The Sociodemographic Biases in Machine Learning Algorithms: A Biomedical Informatics Perspective.

Area of Science:

• Healthcare AI
• Machine Learning Bias
• Health Equity

Background:

• Artificial intelligence (AI) and machine learning (ML) models are increasingly used for healthcare risk assessment and clinical decision-making.
• These algorithms can inadvertently perpetuate societal biases present in training data, leading to discrimination.
• Biases can stem from sociodemographic factors (race, gender, age, insurance, socioeconomic status) and erroneous electronic health records.

Purpose of the Study:

• To outline sociodemographic, training data, and algorithmic biases in AI/ML models used for healthcare risk assessment.
• To highlight how these biases undermine sound medical decision-making and contribute to health disparities.
• To propose recommendations for mitigating these biases in healthcare AI.

Main Methods:

• Review and overview of various biases including gender, race, ethnicity, age, socioeconomic status, and insurance status bias.
• Examination of algorithmic biases, biased evaluations, implicit bias, selection/sampling bias, and biased data distributions.
• Discussion of cultural biases, conformation bias, information bias, and anchoring biases.

Main Results:

• Identified numerous sources of bias in AI/ML models, including sociodemographic factors and data errors.
• Highlighted the potential for these biases to exacerbate existing health inequities and socioeconomic disparities.
• Emphasized the significant social and economic consequences of biased AI in healthcare.

Conclusions:

• Addressing biases in AI/ML models is essential for ensuring fair and accurate healthcare risk assessment.
• Recommendations include improving large language model training data and employing de-biasing techniques.
• Mitigation strategies involve techniques like counterfactual data augmentation, fine-tuning, and algorithmic modifications.