Design and development of a machine learning-based predictive modelling tool to accurately predict thalassemia carrier state using full blood count indices and haemoglobin variants

Abstract

Introduction: This study attempted to use artificial intelligence (AI) techniques to develop a cost-effective and time-saving tool to accurately predict the alpha thalassemia carrier state using a simple blood test, without the need to perform genetic testing. Methods: Ethical clearance was obtained from the Ethics Review Committee of PGIM, University of Colombo (ERC/PGIM/2020/068. Two AI algorithms, random forest and an Artificial Neural Network (ANN) were used to train three models using complete blood count and high-performance liquid chromatography data from a database of 288 cases from the Human Genetics Unit (HGU) of the Faculty of Medicine, Colombo. Performances of the models were assessed using two metrics: accuracy and F1 score, which is a measure of sensitivity and positive predictive value. Results: Dataset included predominantly alpha thalassemia carriers (n = 149), but also individuals with normal red cell indices (n = 55), iron deficiency anaemia (n = 21) and beta thalassemia carriers (n = 17). Model 1 was trained to differentiate between alpha thalassemia carriers and normal individuals, and random forest outperformed the ANN with an Accuracy of 93.5% and an F1 score of 87.5%. Model 2 differentiated between alpha thalassemia silent carriers and alpha thalassemia traits, and random forest again showed superior performance with an accuracy of 88.8% and an F1 score of 86.3%. Model 3 differentiated between thalassemia carriers, iron deficiency, and normal individuals; but did not show satisfactory performance. Conclusions: Model 1 and 2 showed satisfactory performance and can be combined into a diagnostic tool to identify alpha thalassemia carrier states. The tool could be used in situations where there are constraints to genetic testing, after validating prospectively with a larger dataset.