Exception Queue

The primary mechanism that routes documents to the exception queue. When an AI model's prediction probability falls below a configurable threshold, the document is flagged for review.

• Low-Confidence Routing: Documents with classification scores below the threshold are automatically diverted
• Multi-Label Ambiguity: Cases where multiple labels have similar, low probabilities trigger an exception
• Configurable Bounds: Thresholds are tuned per document type to balance automation rate against accuracy
• Example: A pathology report with a 45% confidence score for 'Surgical Pathology' and 42% for 'Cytology' would be routed to the queue for manual adjudication

Every document in the exception queue is tagged with a specific, machine-readable error code that categorizes the failure mode. This enables systematic root cause analysis and workflow optimization.

• PARSE_ERROR: Corrupted file, unsupported encoding, or malformed structure
• CLASSIFICATION_LOW_CONFIDENCE: No label exceeded the confidence threshold
• EXTRACTION_FAILURE: Document classified successfully but key entities could not be extracted
• DUPLICATE_DETECTED: Document flagged as a potential duplicate requiring human verification
• PATIENT_MISMATCH: Patient identifiers on the document do not match the expected record context

Not all exceptions are equal. A mature exception queue implements dynamic prioritization based on clinical urgency and service level agreements to prevent critical documents from languishing.

• Critical Findings Escalation: Documents containing keywords indicating life-threatening conditions are pushed to the top
• SLA-Based Aging: Documents approaching breach of turnaround time targets are visually flagged and escalated
• Document Type Priority: Stat results, ED reports, and stroke alerts receive higher base priority than routine outpatient notes
• Queue Segmentation: Logical sub-queues separate work by error type, specialty, or facility for specialized review teams

Every interaction with an exception queue item is immutably logged to maintain a complete chain of custody. This is essential for compliance, operational analytics, and resolving disputes.

• Status Lifecycle: Documents transition through states: UNREVIEWED → IN_REVIEW → RESOLVED or ESCALATED
• Reviewer Attribution: Every action is stamped with the user ID, timestamp, and resolution action taken
• Idempotency Guarantees: The system prevents the same document from being processed twice or by two reviewers simultaneously
• Replay Capability: The full history of a document's journey through the queue can be reconstructed for audits

The exception queue is not a dead end. Resolved documents must feed back into the system to improve the underlying models and prevent recurring exceptions.

• Human-Corrected Labels: The reviewer's manual classification becomes a new ground-truth label for future model fine-tuning
• Closed-Loop Learning: Resolved exceptions are periodically aggregated and used to retrain or fine-tune the classification model via Supervised Fine-Tuning (SFT)
• Rule Updates: Recurring pattern-based failures trigger updates to deterministic pre-processing rules or Regular Expression Parsing logic
• Dashboard Metrics: Exception rates by error type, document category, and facility are tracked to measure system health over time

A resolved exception must seamlessly re-enter the automated workflow as if it had never been flagged. The queue must integrate cleanly with the Report Routing Engine and other downstream consumers.

• API-Driven Resolution: External review interfaces interact with the queue via a RESTful API to claim, update, and resolve items
• Event-Driven Notification: Upon resolution, an event is emitted to trigger downstream processes such as FHIR Resource Mapping or Critical Results Notification
• Metadata Preservation: All original document metadata and the resolved classification are passed forward without data loss
• HL7 MDM Integration: For clinical document management, resolution actions can be communicated via HL7 Master Document Management messages

The primary gatekeeper that determines whether a document proceeds automatically or lands in the exception queue. When a classification or extraction model's prediction probability falls below a predefined threshold—typically set between 85% and 99% depending on clinical risk tolerance—the document is routed for manual review.

• High-confidence predictions bypass the queue entirely
• Low-confidence predictions trigger exception routing
• Thresholds are often tuned per document type (radiology vs. pathology)

The manual intervention process that resolves items sitting in the exception queue. Clinical reviewers audit AI-generated classifications, correct mislabeled document types, and validate extracted entities before releasing documents back into the automated workflow.

• Reviewers see the original document alongside the AI's prediction
• Corrections feed back into model retraining pipelines
• Turnaround time SLAs typically govern queue resolution speed

Every document that enters and exits the exception queue generates an immutable audit record capturing who reviewed it, what was changed, and when resolution occurred. This is critical for HIPAA compliance and operational analytics.

• Tracks time-in-queue metrics for SLA reporting
• Records the original AI prediction and the final human determination
• Provides data lineage for model improvement feedback loops

Documents in the exception queue are assigned a distinct lifecycle status—typically 'pending review' or 'exception held'—that prevents them from progressing to downstream systems like FHIR resource mapping or clinical decision support until resolution.

• Prevents incomplete or unverified data from entering the legal medical record
• States transition automatically upon reviewer sign-off
• Integrates with amendment and addendum processing workflows

A common trigger for exception queue routing occurs when the system cannot determine whether an incoming document is a true duplicate of an existing record. Near-duplicate documents—such as amended reports or corrected versions—require human judgment to resolve.

• Hash-based deduplication catches exact copies automatically
• Semantic similarity scoring flags near-duplicates for review
• False duplicate flags can delay critical results if not resolved quickly

When a document type doesn't match any category in the trained document type ontology, zero-shot classifiers attempt to assign a label using semantic similarity. If confidence remains low, the document enters the exception queue as an 'unknown type' for manual categorization.

• Handles novel or rare document formats
• Reviewer-assigned labels can be added to the ontology for future automation
• Prevents silent misclassification of unfamiliar clinical documents