Algorithmic Learning of Clinically Acceptable Levels of Laboratory Test Results From Electronic Medical Records: Personalized Reference Intervals

Tez Türü: Doktora

Tezin Yürütüldüğü Kurum: Dokuz Eylül Üniversitesi, Fen Bilimleri Enstitüsü, Bilgisayar Mühendisliği, Türkiye

Tez Danışmanı: Dr. Öğr. Üyesi Özlem Aktaş,Prof. Dr. Süleyman Sevinç

Tezin Onay Tarihi: 2023

Tezin Dili: İngilizce

Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu

Özet:

Laboratory results are an important tool that provides objective data to help physicians in diagnosis and treatment processes. Results must be reported with reference ranges to be useful. Reference intervals should be determined separately according to demographic characteristics such as age and gender. It is aimed to determine reference intervals in an easy, fast, safe and economical way with machine learning algorithms. In our study, firstly, the usability of Gaussian Mixture Model and Hierarchical Clustering algorithms was validated by using the laboratory results obtained by the direct method within the scope of the "Canadian Laboratory Initiative on Pediatric Reference Intervals" study. Then, the results of the neonatal period of the Inorganic Phosphorus, Calcium, Creatinine, Neonatal Bilirubin and Urea Nitrogen tests, which were studied in the Central Laboratory of Dokuz Eylül University in 2018- 2019-2020, were taken from the hospital database, and the age partitions of the relevant tests were obtained using the algorithm we developed. It was determined that the unsupervised machine learning method we developed is a new, modern alternative to indirect methods in determining reference intervals. In this study, an unsupervised machine learning algorithm solution based on mathematical and statistical foundations, which can determine the age ranges, which is the basic step in the calculation of reference intervals, with high resolution is presented. By using the algorithmic method developed in the study, each laboratory will be able to calculate reference intervals compatible with their own population and analytical methods in an easy, fast, safe and economical way.

Keywords: Machine learning, reference intervals, electronic health records, Gaussian mixture model, hierarchical clustering