HealthData Labs wants to track drift in feature attributions (SHAP values) for a clinical risk model. They have computed training-set SHAP values already and stored them in s3://healthdata-baselines/shap/. Production endpoint captures inputs and model outputs to s3://healthdata-capture/. Which approach will let them use SageMaker monitoring to detect changes in SHAP distributions over time?

The correct approach leverages SageMaker Model Monitoring's explainability monitoring capabilities. It involves establishing a baseline of SHAP values using SageMaker Clarify on the training data. This baseline is then stored in S3. For production, data capture is enabled on the endpoint to collect inference requests and responses. Finally, a Model Explainability MonitoringSchedule is created. This schedule periodically triggers Clarify processing jobs to compute SHAP values for the captured production data and compares their distributions against the established baseline, effectively detecting drift in feature attributions. Incorrect options: SageMaker does not have a built-in SHAP export flag in EndpointConfig that automatically annotates captured requests with SHAP values for Model Monitor. CloudWatch metric alarms on FeatureImportance metrics are not directly related to SHAP values and would not automatically show SHAP changes unless the model container explicitly emits SHAP as a custom metric, which is not a standard SageMaker monitoring feature for explainability. Running daily Athena queries to approximate SHAP and pushing custom CloudWatch metrics is a manual and less integrated approach compared to SageMaker's native explainability monitoring, which automates the SHAP computation and comparison process.