Bias Mitigation in Primary Health Care Artificial Intelligence Models: Scoping Review.

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Algorithmic Bias

Background:

• Artificial intelligence (AI) predictive models in primary health care offer potential for population health enhancement.
• However, these models risk perpetuating or amplifying biases against diverse groups.
• A gap exists in understanding strategies to assess and mitigate bias in primary health care algorithms.

Purpose of the Study:

• To describe bias mitigation strategies in primary health care AI models.
• To identify targeted diverse groups and protected attributes.
• To evaluate the impact of these strategies on bias reduction and model performance.

Main Methods:

• A scoping review following Joanna Briggs Institute (JBI) guidelines was conducted.
• Searches included Medline, CINAHL, PsycINFO, and Web of Science databases (Jan 2017–Nov 2022).
• Data extraction and quality appraisal were performed on 17 included studies.

Main Results:

• Race/ethnicity and sex were the most frequently investigated protected attributes.
• Bias mitigation approaches included data/model modification, electronic health record sourcing, human-in-the-loop systems, and ethical principles.
• Algorithmic preprocessing (relabeling, reweighing) and natural language processing showed promise, but sometimes exacerbated errors.

Conclusions:

• Bias mitigation is more effective with open-sourced data, stakeholder engagement, and during the preprocessing stage.
• Further empirical studies are needed, including a broader range of groups.
• Validation and expansion of findings are crucial for equitable AI in primary care.