Rethinking fairness in unsupervised healthcare AI: A methodological scoping review.

Area of Science:

• Healthcare AI
• Machine Learning Ethics
• Unsupervised Learning

Background:

• Fairness in supervised machine learning is well-studied, but its application to unsupervised learning in healthcare is limited.
• Unsupervised learning discovers patient subgroups and disease trajectories without labels, making fairness assessment challenging.
• Existing research lacks clear definitions and evaluation methods for fairness in unsupervised healthcare AI.

Purpose of the Study:

• To systematically review how fairness is conceptualized, operationalized, and evaluated in unsupervised healthcare AI.
• To identify algorithmic mechanisms, evaluation strategies, and methodological assumptions in this field.
• To propose a framework for structuring the diverse approaches to fairness in unsupervised healthcare AI.

Main Methods:

• Conducted a PRISMA-guided methodological scoping review of relevant literature.
• Focused on identifying fairness definitions, algorithmic approaches, and evaluation strategies.
• Analyzed records based on data modalities, unsupervised learning techniques, and fairness definitions used.

Main Results:

• Significant growth in interest in fairness-aware unsupervised healthcare AI, but with considerable heterogeneity and inconsistency.
• Identified five families of fairness approaches: Individual Fairness, Performance Dependence, Welfare-Anchored, Statistical Inference-based, and Representation Parity.
• Recurring challenges include defining fairness without labels, limited clinical expertise integration, and weak alignment with medical validity.

Conclusions:

• Fairness in unsupervised healthcare AI is an emerging field with unsettled conceptual foundations.
• Current approaches use diverse, sometimes incompatible, notions of equity, necessitating clearer theoretical grounding.
• Future progress requires explicit fairness goals, domain expertise integration, participatory evaluation, and alignment with clinical validity.