Identification of hematological biomarkers and assessment of machine learning models for sickle cell anemia severity classification

Abstract Objectives Sickle cell anemia (SCA) is a severe form of sickle cell disease (SCD). Given the rising global disease burden and the unpredictable clinical outcomes, there is a need for development of reliable methods to predict disease severity. Methods Our study involved 481 participants, including 356 SCA patients and 125 healthy controls, who reported at the Korle-Bu Teaching Hospital, Ghana. Using a mixed-methods approach, we performed a biomarker identification analysis followed by assessment of several machine learning (ML) models to predict the severity of SCA. Results Significant correlations were observed between immune cells, erythrocyte indices, and bilirubin, which highlights the chronic inflammatory state and hemolytic nature of the disease. A principal component analysis (PCA) revealed strong correlations between immune cells and erythrocyte indices with PCA1 and PCA2, indicating a significant influence of immune pathways and erythropoiesis. The all-variable model achieved an area under the receiver operating characteristics curve (AUC-ROC) of 0.98 with a 92.4% predictive accuracy. The model identified direct and total bilirubin, reticulocyte count, hydrogen sulfide, and neutrophil count as the top five biomarkers with the highest average importance (scores >1.2). Further ML assessment for prediction of SCA severity exhibited excellent discriminating performance for the C5.0 decision tree (C5.0), Random Forest (RF), XG boost (XGB), and bagged trees (TREEBAG) models, with AUC-ROC ≥80% and area under the precision recall curve (AUC-PR) ≥85%. Conclusions We identified key biomarkers associated with immune response, erythropoiesis, and oxidative stress that could serve as surrogate endpoints in clinical trials.