AUTHORS
Antoine Martin Jr, PhD Candidate, Matthew Mckoy2, Jinrui Xu, PhD3, Jiang Kong, PhD4, Nicole Nash5, Praise Ekomwen6, Sergei Nekhai, PhD7, James Grover Taylor VI, MD8, Victor Gordeuk, MD9, Fayuan Wen, PhD10
AFFILIATIONS
1 Howard University, Washington, District of Columbia, United States
2 Howard University, Washington, District of Columbia, United States
3 Howard University, Washington, District of Columbia, United States
4 Howard University, Washington, District of Columbia, United States
5 Howard University, Washington, District of Columbia, United States
6 Howard University, Washington, District of Columbia, United States
7 Howard University, Washington, District of Columbia, United States, Howard University, Washington, United States
8 Howard University, Washington, District of Columbia, United States, Howard University, Washington, United States
9 Sickle Cell Center, University of Illinois, Chicago, United States
10 Howard University, Washington, District of Columbia, United States
ABSTRACT
Introduction : Iron overload is a major complication in patients with sickle cell disease (SCD) receiving chronic transfusions. Early identification of patients at risk remains challenging because iron accumulation develops gradually, and patients with similar transfusion histories often exhibit different degrees of iron overload. While serum ferritin is commonly used to monitor iron burden, it may be influenced by inflammation and other clinical factors, limiting its accuracy as a standalone marker. Machine learning approaches offer an opportunity to integrate complex demographic, clinical, laboratory, and transfusion-related data to identify patients at greatest risk. By applying machine learning techniques such as Random Forest and Elastic Net, we aim to identify the most influential predictors of iron overload.
Methods: Clinical data with from 370 pediatric SCD participants in the Howard PUSH cohort (mean age11.75±5.24 years, range 3-20 years) were analyzed. Random Forest and Elastic Net machine learning models were developed using 617 demographic, clinical, laboratory, treatment, and transfusion-related variables to identify predictors of iron overload. Variables included transfusion history, serum ferritin levels, hematologic biomarkers, hydroxyurea use, chronic transfusion therapy, stroke history, and other phenotypic characteristics. The dataset was randomly split into training (70%) and testing (30%) subsets for model development and evaluation. Random Forest feature importance was evaluated using the mean decrease in Gini, while Elastic Net predictors were ranked by the magnitude of their regression coefficients.
Results: Iron overload prevalence increased substantially with cumulative lifetime transfusion exposure, rising from 0% among patients who received with fewer than four lifetime transfusions to 63% among those receiving at least 20 lifetime transfusions. Machine learning analyses using Random Forest and Elastic Net consistently identified transfusion burden as the primary predictor of iron overload. Random Forest analysis identified ferritin as the strongest predictor of iron overload, followed by iron chelation medication Exjade treatment, lifetime red blood cell transfusion count, chronic transfusion program participation, Desferal treatment, and stroke history. Additional predictors included stroke history, Hemoglobin A1 by electrophoresis (%), BUN, Vascular Endothelial Growth Factor level, Absolute Granulocyte Count (AGC), white blood cell count (WBC) and age. Elastic Net analysis corroborated these findings, ranking transfusion frequency, transfusion type, lifetime RBC transfusions, and chronic transfusion status among the most influential predictors.
Discussion: The integrated modeling framework developed here offers a roadmap for precision risk assessment and early identification of high-risk pediatric SCD patients for iron overload. Random Forest and Elastic Net modeling effectively identified clinically meaningful predictors of iron overload in pediatric SCD. Machine learning approaches may facilitate early risk stratification and help identify patients who would benefit from closer monitoring and timely intervention.
ACKNOWLEDGMENTS: This work was supported by NIH Research Grants (U24MD015970 and 1R01HL125005) NSF Grant 2406155.
DISCLOSURES
Antoine Martin Jr, PhD Candidate: No relevant conflicts of interest to declare.