DEVELOPMENT AND VALUE EVALUATION OF PREDICTIVE MODELS FOR DETERMINING BLOOD SERUM FERRITIN LEVELS IN DIFFERENT CLINICAL GROUPS USING MACHINE LEARNING

Studying the capabilities of artificial intelligence (AI) is important to develop an optimal screening strategy, identify risk groups, and create less expensive laboratory tests to assess the iron status. The aim of the study is to evaluate the effectiveness of using machine learning (ML) tools to assess the iron status by the predicted serum ferritin (SF) level based on demographic data (gender and age), complete blood count (CBC), C-reactive protein (CRP) content and the historic data on the SF level. To perform ML using AI, a dataset of 52,158 patients was used. The obtained data were presented in the form of the First regression model to determine the pre dicted SF concentration and the Second model for classifying patient groups depending on different iron status by the level of known SF: 0) <15.0, 1) 15.1-100.0, 2) 100.1-300.0, 3) >300.1 μg/L. As a result, the First model demonstrated adequate predictive ability (R2=0.717), and its quality is better, the lower the SF value (the average absolute error was 2.4 μg/L for the class of patients with SF <15.0 μg/L) in the test sample. The Second model showed an even higher diagnostic ability with accuracy for different clinical groups (AUC-ROC indicator: 0.914, 0.807, 0.812, 0.891, respectively), which is important for determining patient management tactics. As a result of the study, it can be concluded that the determination of SF content using the models developed can be used as an accurate and clinically significant tool for assessing iron status in clinical practice. © 2025 Elsevier B.V., All rights reserved.