarket submissions
• Complements Model Cards for transparency documentation

Biomarker saliency generates per-patient explanations that identify exactly which biological measurements drove a specific diagnosis. Unlike population-level feature importance, saliency maps reveal that Patient A's sepsis alert was triggered by procalcitonin and lactate, while Patient B's alert stemmed from creatinine and bilirubin trends. This granularity is essential for FDA submissions requiring demonstration that the model's reasoning aligns with established clinical knowledge for each individual case.

SHAP (SHapley Additive exPlanations) values decompose a diagnostic prediction into the additive contribution of each biomarker. In oncology applications, SHAP analysis can distinguish between competing diagnoses by showing that elevated CA-125 contributed positively to an ovarian cancer prediction while simultaneously contributing negatively to a colorectal cancer prediction. This enables clinicians to understand why the model favored one diagnosis over another, supporting complex differential diagnosis workflows.

When diagnostic models process longitudinal patient data, temporal biomarker saliency reveals when specific measurements became clinically significant. For chronic kidney disease models, saliency analysis might show that eGFR trends dominated predictions in early stages, while potassium and phosphate levels became salient only in later stages. This temporal attribution helps clinicians understand disease trajectory and identify critical intervention windows where specific biomarkers signal impending deterioration.

Counterfactual biomarker saliency answers the question: 'Which biomarkers would need to change, and by how much, to alter the diagnosis?' For a patient classified as high-risk for cardiac events, counterfactual analysis might reveal that reducing LDL cholesterol by 35 mg/dL and systolic blood pressure by 12 mmHg would reclassify them as low-risk. This directly informs personalized treatment targets and provides clinicians with quantifiable intervention goals derived from the model's decision boundary.

Modern diagnostic models fuse genomic, proteomic, and metabolomic data streams. Biomarker saliency methods must attribute importance across these heterogeneous modalities. Integrated Gradients can reveal that a metabolic disorder prediction relied 60% on metabolomic markers (e.g., acylcarnitines), 30% on genomic variants (e.g., ACADM mutations), and 10% on proteomic signals. This cross-modal attribution validates that the model is using biologically coherent evidence rather than spurious correlations.

FDA guidance on Good Machine Learning Practice (GMLP) emphasizes that AI/ML-enabled devices must provide performance evidence across intended patient populations. Biomarker saliency maps serve as explanatory evidence in regulatory submissions by demonstrating that the model's internal logic aligns with established pathophysiology. For a sepsis prediction SaMD, saliency analysis across demographic subgroups proves that the model consistently weights lactate and white blood cell count appropriately regardless of patient age or sex, supporting claims of algorithmic fairness.