SaMD Audit Trail

The audit trail must be cryptographically secured to prevent post-hoc modification or deletion. Once a clinical AI decision is logged, the record becomes append-only and immutable.

• Uses hash chaining to link sequential records, where each entry contains the hash of the previous entry
• Employs digital signatures to verify the identity of the system component that generated each log entry
• Implements write-once-read-many (WORM) storage policies at the infrastructure level
• Any alteration attempt must generate an immediate alert and leave a forensic trace
• Critical for demonstrating data integrity during FDA inspections and ISO 13485 audits

Every clinical output must be traceable back through the entire inference pipeline to its originating inputs, model version, and explanation artifacts.

• Records the exact input data including DICOM headers, pixel data hashes, and preprocessing parameters
• Logs the model identifier, version, and checksum of the deployed neural network weights
• Captures all intermediate representations including feature vectors and attention maps
• Associates each prediction with its corresponding saliency map or feature attribution output
• Enables retrospective analysis: 'Why did the model flag this region as malignant on this specific study?'

All events across distributed diagnostic system components must be recorded with precise, synchronized timestamps to reconstruct the exact sequence of clinical decision-making.

• Uses Network Time Protocol (NTP) or Precision Time Protocol (PTP) to synchronize clocks across all system nodes
• Timestamps recorded at millisecond granularity to resolve event ordering in high-throughput inference pipelines
• Includes both wall-clock time and monotonic system time to detect clock drift anomalies
• Logs the time of image acquisition, preprocessing completion, inference start/end, and result display
• Essential for correlating audit events with clinical workflows and identifying latency bottlenecks

The audit trail must capture not just the diagnostic output, but the complete explanation artifacts that justify the model's decision to clinicians and regulators.

• Stores Grad-CAM heatmaps, SHAP value matrices, and Integrated Gradients attributions as structured data
• Records the specific explanation method and its version used for each prediction
• Logs counterfactual explanations showing what minimal changes would alter the diagnosis
• Captures uncertainty quantification metrics including epistemic and aleatoric uncertainty scores
• Enables post-market surveillance teams to verify that model decisions were based on clinically relevant anatomical regions, not confounding artifacts

Access to the audit trail itself must be governed by strict role-based controls, with all access events themselves logged to create a meta-audit trail.

• Implements role-based access control (RBAC) distinguishing radiologists, administrators, regulators, and service engineers
• All read access to audit records is logged with user identity, timestamp, and purpose of access
• Supports regulatory read-only modes where inspectors can query but never modify the audit trail
• Generates cryptographic proof of completeness to demonstrate that no records have been deleted
• Aligns with HIPAA audit control requirements and FDA's Postmarket Management of Cybersecurity in Medical Devices guidance

The audit trail infrastructure should include continuous monitoring that automatically flags deviations from expected clinical AI behavior for immediate investigation.

• Detects concept drift by monitoring shifts in feature attribution patterns over time
• Flags out-of-distribution inputs where the model's confidence and explanation patterns indicate unfamiliar data
• Monitors for attribution attacks where adversarial inputs produce correct classifications but misleading explanations
• Tracks explanation faithfulness scores over time to detect degradation in interpretability quality
• Generates automated alerts when saliency maps deviate from domain-specific anatomical priors