Determination of the effect of red blood cell parameters in the discrimination of iron deficiency anemia and beta thalassemia via Neighborhood Component Analysis Feature Selection-Based machine learning

Abstract

Differential diagnosis of iron deficiency anemia (IDA) and β-thalassemia is a time-taking and costly procedure. Complete blood count (CBC) is a quick, inexpensive, and easily accessible test which is used as the primary test for the diagnosis of anemia. However, as CBC cannot successfully discriminate between IDA and β-thalassemia, advanced techniques are needed. To date, numerous red blood cell (RBC) indices have been investigated and various parameters have been proposed for each index. In the present study, a differential diagnosis of IDA and β-thalassemia was performed by using RBC indices and machine learning techniques including Support Vector Machine (SVM) and K-Nearest Neighbor (KNN). The RBC indices were used as input parameters for the classifier and the performances of SVM and KNN were evaluated separately, in order to determine the effectivity of both techniques. Fewer parameters were given as an inputs to machine learning algorithms, and higher performance was achieved. On the other hand, a feature selection technique, the Neighborhood Component Analysis Feature Selection (NCA) algorithm, was used for selecting features from the datasets, and the parameters selected via NCA provided high performance (97% Area Under the ROC curve [AUC]). Taken together, the results indicated that the RBC indices used in the study showed higher performance compared to those reported in the literature. By using these indices, not only the individual effect of each index parameter on the machine learning model was investigated but also a different subset of features from those employed in the literature was established. In addition, as distinct from the literature, the study revealed that different CBC parameters were efficient in distinguishing between IDA and β-thalassemia in male and female patients. Accordingly, the RBC indices employed in the study can be easily and inexpensively used in clinical and daily practice for the discrimination of IDA and β-thalassemia.