Comparison of reference distributions acquired by direct and indirect sampling techniques: exemplified with the Pediatric Reference Interval in China (PRINCE) study

Abstract

Background Our study aimed to compare the reference distributions of serum creatinine and urea obtained by direct sampling technique and two indirect sampling techniques including the Gaussian Mixture Model (GMM) and the Self-Organizing Map (SOM) clustering based on clinical laboratory records, so that the feasibility as well as the potential limitations of indirect sampling techniques could be clarified. Methods The direct sampling technique was used in the Pediatric Reference Interval in China (PRINCE) study, in which 15,150 healthy volunteers aged 0 to 19 years were recruited from 11 provinces across China from January 2017 to December 2018. The indirect sampling techniques were used in the Laboratory Information System (LIS) database of Beijing Children’s Hospital, in which 164,710 outpatients were included for partitioning of potential healthy individuals by GMM or SOM from January to December 2016. The reference distributions of creatinine and urea that were established by the PRINCE study and the LIS database were compared. Results The density curves of creatinine and urea based on the PRINCE data and the GMM and SOM partitioned LIS data showed a large overlap. However, deviations were found in reference intervals among the three populations. Conclusions Both GMM and SOM can identify potential healthy individuals from the LIS data. The performance of GMM is consistent and stable. However, GMM relies on Gaussian fitting, and thus is not suitable for skewed data. SOM is applicable for high-dimensional data, and is adaptable to data distribution. But it is susceptible to sample size and outlier detection strategy.