Machine learning approach for differentiating iron deficiency anemia and thalassemia using random forest and gradient boosting algorithms

Abstract

Formulas based on red blood cell indices have been used to differentiate between iron deficiency anemia (IDA) and thalassemia (Thal). However, they exhibit varying efficiencies. In this study, we aimed to develop a tool for discriminating between IDA and Thal by using the random forest (RF) and gradient boosting (GB) algorithms. Complete blood count data from 1143 patients with anemia and low mean corpuscular volume were collected (382 patients with IDA, 635 with Thal, and 126 with IDA and Thal). The data were randomly divided into the training and testing datasets in a ratio of 80:20. The RF and GB models had good diagnostic performances for predicting IDA and Thal in the training and testing datasets. In the testing dataset for predicting binary outcomes, GB and RF both had an accuracy of 90.7%, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.953. A lower diagnostic performance was observed when patients with IDA and Thal were included. GB and RF showed accuracies of 80.4% and 82.2%, respectively, and AUC-ROC values of 0.910 and 0.899, respectively. In conclusion, we developed a machine learning approach using GB algorithm. This tool is potentially useful in Thal- and IDA-endemic regions.

Keywords: Gradient boosting; Iron deficiency anemia; Machine learning; Random forest; Thalassemia.

Fig. 1

Study population flow and distribution of cases in the training and testing dataset. CBC, complete blood count; MCV, mean corpuscular volume

Fig. 2

Variable importance from multiclass outcomes (Thal, IDA, IDA with Thal) of the random forest model (a) and gradient boosting model (b).