Prediction of β-thalassemia carrier using federated learning and explainable AI

TL;DR

This paper introduces a privacy-preserving method for β-thalassemia carrier screening using federated learning and explainable AI, achieving high accuracy while protecting patient data.

Contribution

The novel contribution is a federated multi-kernel SVM framework with XAI techniques for β-thalassemia screening that maintains privacy and achieves high diagnostic performance.

Findings

The federated model achieved 98.4% accuracy, 99.2% sensitivity, and 98.8% specificity on blood count records.

Hemoglobin level and mean corpuscular volume were identified as the most significant predictors.

The model performs as well as centralized methods while preserving data privacy.

Abstract

Millions of people worldwide suffer from β-thalassemia, an inherited blood disorder that requires precise carrier screening to avoid serious health issues. Conventional centralized screening techniques rely on combining patient data, which raises privacy and legal issues under regulations like General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA). Although machine learning has increased the accuracy of diagnoses, its reliance on shared data raises issues with data security and makes it more difficult to develop collaborative models. Federated learning provides a solution by allowing multi-center collaboration and protecting privacy by training models locally at each clinical site and sharing only model parameters. In this work, a federated multi-kernel support vector machine (SVM) framework is developed, which aggregates updates through federated averaging and deploys linear, polynomial, radial basis function, and Deep kernel (DK) on client devices. We also incorporate explainable AI methods SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) to decipher forecasts and pinpoint important hematological characteristics. Our federated model performs on par with centralized methods, achieving 98.4% accuracy, 99.2% sensitivity, and 98.8% specificity when tested on 5,066 complete blood count records. The most significant predictors, according to the explainability analyses, are hemoglobin level and mean corpuscular volume. Our results open the door for scalable, transparent, and compliant β-thalassemia screening across dispersed healthcare systems by demonstrating that federated multi-kernel SVMs with Explainable Artificial Intelligence (XAI) can provide high diagnostic performance while protecting patient privacy.

Figures