Report Routing Engine

The engine parses structured identifiers—such as ordering provider NPI, facility code, and document type—to determine the final destination. It does not rely on manual folder sorting. Instead, it uses rule-based logic and probabilistic matching against an Enterprise Master Patient Index (EMPI) to resolve ambiguous recipient identities.

• Extracts HL7 v2 PID segments for patient context
• Maps CDA header metadata to routing tables
• Resolves provider identity via NPI registry lookups

When a document fails automatic classification or the recipient cannot be uniquely identified, the engine places it into a prioritized exception queue. This worklist is sorted by clinical urgency and document age, presenting human reviewers with the AI's top-k predictions for rapid validation.

• Flags confidence scores below a configurable threshold
• Surfaces near-duplicate matches for manual merging
• Tracks mean time to resolution (MTTR) for SLA compliance

The engine integrates with critical results alerting protocols to bypass standard routing when life-threatening findings are detected. If a radiology report contains an impression indicative of an acute condition, the system triggers an immediate, multi-channel notification to the responsible clinician.

• Integrates with SMTP, SMS, and pager gateways
• Requires acknowledgment receipts for closed-loop communication
• Logs all notification attempts in an immutable audit trail

Before routing, the engine performs hash-based deduplication and semantic similarity checks to prevent redundant documents from entering the patient record. Exact binary matches are blocked, while near-duplicates—such as an amended report—are linked to the original via document lifecycle state management.

• Generates SHA-256 fingerprints for exact matching
• Uses MinHash and LSH for fuzzy near-duplicate detection
• Applies addendum processing logic to link versions

The engine transforms classified documents into the target system's required format at the point of delivery. It can convert a Clinical Document Architecture (CDA) file into a FHIR DocumentReference resource, or embed extracted metadata into an HL7 MDM message for legacy EHR consumption.

• Supports XDS.b repository submission
• Generates DICOM Structured Reports for PACS integration
• Validates output against FHIR R4 profiles

Every routing decision is recorded in a tamper-proof audit log that captures the document identifier, the classified type, the resolved recipient, the routing timestamp, and any human overrides. This provides a complete chain of custody for HIPAA disclosure accounting and internal compliance reviews.

• Records user ID for all manual interventions
• Tracks document lifecycle state transitions
• Supports immutable append-only storage backends

A formal, hierarchical classification system defining the semantic categories of clinical documents. The routing engine uses this ontology as its decision map, distinguishing between a discharge summary, an operative note, and a pathology report to determine the correct destination. Without a robust ontology, routing logic collapses into ambiguous keyword matching.

• Maps LOINC document codes to internal categories
• Defines parent-child relationships (e.g., 'Cardiac Imaging' is a child of 'Radiology Report')
• Enables rule-based routing policies per document class

A machine learning algorithm trained to automatically assign predefined category labels to unstructured clinical text. This model serves as the intelligent gatekeeper for the routing engine, ingesting raw documents and outputting a predicted document type with a confidence score. Modern implementations often use transformer-based architectures fine-tuned on clinical corpora.

• Input: Raw text or OCR output
• Output: Document type label + confidence probability
• Common architectures: ClinicalBERT, BioBERT, or fine-tuned Llama variants

A filtering mechanism that routes AI predictions with low probability scores to a manual review queue. The routing engine applies dual thresholds: high-confidence classifications proceed directly to distribution, while borderline predictions are diverted to an exception queue for human validation. This ensures that a misclassified radiology report never silently reaches the wrong specialist.

• Typical auto-routing threshold: >0.95 confidence
• Typical review threshold: 0.70–0.95 confidence
• Below 0.70: Rejected for full manual triage

A FHIR resource used to index and reference a clinical document, including its location, type, and metadata, without embedding the full content. The routing engine generates or updates a DocumentReference as its final output, registering the document's availability in the EHR ecosystem. This resource contains the master identifier, status, type, subject, and attachment URL.

• Contains docStatus (preliminary, final, amended)
• Links to the binary content via content.attachment.url
• Indexed by type using LOINC codes

A dedicated worklist for documents that could not be automatically processed or classified, requiring manual intervention. The routing engine places documents here when encountering unparseable formats, confidence scores below threshold, or metadata conflicts (e.g., mismatched patient identifiers). This queue is the safety net ensuring no document is silently dropped.

• Prioritized by clinical urgency flags
• Tracks time-in-queue for SLA compliance
• Provides audit trail of all manual resolutions

The immutable recording of all system interactions, data modifications, and access events related to a clinical document. For the routing engine, this captures the full provenance chain: who sent the document, how it was classified, which rule determined the destination, and when delivery was confirmed. This is non-negotiable for HIPAA compliance and medicolegal defensibility.

• Logs classification decision with model version
• Records routing rule evaluation path
• Timestamps every state transition (received → classified → routed → delivered)