van der Meijden S L, Wang Y, Arbous M S, Geerts B F, Steyerberg E W, Hernandez-Boussard T
Department of Intensive Care Medicine, The Leiden University Medical Center, Leiden, The Netherlands.
Healthplus.ai B.V.
medRxiv. 2025 Mar 24:2025.03.24.25324500. doi: 10.1101/2025.03.24.25324500.
Artificial Intelligence (AI)-based prediction models, including risk scoring systems and decision support systems, are increasingly adopted in healthcare. Addressing AI fairness is essential to fighting health disparities and achieving equitable performance and patient outcomes. Numerous and conflicting definitions of fairness complicate this effort. This paper aims to structure the transition of AI fairness from theory to practical application with appropriate fairness metrics. For 27 definitions of fairness identified in the recent literature, we assess the relation with the model's intended use, type of decision influenced and ethical principles of distributive justice. We advocate that due to limitations in some notions of fairness, clinical utility, performance-based metrics (area under the receiver operating characteristic curve), calibration, and statistical parity are the most relevant group-based metrics for medical applications. Through two use cases, we demonstrate that different metrics may be applicable depending on the intended use and ethical framework. Our approach provides a foundation for AI developers and assessors by assessing model fairness and the impact of bias mitigation strategies, hence promoting more equitable AI-based implementations.
包括风险评分系统和决策支持系统在内的基于人工智能（AI）的预测模型在医疗保健领域的应用越来越广泛。解决人工智能的公平性问题对于消除健康差距、实现公平的绩效和患者治疗效果至关重要。众多相互冲突的公平性定义使这项工作变得复杂。本文旨在通过适当的公平性指标构建人工智能公平性从理论到实际应用的转变。对于近期文献中确定的27种公平性定义，我们评估了它们与模型预期用途、受影响决策类型以及分配正义伦理原则的关系。我们主张，由于某些公平性概念存在局限性，临床效用、基于性能的指标（受试者操作特征曲线下面积）、校准和统计均等是医疗应用中最相关的基于群体的指标。通过两个用例，我们证明了根据预期用途和伦理框架，可能适用不同的指标。我们的方法通过评估模型公平性和偏差缓解策略的影响，为人工智能开发者和评估者提供了一个基础，从而促进更公平的基于人工智能的实施。
