SHAP

SHAP values are derived from Shapley values, a concept from cooperative game theory introduced by Lloyd Shapley in 1953. The core idea is to fairly distribute the 'payout' (the prediction) among the 'players' (the input features) based on their marginal contribution across all possible coalitions of features. In a pharmacovigilance context, this means quantifying exactly how much a specific word like 'anaphylaxis' contributed to classifying a clinical note as a serious adverse event, compared to a baseline prediction. The method satisfies four axioms: efficiency (contributions sum to the prediction difference), symmetry (identical contributions get identical values), dummy (zero-contribution features get zero), and additivity (values can be summed across models).

SHAP belongs to the class of additive feature attribution methods, meaning it explains a model's output as a linear sum of individual feature effects. The explanation model g is expressed as: g(z') = φ₀ + Σ φᵢz'ᵢ, where φ₀ is the base value (average model output), φᵢ is the SHAP value for feature i, and z' indicates feature presence. This linear decomposition makes SHAP inherently human-interpretable, even when explaining highly non-linear models like gradient-boosted trees or deep neural networks used for adverse event classification. The additive property ensures that the sum of all SHAP values plus the base value equals the model's exact prediction for that instance.

SHAP is fully model-agnostic, meaning it can explain the output of any machine learning model without requiring access to internal weights or gradients. This is critical in pharmacovigilance workflows where heterogeneous model types are deployed: a logistic regression for initial screening, a random forest for signal prioritization, and a transformer-based NLP model for adverse event extraction from clinical notes. SHAP provides a consistent explanation framework across all of them, enabling direct comparison of feature importance. This contrasts with model-specific methods like integrated gradients (deep learning only) or Gini importance (tree-based models only).

KernelSHAP is the model-agnostic implementation that estimates Shapley values using a specially weighted linear regression. It works by sampling feature coalitions, evaluating the model with features either present or masked, and fitting a local explanation model. The key innovation is the Shapley kernel weighting function: π(z') = (M-1) / (M choose |z'|) * |z'| * (M - |z'|), which gives higher weight to coalitions that are either very small or very large, as these provide the most information about individual feature contributions. For pharmacovigilance models processing high-dimensional clinical text, KernelSHAP can be computationally expensive but provides unbiased estimates.

TreeSHAP is a specialized algorithm for tree-based models (XGBoost, LightGBM, random forests) that computes exact SHAP values in polynomial time instead of exponential. It exploits the internal structure of decision trees by tracking the proportion of training samples flowing through each node. This reduces computation from O(TL2^M) to O(TLD^2), where T is the number of trees, L is the maximum number of leaves, and D is the maximum tree depth. For a pharmacovigilance signal detection model using an ensemble of 500 trees, TreeSHAP can compute exact SHAP values for thousands of drug-event combinations in seconds, making it practical for real-time explainability dashboards.

SHAP operates at two complementary levels of explanation. Local explanations explain a single prediction—for instance, why a specific clinical note triggered a high adverse event score, highlighting that 'Stevens-Johnson syndrome' contributed +0.45 while 'headache' contributed only +0.02. Global explanations aggregate SHAP values across many predictions to reveal overall model behavior, such as identifying that seriousness criteria terms are consistently the most influential features across all ICSR classifications. This dual capability is essential for both auditing individual pharmacovigilance decisions and validating that the model has learned clinically meaningful patterns rather than spurious correlations.

SHAP (SHapley Additive exPlanations) is a game-theoretic method that assigns each input feature an importance value for a specific prediction. It computes the marginal contribution of a feature by comparing model outputs with and without that feature, averaged over all possible feature subsets. The result is an additive feature attribution where the sum of all SHAP values equals the difference between the model's prediction and the average prediction. This satisfies three desirable properties: local accuracy, missingness (missing features get zero attribution), and consistency (if a model changes so a feature contributes more, its SHAP value never decreases).

For a model classifying whether a clinical note contains an adverse event mention, SHAP values reveal exactly which tokens drove the decision. Key interpretability outputs include:

• Token-level attribution: Words like 'anaphylaxis' or 'rash' receive high positive SHAP values for the AE class
• Negation handling: The phrase 'no evidence of' assigns negative SHAP values, pushing prediction away from AE
• Contextual modifiers: Terms like 'history of' or 'resolved' receive intermediate values, indicating temporal context This granularity allows safety officers to validate that the model relies on clinically relevant signals rather than spurious correlations.

SHAP operates at two complementary levels:

• Local explanations: Explain a single prediction. For a specific ICSR, SHAP shows which sentences or terms most influenced the causality assessment score
• Global explanations: Aggregate local SHAP values across thousands of predictions to identify which features the model consistently considers important. A SHAP summary plot ranks features by mean absolute SHAP value, revealing that 'drug name' and 'temporal relationship' dominate the model's reasoning This dual capability supports both individual case review and systemic model auditing for regulatory compliance.

Compared to alternative explainability techniques:

• LIME: Perturbs inputs locally but lacks the theoretical guarantees of Shapley values; SHAP provides consistent, axiomatically justified attributions
• Integrated Gradients: Requires a baseline reference and is specific to neural networks; SHAP is model-agnostic
• Attention weights: Indicate where a transformer looks, not what contributes to the output; SHAP quantifies actual impact on prediction For pharmacovigilance, SHAP's theoretical foundation makes it the preferred method when explanations must withstand regulatory scrutiny.

Exact SHAP computation requires evaluating all 2^N feature subsets, which is intractable for high-dimensional clinical text. Practical implementations use:

• KernelSHAP: Model-agnostic approximation using weighted linear regression on sampled coalitions
• TreeSHAP: Exact, polynomial-time algorithm for tree-based models like XGBoost, commonly used in disproportionality analysis
• DeepSHAP: Combines SHAP with DeepLIFT for efficient approximation on deep neural networks For transformer-based AE classifiers processing 512-token sequences, Partition SHAP groups tokens hierarchically to produce interpretable feature-level explanations without combinatorial explosion.

ICH E2E and E2C(R2) guidelines emphasize the need for transparent, auditable pharmacovigilance processes. SHAP directly supports:

• Signal validation: Demonstrating that a detected drug-event association stems from genuine clinical features rather than reporting artifacts
• Model documentation: Providing quantitative evidence of feature importance for regulatory submissions describing AI-assisted case processing
• Bias detection: Identifying if protected attributes like patient demographics inadvertently influence predictions, supporting fairness audits By decomposing predictions into additive feature contributions, SHAP transforms black-box classifiers into auditable decision-support tools suitable for GxP-validated environments.