Phenotype Harmonization

The process of translating heterogeneous clinical codes from different health systems into a unified semantic framework. This involves mapping local billing codes, such as ICD-9-CM and ICD-10-CM, to standardized terminologies like SNOMED CT or the Human Phenotype Ontology (HPO).

• Lexical Matching: Algorithms identify equivalent terms through string similarity and synonym detection.
• Logical Subsumption: Hierarchical reasoning ensures a local 'Type II Diabetes' code is correctly classified under the broader 'Diabetes Mellitus' parent concept.
• Post-Coordination: Combines multiple atomic concepts to represent complex clinical statements that lack a single pre-existing code.

The mathematical transformation of quantitative laboratory measurements and vital signs to a common scale and unit of measure. This step is essential because federated sites may report the same biomarker in different units, such as mg/dL versus mmol/L for glucose.

• Linear Conversion: Applies standard conversion factors to align units.
• Harmonized Reference Ranges: Normalizes values against site-specific normal ranges to compute z-scores or multiples of the median.
• Logical Observation Identifiers Names and Codes (LOINC): The universal standard for identifying laboratory and clinical observations, ensuring that a serum sodium test is recognized identically regardless of local naming conventions.

The use of rule-based or machine learning classifiers to infer a specific disease state or clinical condition from raw electronic health record data. This moves beyond simple diagnostic codes to incorporate evidence from multiple data modalities.

• Rule-Based Definitions: Computable phenotypes using explicit logic combining diagnosis codes, medication orders, lab results, and clinical notes.
• Probabilistic Phenotypes: Machine learning models that output a likelihood score for a condition, handling the inherent uncertainty and missing data in real-world records.
• PheKB: A widely used knowledge base of validated computable phenotype algorithms for conditions like Type 2 Diabetes Mellitus and resistant hypertension.

The synchronization of clinical events along a common patient timeline to define a consistent index date for analysis. This is critical for cohort studies where the 'start' of a disease must be uniformly defined across sites with different recording practices.

• Index Date Definition: Standardizing the first qualifying event, such as the date of first diagnosis, first medication, or first abnormal lab value.
• Look-Back Windows: Defining a uniform period prior to the index date to establish baseline characteristics and exclude prevalent cases.
• Censoring Logic: Applying consistent rules for handling patients who leave the health system or die during the follow-up period.

The systematic evaluation of harmonized data to identify missingness, implausible values, and logical inconsistencies before analysis. This ensures that federated model training is not corrupted by site-specific data artifacts.

• Completeness Checks: Quantifying the proportion of missing data for each variable per site to identify systematic data capture failures.
• Plausibility Rules: Flagging values that fall outside biologically possible ranges, such as a negative age or a body temperature of 0°C.
• Temporal Consistency: Verifying that events are recorded in a logical chronological order, such as a procedure date not preceding a patient's birth date.

The structural transformation of local database schemas into a standardized, vendor-neutral format. This provides a consistent technical foundation upon which all other semantic harmonization is built.

• OMOP CDM: The Observational Medical Outcomes Partnership Common Data Model, an open community standard for observational health data.
• Standardized Vocabularies: Forces all clinical data into a single set of controlled terminologies within the CDM structure.
• Extract, Transform, Load (ETL): The technical process of converting source data into the CDM, which is often the most resource-intensive phase of harmonization.

The ability of two systems to exchange clinical data with shared, unambiguous meaning. A hemoglobin A1c value of 7.0% must be interpreted identically across sites, including the unit of measure, assay methodology, and temporal context. This requires mapping local lab codes to LOINC and harmonizing unit conversions.

• Goes beyond syntactic (format) compatibility
• Resolves 'HbA1c' vs 'Glycated Hemoglobin' synonyms
• Critical for continuous trait harmonization

Phenotype definitions often require time-window constraints to distinguish acute from chronic conditions. A 'Major Depressive Episode' phenotype may require two distinct diagnosis codes separated by 30-180 days, with no bipolar disorder codes in the preceding year. This logic is encoded in tools like ATLAS.

• Uses index dates and look-back periods
• Handles 'washout' windows for incident cases
• Prevents misclassification of single encounters

The assessment of local data fitness before federated execution. A site may have complete diagnosis codes but sparse lab results, biasing a phenotype relying on lab values. Data Quality Dashboards profile missingness, implausible values, and temporal gaps to determine if a site can contribute valid cases to a specific study.

• Checks for logical inconsistencies (e.g., male patients with ovarian cancer codes)
• Evaluates record continuity
• Informs study feasibility assessments