Predicting thalassemia using deep neural network based on red blood cell indices

Area of Science:

• Computational biology
• Medical diagnostics
• Machine learning in healthcare

Background:

• Traditional thalassemia screening relies on red blood cell (RBC) indices.
• Machine learning approaches are emerging as superior alternatives.

Purpose of the Study:

• To develop and evaluate deep neural networks (DNNs) for thalassemia prediction.
• To compare DNN model performance against traditional statistical screening methods.

Main Methods:

• Constructed 11 DNN models and 4 traditional statistical models using a dataset of 8693 records.
• Included genetic tests and 11 other features for model development.
• Analyzed feature importance to interpret DNN model predictions.

Main Results:

• The best DNN model achieved an area under the receiver operating characteristic curve of 0.960 and accuracy of 0.897.
• DNN models demonstrated significant improvements in sensitivity, specificity, and predictive values over traditional methods.
• Model performance decreased without key features like age, RBC distribution width (RDW), sex, or white blood cell (WBC) and platelet (PLT) counts.

Conclusions:

• The developed DNN model significantly outperformed existing thalassemia screening methods.
• RDW and age were identified as the most crucial features for DNN model accuracy.
• Sex and the combination of WBC and PLT also contributed to model performance, while other features were less impactful.