Unstructured Criteria Extraction

Transforms narrative eligibility clauses into structured, machine-readable logical forms. This process goes beyond keyword spotting to understand the intent of a criterion.

• Identifies the clinical concept (e.g., 'HbA1c'), the operator (e.g., 'greater than'), and the value (e.g., '7.0%').
• Resolves complex linguistic structures like anaphora ('it', 'the patient') to link conditions to the correct entity.
• Example: Parsing 'Patient must not have had a myocardial infarction within the last 90 days' into a structured object with a negated temporal constraint.

Critically distinguishes between affirmed, negated, and uncertain clinical conditions to prevent catastrophic screening errors. A simple keyword match for 'cancer' is insufficient.

• Uses contextual embeddings to analyze the linguistic scope of negation triggers like 'no evidence of' or 'ruled out'.
• Identifies hedging language such as 'suspected,' 'possible,' or 'cannot be ruled out' to flag uncertain findings.
• Prevents false exclusions: Correctly interpreting 'history of breast cancer' as negated in 'no history of breast cancer' is essential for accurate eligibility scoring.

Extracts and standardizes time-based constraints from criteria into a computable format for comparison against a patient's longitudinal record.

• Recognizes and normalizes relative expressions like 'within the last 6 weeks' or 'prior to enrollment' into absolute date ranges.
• Maps clinical events to a standardized timeline to validate constraints such as 'disease progression after platinum-based chemotherapy'.
• Handles complex durations: Parsing 'a washout period of 5 half-lives' requires linking the concept to a specific drug's pharmacokinetic data.

Maps extracted clinical terms to standard medical ontologies like SNOMED CT, LOINC, and RxNorm to resolve semantic ambiguity.

• Normalizes synonymous terms: 'High blood pressure' and 'Hypertension' are mapped to the same SNOMED code.
• Links lab tests to LOINC codes to ensure the correct analyte is being evaluated, regardless of local naming conventions.
• Enables semantic interoperability between trial criteria and diverse EHR systems, which is the foundational step for accurate criteria-to-query translation.

Breaks down a complex, multi-part eligibility criterion into its indivisible, independently evaluable logical components.

• A single sentence like 'Patient must have EGFR-mutant NSCLC and no prior TKI therapy' is decomposed into two atomic criteria.
• Each atomic unit contains a single clinical fact (e.g., diagnosis = NSCLC), a single operator, and a single value.
• This logical atomization is a prerequisite for execution by a deterministic rule engine and for assigning granular weights to each condition.

Isolates and structures lists of prohibited and permitted medications directly from the protocol narrative.

• Identifies specific drug names, classes, and routes of administration mentioned in exclusion criteria.
• Normalizes drug terms to RxNorm to enable cross-referencing with a patient's structured medication list.
• Handles complex rules: Parsing 'concurrent use of strong CYP3A4 inhibitors' requires inferring a class of drugs rather than matching a single named entity.

The process of converting parsed, structured eligibility criteria into executable database queries, such as SQL or FHIR API calls, to screen patient repositories. This bridges the gap between the extracted unstructured criteria and the operational data layer.

• Translates logical rules to query languages
• Maps criteria to specific data fields (e.g., LOINC codes)
• Enables direct execution against EHR databases

The AI capability to interpret and validate time-dependent clinical constraints against a patient's longitudinal record. Extracting a criterion like 'no myocardial infarction within 6 months' requires not just entity recognition but also sophisticated temporal reasoning.

• Resolves relative time expressions (e.g., 'within 3 months')
• Validates event sequences against timelines
• Handles duration-based constraints (e.g., 'treatment for 2 years')

The process of mapping and harmonizing disparate medical terminologies such as SNOMED CT, ICD-10-CM, LOINC, and RxNorm. Unstructured criteria extraction relies on this alignment to normalize the varied language used in protocols to a standard set of concepts.

• Resolves synonyms (e.g., 'high blood pressure' to HTN)
• Maps brand names to generic drug ingredients
• Enables semantic interoperability between systems

The NLP task of distinguishing between affirmed, negated, and uncertain clinical findings in narrative text. A criterion stating 'patients without active infection' requires accurate negation detection to avoid incorrectly excluding a patient with a historical note of infection.

• Identifies negation cues (e.g., 'no evidence of')
• Detects uncertainty (e.g., 'suspected for')
• Prevents false-positive exclusions in screening

A machine-processable definition of a clinical condition, expressed as a set of logical expressions and data queries, used to identify patient cohorts. Unstructured criteria extraction is the mechanism that feeds these definitions by converting narrative protocol text into the logical rules that form a computable phenotype.

• Combines structured and unstructured data rules
• Defines a cohort for screening or research
• The executable output of criteria extraction