Fairness-Aware and Trustworthy Machine Learning Framework for Secure Stroke Risk Prediction in Privacy-Sensitive Clinical Data Environments

Abstract

Worldwide stroke is still among the top ten causes of death and permanent disability, thus the importance of early and reliable risk prediction systems. Current machine learning (ML) models in healthcare, however, have been plagued with issues related to data privacy leakage, algorithmic bias, and lack of interpretability, and have failed to reach clinical use. To predict stroke risk securely in privacy-sensitive clinical settings, this study proposes a Fairness-Aware and Trustworthy Machine Learning Framework (FAT-MLF). This framework combines privacy-preserving data management with differential privacy, fairness-oriented optimization and fairness constraints, such as demographic parity, and explainable AI with interpretability, such as SHAP. To capture the nonlinear interactions between clinical features, a hybrid predictive model using the Gradient Boosting Machine (GBM) and deep neural networks is used. The experimental results on the stroke benchmark dataset with 5000 patient data records show the proposed model with an accuracy of 96.8%, a precision of 95.4%, a recall of 94.9%, and an F1-score of 95.1%, which reduces the demographic bias by 32% compared to baseline ML models. Under ε-differential privacy constraints, privacy analysis verifies that the data is resistant to a data reconstruction attack. The findings suggest that embedding fairness, privacy, and trustworthiness substantially enhances predictive accuracy and ethical adherence of clinical AI systems. The framework, as proposed, is scalable and can be implemented in real-world healthcare systems to safely and fairly predict stroke risk.