Entropy to Ethics: A Unified Information-Theoretic Framework for Interpretable Healthcare Data Mining

Abstract

The increasing deployment of machine learning models in healthcare has brought about transformative potential, but also critical challenges related to model interpretability, fairness, and ethical compliance. In this paper, we present a unified framework grounded in information theory to rigorously quantify and enhance interpretability, fairness, and privacy within data mining systems applied to healthcare contexts. By leveraging core principles such as mutual information, conditional entropy, and Kullback-Leibler divergence, we formulate a composite loss function that enables trade-offs between predictive performance, transparency and ethical guarantees. The framework is instantiated in the form of entropy-regularized models trained on three real-world healthcare datasets: MIMIC-III, eICU, and UCI Heart Disease. We evaluate our models on classification performance, explanation fidelity, demographic parity, and privacy leakage sensitivity. Empirical findings reveal that our framework achieves higher interpretability and fairness with only marginal impact on classification accuracy, demonstrating that ethical AI in healthcare need not come at the cost of performance.

Ähnliche Arbeiten