Bias Mitigation Across the Healthcare Artificial Intelligence Lifecycle: A Structured Narrative Review

Bias remains a difficult issue when applying artificial intelligence (AI) in healthcare. Many types of bias can appear at various stages of AI development, and bias has been shown to perpetuate disparities in vulnerable populations. Current literature tends to either be an exploration of a particular method or a review discussing multiple approaches but fails to explore how the approaches work together. This narrative review summarizes recent approaches to mitigating bias in healthcare AI and organizes them within a practical lifecycle framework spanning data acquisition, preprocessing, model training, validation, and deployment. A structured literature search was conducted using PRISMA-informed principles. Keywords "Artificial intelligence", "bias mitigation", and "artificial intelligence bias mitigation" were searched on PubMed, Embase, Web of Science, Semantic Scholar, MedRxiv, and Google Scholar. Articles from 2018 to 2024 in English that explained a method for mitigating bias in AI were included. Reviews were excluded. The search identified 340 records. After removal of duplicates and screening, 16 articles were included. Many techniques were designed to guide the composition of the training data, mainly sampling methods. Datasets are also augmented with newer methods such as adversarial learning and subpopulation techniques. New model structures, loss functions, and algorithms are being applied in medical settings. The review explains the mechanics behind these methods and the settings in which they can be used. Addressing AI bias in healthcare requires multiple approaches targeting many stages of AI development, including data preprocessing, model training, and validation. Recent innovations include adversarial learning and hybrid models. We propose a practical lifecycle protocol encouraging multifaceted approaches that address bias at all stages of development. Ongoing studies are needed to validate these methods clinically.