Ethical Frameworks for Deploying AI in Predictive Population Health Management

Abstract

The potential generated by the inclusion of Artificial Intelligence (AI) in public health systems, to help improve disease surveillance, diagnosis and healthcare delivery has raised considerable interest. Yet, fairness, transparency, data privacy, and public trust are all critical issues that stand in the way of widespread and ethical MBC deployment. Ultimately, this paper aims to introduce an Ethical AI Framework designed for public health, which works towards cultivating ethical and inclusive applications of AI. We take these ethical principles and implementable practices such a FAIR, Accountability, Transparency, Privacy - and convert that into the architecture for this approach. We evaluate the framework on a real-world clinical dataset and show that it translates into desirable properties conducive to improving AI performance, interpret ability, and trust when overturning the ethical rock. The model uses differential privacy and interprets able predictions, plus fairness over demographic groups via SHAP, AIF360 and Opacus. Our results show that, even though a slight degradation in accuracy is incurred, there are significant improvements in gender fairness (from 12.5 % to 3.4 %), interpret ability (SHAP score: 0.88) and user trust levels (increased from 62 % to 87 %). These results showed that incorporating ethical principles in AIembedded systems is not only a possibility but also a necessity for an equitable and sustainable AI deployment into the public health settings.

Ähnliche Arbeiten