Criteria Decomposition

Each decomposed unit must represent a single, indivisible clinical fact that can be evaluated as a boolean true/false against a patient record. A criterion like 'Patient must have stage III or IV non-small cell lung cancer' should be split into disease type (NSCLC) and staging (III/IV) as separate evaluable components. This prevents partial matches from being incorrectly classified and allows for precise screen failure analysis when a patient fails only one atomic sub-criterion.

The explicit boolean operators (AND, OR, NOT) and temporal relationships (BEFORE, AFTER, WITHIN) connecting atomic criteria must be preserved and structured into a formal logic tree. For example, 'Prior therapy with cisplatin or carboplatin, unless contraindicated' decomposes into:

• Prior therapy with cisplatin OR prior therapy with carboplatin
• UNLESS contraindication to platinum-based therapy exists This preserves the original intent while making the logic executable by a phenotype execution engine.

Every clinical term in a decomposed criterion must be mapped to a standardized concept identifier from a formal ontology like SNOMED CT, RxNorm, or LOINC. The phrase 'heart attack within 6 months' becomes:

• Concept: 22298006 | Myocardial infarction (disorder) |
• Temporal constraint: Occurrence < 6 months before enrollment This grounding enables criteria-to-query translation across heterogeneous EHR systems that may use different local terminologies for the same condition.

Decomposed criteria must explicitly capture all time-window constraints relative to a defined index date (typically enrollment or randomization). A requirement like 'stable corticosteroid dose for 2 weeks prior to enrollment' decomposes into:

• Medication: Corticosteroid (RxNorm class)
• Status: Active prescription
• Temporal constraint: Dose unchanged for >= 14 days BEFORE index date This anchoring is critical for temporal reasoning for eligibility engines that reconstruct patient timelines from longitudinal records.

Quantitative thresholds must be decomposed into a standard unit and a comparator. 'Serum creatinine < 1.5 mg/dL' becomes:

• Lab test: 2160-0 | Creatinine [Mass/volume] in Serum or Plasma | (LOINC)
• Value: 1.5
• Unit: mg/dL
• Comparator: < This normalization allows the eligibility rule engine to convert and compare values regardless of whether the source EHR stores results in mg/dL or µmol/L.

The decomposition must preserve the distinction between affirmed, negated, and uncertain clinical statements from the original protocol text. 'No active autoimmune disease requiring systemic treatment' decomposes into:

• Condition: Autoimmune disease (SNOMED CT)
• Status: NOT present OR inactive
• Qualifier: If present, NOT requiring systemic therapy This prevents negation and uncertainty detection errors where a patient with a resolved or controlled condition is incorrectly excluded from a trial.

The AI capability to interpret and validate time-dependent clinical constraints against a patient's longitudinal record. Decomposed criteria often contain explicit temporal windows that must be independently verified.

• Evaluates constraints like washout periods ('no chemotherapy in prior 4 weeks').
• Validates disease progression timelines ('documented progression after first-line therapy').
• Requires clinical event sequencing to order discrete medical events chronologically before applying logic.

The process of converting parsed, structured eligibility criteria into executable database queries. Each atomic criterion from the decomposition phase maps to a specific query predicate.

• Translates a criterion like 'Platelet count ≥ 100,000/μL' into a SQL WHERE clause or a FHIR API call.
• Handles unit normalization (e.g., converting /mm³ to /μL) to ensure accurate comparison.
• Enables direct, high-speed screening of millions of patient records in a data warehouse.

Distinguishing between affirmed, negated, and uncertain clinical findings in narrative text. This is essential for accurately evaluating a decomposed criterion against unstructured patient notes.

• NegEx and ConText algorithms identify phrases like 'no evidence of metastasis'.
• Prevents a patient from incorrectly matching a criterion requiring 'history of myocardial infarction' based on a note stating 'no history of MI'.
• Ensures the atomic fact extracted for a criterion is contextually true for the patient.

A deterministic software system that evaluates a set of patient facts against a predefined library of clinical trial eligibility rules to produce a pass/fail decision. It executes the logic assembled from decomposed criteria.

• Operates on a Rete algorithm or similar pattern-matching engine for high efficiency.
• Combines the boolean results of all atomic criteria to determine overall trial eligibility.
• Provides a fully auditable decision trail, showing exactly which criteria were met or failed.