Strategies for mitigating artificial intelligence bias in healthcare: a systematic review

Abstract Objectives Artificial intelligence is used in healthcare to identify and manage health conditions across diverse patient populations and clinical settings, but biases in algorithms can perpetuate and exacerbate inequalities in health and healthcare delivery. While some research has focused on addressing this critical issue, there is a lack of comprehensive information on the types of strategies employed for bias mitigation in the use of artificial intelligence in healthcare and the effectiveness of these strategies. The objective of this review was to address this lack of information by identifying, categorizing, and describing the effectiveness of bias reduction strategies and fairness metrics in healthcare algorithms. Materials and Methods Following the Preferred Reporting Items for Systematic reviews and Meta-Analyses guidelines, this study categorizes and evaluates the effectiveness of bias reduction strategies and fairness metrics identified in a previous scoping review. Results The review included 35 studies. The findings are related to the stages of the algorithm lifecycle and are aligned with the authors’ institutional governance structure for algorithm development, silent evaluation, effectiveness evaluation, and deployment. The majority (60%) of the identified strategies were implemented during the algorithm development and silent evaluation stages, and all studies utilized group fairness metrics for performance measurement. Most studies (85%) reported effective bias reduction, while only a few reported ineffectiveness (8%) or no effect (5%). Conclusion There is a significant opportunity for model developers and end users to identify and reduce bias, particularly during model design. When evaluating strategy effectiveness, efforts should be measured using evidenced-based fairness metrics—such as group-based metrics—to ensure effectiveness and interpretability.