A novel indirect method for deriving reference intervals through iterative data cleaning guided by self-organizing...

Area of Science:

• Clinical Chemistry
• Bioinformatics
• Data Mining

Background:

• Existing methods for deriving reference intervals (RIs) often use univariate approaches with limited data cleaning.
• Multivariate data-mining strategies offer potential for more rigorous RI derivation.

Purpose of the Study:

• To develop and evaluate novel software, SOM-clean, for indirectly deriving reference intervals (RIs) from routine laboratory databases.
• To employ self-organizing map (SOM) clustering for iterative exclusion of atypical multi-test patterns in laboratory data.

Main Methods:

• Retrieved 22 major health-screening tests (HSTs) from a Saudi Arabian laboratory database (37,285 records).
• Applied parametric Box-Cox transformation and standardization of values against initial RIs.
• Utilized self-organizing map (SOM) clustering for iterative identification and exclusion of atypical records.
• Recalculated RIs iteratively until stabilization, optimizing SOM parameters for best goodness-of-fit.

Main Results:

• SOM-clean achieved excellent goodness-of-fit (GOF) for RIs across nearly all HSTs, indicating conformity to healthy status.
• Compared to the refineR method, SOM-clean RIs were less broad and biased, especially for skewed distributions.
• GOF was optimized by referencing direct study RIs to determine optimal SOM map size and cell exclusion criteria.

Conclusions:

• SOM-clean is a practical and robust parametric tool for indirect RI estimation.
• The software utilizes a novel multivariate data cleaning scheme for improved accuracy.
• This approach enhances the reliability of RIs derived from routine laboratory data.