Poster #74: Katherine Cyr MS

A Data Harmonization and Statistical Pipeline for Continuous, Age- and Sex-Specific Pediatric Laboratory References

Katherine Cyr, Benjamin Jung, Isaac Freeman, Robert Grundmeier, Barbara Chaiyachati, Russell T Shinohara, Justine Y Hansen, Aaron Alexander-Bloch

Laboratory reference intervals are the basis for clinical interpretation of blood test results but are poorly standardized in many electronic health records (EHRs). The same analyte may be measured with different names and evaluated against differing intervals. Results are flagged as abnormal without indicating the extremity of the abnormality. In pediatric settings, references based on discrete age ranges yield unrealistic, abrupt changes in reference intervals. Continuous reference curves offer a principled alternative, enabling standardized interpretation of individual deviation from population norms. Such curves would have immediate relevance for retrospective clinical research and, ultimately, clinical care. But their construction requires overcoming missing or deprecated analyte identifiers, inconsistent measurement units, and pathology enrichment in hospital-based samples. We developed a pipeline for continuous reference curves from The Children's Hospital of Philadelphia’s EHR (93,979 patients; 39,376,265 observations). Analytes were identified using Logical Observation Identifiers Names and Codes (LOINC). Observations with missing LOINCs or equivalent LOINCs were remapped via deterministic, rule-based mapping using component names and procedural codes (48.3% combined). Units were harmonized within each analyte. Since hospital-based observations are enriched for pathology, we downweighted observations based on their predicted pathology from an eXtreme Gradient Boosting model (XGBoost median test-set AUC = 0.99). We then fitted Generalized Additive Models for Location, Scale, and Shape (GAMLSS) per analyte, with distribution families selected by Bayesian Information Criterion, yielding age- and sex-specific centile curves for 73 blood analytes. Fitted centiles were well calibrated against empirical percentiles. A median of 4.3% of clinically normal observations fell beyond the 2.5th–97.5th centiles (~5% expected). Although recorded reference ranges are themselves non-standardized, fitted centiles were concordant with clinical flags, with abnormal values concentrating at centile extremes. By transforming routinely collected EHR blood analytes into continuous, harmonized norms this work lays the groundwork for continuous reference curves using real-world clinical cohorts.