Hybrid Matching Architecture

The first-pass filter that applies hard constraints with absolute certainty. This layer handles binary inclusion/exclusion criteria such as:

• Age ranges and sex assignments
• Lab value thresholds (e.g., platelet count > 100,000/μL)
• Concomitant medication exclusions
• Prior therapy washout periods

Rules are compiled from structured eligibility criteria parsing output and executed against normalized patient data. This layer guarantees 100% precision on explicitly defined criteria and eliminates patients who definitively do not qualify before more expensive semantic processing occurs.

The second-pass layer that handles ambiguous or narrative criteria where exact matching fails. This component:

• Converts patient records and trial criteria into dense vector embeddings using domain-adapted language models
• Computes cosine similarity between patient profiles and eligibility requirements
• Identifies patients with conditions that are clinically similar but not identically phrased (e.g., 'renal insufficiency' vs. 'CKD stage 3')

This layer dramatically improves recall by surfacing candidates that a pure rule-based system would miss due to terminology gaps or implicit clinical relationships.

A unified scoring layer that combines outputs from both matching passes into a single, interpretable eligibility score per patient-trial pair. The mechanism:

• Assigns binary pass/fail weights to deterministic rule results
• Converts semantic similarity scores into probabilistic confidence values
• Applies criteria weighting where critical inclusion criteria carry higher importance
• Produces a ranked candidate list with explainable score breakdowns

Thresholds are configurable per trial, allowing study teams to balance precision vs. recall based on recruitment urgency and screen failure tolerance.

A specialized sub-engine that validates time-dependent criteria against longitudinal patient records. This module:

• Reconstructs patient timelines from timestamped clinical events
• Evaluates sequence constraints (e.g., 'progression after platinum-based chemotherapy')
• Validates washout windows and recency requirements
• Handles relative time expressions like 'within 6 months of screening'

Integrates with both deterministic rules for exact date math and semantic matching for fuzzy temporal expressions found in unstructured criteria text.

A preprocessing pipeline that harmonizes disparate clinical terminologies before matching occurs. This layer:

• Maps local lab codes to LOINC standards
• Normalizes drug names to RxNorm concepts
• Aligns diagnoses across ICD-10-CM and SNOMED CT
• Resolves unit-of-measure discrepancies (e.g., mg/dL vs. mmol/L)

By normalizing both patient data and parsed trial criteria to common ontologies, the architecture eliminates false negatives caused by coding system mismatches and enables cross-institutional screening across heterogeneous EHR systems.

An automated monitoring system that detects and integrates protocol amendments into the active screening logic without manual reconfiguration. Capabilities include:

• Version comparison of structured eligibility criteria
• Identification of added, removed, or modified constraints
• Automatic recompilation of rule sets and re-indexing of semantic embeddings
• Flagging of previously screened patients for re-evaluation under amended criteria

This ensures screening accuracy persists across the trial lifecycle, preventing the common problem of sites using outdated eligibility logic after protocol revisions.

The rule-based filtering layer that applies hard constraints to eliminate patients who definitively fail binary criteria. This engine evaluates structured data against explicit thresholds—such as lab values, age ranges, or genomic markers—with zero ambiguity.

• Operates on structured, normalized data from FHIR resources or OMOP CDM
• Executes Boolean logic (AND/OR/NOT) against atomic criteria
• Guarantees 100% precision on exclusionary rules
• Handles concomitant medication checking and allergy cross-referencing
• Typically implemented via Drools, Red Hat Decision Manager, or custom SQL-based evaluators

The probabilistic layer that computes similarity between a patient's clinical narrative and trial criteria using dense vector embeddings. This component handles the ambiguity of natural language, matching concepts like 'uncontrolled hypertension' to specific blood pressure patterns in unstructured notes.

• Uses transformer-based encoders (e.g., ClinicalBERT, PubMedBERT) to generate embeddings
• Performs cosine similarity or dot-product scoring between patient and criteria vectors
• Enables fuzzy matching for synonymous conditions and gradable findings
• Powers patient vector embedding for nearest-neighbor search across trial databases
• Typically backed by vector databases like Pinecone, Weaviate, or pgvector

The preprocessing stage that breaks complex free-text eligibility criteria into atomic, machine-evaluable units. A single paragraph like 'Patients must have failed at least 2 prior lines of therapy, with ECOG 0-1, and no active CNS metastases' becomes multiple discrete logical clauses.

• Combines medical NER with syntactic dependency parsing
• Identifies temporal constraints (e.g., 'within 28 days of enrollment')
• Classifies each atomic criterion as inclusion or exclusion
• Routes binary criteria to the rule engine and narrative criteria to the semantic matcher
• Outputs a structured criteria graph for downstream execution

The decision integration module that combines outputs from the deterministic and probabilistic paths into a single eligibility score. This layer applies calibrated weights based on each component's confidence, ensuring that a high semantic match does not override a hard exclusion.

• Implements a weighted scoring function: FinalScore = (w1 × RuleScore) + (w2 × SemanticScore)
• Applies hard override logic for deterministic failures regardless of semantic score
• Uses calibrated confidence thresholds to flag borderline cases for human review
• Generates explainable audit trails showing which criteria contributed to the final decision
• Enables human-in-the-loop review for cases falling in the uncertainty band

The specialized module that validates time-dependent constraints against a patient's longitudinal record. This engine interprets criteria like 'no chemotherapy within 4 weeks of enrollment' by reconstructing the patient's clinical timeline and checking event sequencing.

• Performs clinical event sequencing from timestamped EHR data
• Validates washout periods, disease progression timelines, and recency requirements
• Handles relative time expressions (e.g., 'within 6 months of diagnosis')
• Integrates with patient timeline reconstruction pipelines
• Uses Allen's interval algebra for complex temporal logic resolution

The bridge component that converts structured eligibility criteria into executable database queries against clinical data warehouses. This translator generates FHIR API calls, SQL statements, or OMOP CDM queries to retrieve candidate patient cohorts.

• Maps parsed criteria to standardized terminologies (SNOMED CT, LOINC, RxNorm)
• Generates parameterized queries for EHR and claims databases
• Supports bulk FHIR export and CDS Hooks integration patterns
• Enables site feasibility assessment by executing queries across multiple data sources
• Implements query optimization for large-scale patient repositories