Explainable Artificial Intelligence for Trustworthy Machine Learning in High-Stakes Applications

Abstract

Machine learning adoption in domains such as healthcare, finance, autonomous driving, and law demands not only predictive accuracy but also transparency and trustworthiness. Conventional black box approaches, while powerful, often lack interpretability, limiting their acceptance in mission critical applications. The selected application domains—healthcare, autonomous driving, finance, and legal systems—are systematically chosen as representative high-stakes environments where decision errors can lead to severe consequences, including risks to human life, financial loss, and legal or ethical violations. The proposed LEAF-TML framework is specifically designed to address these contexts by ensuring transparency, accountability, and reliability under conditions where trust and interpretability are critical for decision-making. To address this challenge, we propose LEAF TML, a layered explainable AI framework for trustworthy machine learning. LEAF TML integrates governance, interpretable and hybrid modeling, multi-level explanation generation, human in the loop oversight, assurance monitoring, and adaptive feedback into a closed loop architecture. The framework was validated on four representative datasets: intensive care unit electronic health records, urban driving logs, financial transaction data, and legal case archives. Results demonstrated that LEAF TML consistently outperformed baselines including Logistic Regression, Decision Trees, Explainable Boosting Machines, Random Forests, and XGBoost. It achieved accuracy and fidelity scores of 0.930 and 0.940 in healthcare, 0.880 and 0.910 in driving, 0.920 and 0.930 in finance, and 0.880 and 0.920 in law. Additionally, trust perception increased by 22.84 percent, while explanation latency was minimized at 0.039 seconds. Statistical analyses confirmed the robustness, stability, and significance of improvements, establishing LEAF TML as a scalable, ethical, and transparent solution.