Distribution-Based Similarity Measures Applied to Laboratory Results Matching

Area of Science:

• Medical Informatics
• Computational Linguistics
• Health Data Science

Background:

• International laboratory terminologies are crucial for data reuse across hospital information systems.
• Current semantic interoperability methods often rely on language-based matching.
• Distribution matching, based on statistical similarity, offers an alternative strategy for terminology alignment.

Purpose of the Study:

• To design and evaluate a structured framework for distribution matching of concepts described by continuous variables.
• To improve the accuracy and efficiency of aligning disparate laboratory terminologies within healthcare databases.

Main Methods:

• A framework combining distribution matching and machine learning techniques was developed.
• A training sample of correct/incorrect terminology correspondences was used to build a match probability score.
• The model returns and ranks best candidate term matches based on computed probabilities.

Main Results:

• The proposed framework achieved a 95% success rate, correctly matching 96 out of 101 terms within the top 5 candidates.
• The study successfully mapped terms from Lille University Hospital's concepts to the MIMIC-III database.
• The developed model provides a probability score for term correspondence.

Conclusions:

• The developed open-source framework effectively aligns international laboratory terminologies using distribution matching and machine learning.
• This approach facilitates easier expert validation of terminology alignment, crucial for inter-hospital data analysis.
• The system's top-k suggestions enhance the practical application of terminology mapping in healthcare settings.