A Machine Learning Model for Predicting Fetal Hemoglobin Levels in Sickle Cell Disease Patients

Abstract

Sickle cell disease is one of the commonest genetic diseases and is defined as a decrease in hemoglobin concentration in the blood. The main known factor that can alleviate the disease is the persistence of fetal Haemoglobin (HbF) and thus the aim of our research is to build a model to predict the HbF% of patients based on the 3 regulating genes of the disease (BCL11A, Xmm1-HBG2, HBS1L-MYB). A machine-learning approach is employed in order to improve the accuracy of the model, with various algorithms of that type being explored. At the end, the K-Nearest Neighbors algorithm is chosen and an initial version of it is implemented and tested. Finally, the algorithm is optimized enabling our optimized model to predict the HbF% of a patient with 87.25% accuracy, a major improvement over the existing alternative that has a mean error of 336.33%. Furthermore, 93.45% of our predictions have a sheer error that is less than 0.5 and all these facts reinforce the strength of our model as a quick and accurate estimation tool for small and medium-sized clinical trials, where fast HbF% predictions can help adjust for genetic background variability that obscures test outcomes.

A comparative analysis of haemoglobin variants using machine learning algorithms

“…Additionally, their study revealed that complete blood count (CBC) parameters were effective in discriminating between IDA and β-thalassemia for patients. Oikonomou et al [10] built a model that can predict the percentage fatal haemoglobin (HbF%) of patients. The authors explored and compared the accuracy of various machine learning algorithms like Decision Tree, Gradient Boosted Trees, Linear Regression, K-Nearest Neighbors, Neural Network, Random Forest, and Gaussian Process.…”