Bias in medical AI: Implications for clinical decision-making

Summary: Reviews how algorithmic bias in clinical decision-support tools can worsen healthcare disparities and outlines mitigation approaches.

Relevance: Fits Health & Care and Inequality & Algorithmic Harm: reviews how bias baked into clinical decision-support algorithms produces worse predictions for underrepresented patients, directly relevant to social workers advocating for clients navigating medical systems that increasingly rely on these tools.

Key findings: The deployed Epic Sepsis Model showed significantly worse AUC, sensitivity, and specificity after real-world deployment and missed roughly two-thirds of sepsis cases; an ICU mortality-prediction study using the MIMIC-III dataset found recall rates as low as 25% across racial groups. Proposed mitigations include diversifying training data, subgroup bias audits (equalized odds, predictive parity), and interpretability tools like SHAP and LIME.

Read the original: https://pmc.ncbi.nlm.nih.gov/articles/PMC11542778/