Inclusion Criteria

Criteria must map to controlled terminologies to eliminate the ambiguity of free-text clinical notes.

• SNOMED CT for diagnoses and procedures
• LOINC for laboratory tests and vital signs
• RxNorm for medication orders
• ICD-10-CM for billing and administrative codes Using these standards ensures that a 'type 2 diabetic' is identified identically across different EHR systems.

A robust criterion specifies not just what happened, but when. This defines the index date and lookback windows.

• Incident vs. Prevalent: First-ever diagnosis vs. any history of the condition
• Lookback Period: e.g., 'No HbA1c > 6.5% in the 365 days prior to enrollment'
• Temporal Logic: 'Procedure A must occur after Diagnosis B but before Medication C' Without temporal anchors, cohorts mix new cases with long-standing chronic patients.

Inclusion criteria must be translated into a computable phenotype—a machine-executable algorithm combining multiple data types.

• Rule-Based Logic: Boolean operators (AND, OR, NOT) linking structured codes
• Value-Based Thresholds: Laboratory results (e.g., 'eGFR < 60 mL/min')
• Composite Scores: Calculated indices like BMI or Charlson Comorbidity Index This logic allows a distributed query engine to execute the exact same cohort definition across OMOP-mapped databases.

Explicitly defining the population scope prevents confounding and ensures generalizability.

• Age Ranges: '≥ 18 years at index date'
• Sex/Gender: Aligning with study design and biological hypotheses
• Geographic Constraints: Limiting to specific sites or regions
• Vital Signs: Weight, height, or blood pressure ranges These boundaries are critical for matching cohorts to the target product label or research question.

A patient must have sufficient observational data to be evaluated. This prevents immortal time bias.

• Minimum Observation Period: e.g., '≥ 365 days of continuous enrollment prior to index'
• Data Completeness: Requiring at least one recorded visit or lab result in a window
• Site Capability: Ensuring the institution records the specific data types needed (e.g., structured pathology reports) This ensures the absence of a criterion is a true negative, not missing data.

Exclusion criteria are the inverse of inclusion—they refine the cohort by removing confounders.

• Contraindications: Removing patients with conditions that make the exposure unsafe
• Competing Risks: Excluding patients with conditions that prevent the outcome (e.g., bilateral amputation prevents a foot ulcer outcome)
• Data Quality Exclusions: Removing patients with implausible values (e.g., negative age) A clean cohort is defined as much by who is left out as who is let in.