Criteria-to-Query Translation

The parsed, structured eligibility criteria are compiled into an Abstract Syntax Tree (AST)—a hierarchical, intermediate representation that models the logical relationships between clinical facts. This AST explicitly captures AND/OR conjunctions, negation scopes, and temporal precedence before any database-specific code is generated. The tree structure allows for logical optimization, such as predicate pushdown, where highly selective filters are evaluated first to minimize the data scanned in subsequent operations.

A direct one-to-one mapping of criteria to codes is insufficient. The translation engine must perform ontology-grounded semantic expansion using terminologies like SNOMED CT and RxNorm. For a criterion like 'history of myocardial infarction,' the system expands the query to include all child concepts—subendocardial infarction, STEMI, NSTEMI—and synonymous terms via the ontology's transitive closure. This ensures high recall by accounting for the lexical variability in how clinicians document the same condition.

The system synthesizes the optimized AST into multiple target query dialects simultaneously to support a heterogeneous data fabric. A single eligibility rule set generates:

• SQL for relational clinical data warehouses
• FHIR API calls with chained parameters for modern EHRs
• SPARQL for RDF-based knowledge graphs
• Elasticsearch DSL for free-text narrative search This multi-dialect capability allows a single trial protocol to screen patients across legacy and modern systems without manual query rewriting.

Complex time-window criteria like 'disease progression within 6 months after platinum-based chemotherapy' are compiled into temporal logic expressions with explicit interval algebra. The translator generates queries that perform clinical event sequencing—joining a patient's timestamped diagnosis and medication administration records to validate the chronological order and the duration constraint. This involves generating SQL window functions or FHIR reverse chaining to ensure the sequence of events matches the protocol's temporal logic.

Quantitative criteria require value set binding to standard terminologies and unit normalization to ensure comparability. A criterion like 'serum creatinine > 1.5 mg/dL' must be translated to query for the correct LOINC code (e.g., 2160-0) and dynamically convert stored values in SI units (μmol/L) to the required conventional units. The translation layer embeds a unit conversion engine that normalizes all numeric lab results at query time to match the protocol's specified units.

Before execution, the generated query undergoes a cost-based optimization specific to patient screening workloads. The optimizer reorders join operations and filter application based on the cardinality of clinical concepts. A highly selective genomic biomarker filter is executed before a broad age-range filter. The system also generates incremental screening queries that cache intermediate patient cohorts, allowing for rapid re-screening when a single criterion in a protocol amendment changes, rather than re-executing the entire query from scratch.

A machine-processable definition of a clinical condition expressed as a set of logical expressions and data queries. Computable phenotypes are the direct output of criteria-to-query translation, enabling the identification of patient cohorts from electronic health records.

• Combines ICD-10-CM codes, lab values, and medication orders
• Uses logical operators (AND, OR, NOT) to define inclusion boundaries
• Must handle temporal constraints like 'diagnosis within 6 months'

The automated extraction and structuring of complex free-text inclusion and exclusion requirements from clinical trial protocols into a machine-readable format. This upstream process feeds the criteria-to-query pipeline with structured logical predicates.

• Transforms narrative text like 'EGFR-mutant non-small cell lung cancer' into structured fields
• Resolves anaphora and co-reference in multi-sentence criteria
• Outputs JSON or OMOP-compatible structured criteria

A software component that runs a computable phenotype definition against a clinical data repository, resolving logical expressions and temporal constraints to return a patient cohort. This is the runtime environment where translated queries execute.

• Translates phenotype logic into SQL, FHIR, or OMOP CDM queries
• Optimizes query plans for large-scale patient databases
• Returns both cohort lists and individual patient-level eligibility evidence

The AI capability to interpret and validate time-dependent clinical constraints against a patient's longitudinal record. Criteria-to-query translation must encode complex temporal logic such as washout periods, disease progression timelines, and sequence dependencies.

• Handles constraints like 'no chemotherapy within 28 days of enrollment'
• Requires patient timeline reconstruction from timestamped events
• Uses Allen's interval algebra for temporal relationship validation

The process of mapping synonymous clinical terms and varying units of measure within trial criteria to a standard ontology before query generation. Normalization ensures that a criterion like 'HbA1c > 7%' correctly maps to the local lab codes and units in the target database.

• Maps to standard terminologies: SNOMED CT, LOINC, RxNorm
• Handles unit conversion (mg/dL vs mmol/L)
• Resolves semantic equivalence ('heart attack' = 'myocardial infarction')

A standardized, machine-readable syntax for expressing the complex inclusion and exclusion rules that define a patient group. CDL serves as the intermediate representation between parsed criteria and executable queries, enabling portability across different data models.

• Examples include OHDSI's ATLAS and CQL (Clinical Quality Language)
• Provides version control for phenotype definitions
• Enables sharing of computable definitions across institutions