Presentation (CTI)

OBJECTIVES: Artificial intelligence (AI) is transforming healthcare through diagnostics, clinical decision support, risk prediction, imaging, and resource allocation. However, AI trained on historical healthcare data may amplify existing health inequities. This research assesses how such biases manifest across populations, applications, and bias types, and evaluates current mitigation strategies.
METHODS: A structured literature review was conducted using PubMed, Scopus, Web of Science, Google Scholar, and selected grey literature. Studies published between January 2020, and May 2026 were screened against predefined criteria focusing on AI in healthcare, algorithmic bias, health disparities, and training data. Twenty-eight studies were included.
RESULTS: Historical healthcare data contribute to inequities by embedding systemic biases, including underrepresentation, flawed proxy measures, and prior clinical disparities, into AI outputs. Racial and ethnic minorities were most frequently studied, with differences in diagnosis rates and clinical recommendations reported. Socioeconomically disadvantaged groups (e.g., low-income, uninsured, rural populations) were affected by bias linked to proxy variables, data gaps, and misclassification. Disparities were also observed in female and other intersectional populations. Common bias types included representation bias, measurement/proxy bias, and broader historical and algorithmic biases. The highest risks were seen in risk prediction and resource allocation tools, as well as diagnostic and clinical decision-support systems directly influencing care. Mitigation strategies focused on improving dataset diversity, model auditing, and fairness-aware approaches, though real-world effectiveness remains limited.
CONCLUSIONS: AI bias reflects inequities embedded in historical health data, contributing to underdiagnosis, skewed decisions, and misclassification across populations. These biases are often amplified by AI systems, highlighting the need for improved data quality, representation, and evaluation.