P127: Extracting Lab Value Reference Ranges from Real World Neonatal ICU Data

Poster Presenter

Nick Henscheid

Senior Scientist
Critical Path Institute
United States

Objectives

Establishing reliable lab value reference ranges for neonates is challenging, which impedes progress in clinical care and trial design. Our goal is to develop a pipeline to produce reliable reference ranges from real world Electronic Health Record (EHR) data.

Method

Data for six priority labs was extracted from the INC database of EHRs, retaining relevant covariates. A combination of the refineR technique (Ammer et al. 2021) and GAMLSS regression were used to build reference range curves as a function of the covariates.

Results

To develop our pipeline, data were extracted from three EHR databases containing hepatic labs (AST, ALT & GTT), renal labs (BUN & Creatinine) and total platelet count. Gestational age in weeks, postnatal age in days at the time of measurement, birth weight and sex were retained. Patients for whom a time of death was recorded were excluded. For each analyte, a GAMLSS regression model was built with distributional family being a modified Box-Cox transformed Gaussian that allows for skewness and kurtosis, as well as assuming that some unknown percentage of subjects are abnormal. The resulting reference curves smoothly predict the 2.5th and 97.5th percentiles as a function of the covariates. Preliminary comparison to existing tabulated ranges for normal neonates indicates a qualitative concordance, indicating soundness of our methodological pipeline.

Conclusion

In order for lab values to be reliably incorporated in to clinical and research practice, it is necessary to understand what normal ranges for those values are and how they vary as a function of continuous and discrete covariates. We have demonstrated that it is feasible to use Real World Data from EHRs together with purpose-built statistical techniques to produce plausible reference range curves. In the future, our methods could be used to develop severity scales for adverse events, to design better trial endpoints for neonates, and to improve clinical care in NICUs. Further work will aim to investigate additional covariate effects such as medications and subsequent diagnosis, to propose a validation framework for RWD-based reference ranges, and to explore an interactive platform to better display results.