Back

Detecting Algorithmic Bias in ICU Clinical Decision: A Doubly-Robust Framework for Auditing Treatment Recommendation Disparities

Zhou, Y.-H.; Sun, G.

2025-11-17 intensive care and critical care medicine
10.1101/2025.11.16.25340359 medRxiv
Show abstract

Clinical decision support systems increasingly guide ICU care, but may perpetu-ate or amplify existing healthcare disparities. We develop a doubly-robust statistical framework for detecting and quantifying algorithmic bias in ICU treatment recommen-dations. Rather than prescribing treatment allocation, our approach audits existing clinical decision support systems to identify disparities in predicted treatment benefits across demographic groups. Analyzing 193,683 patients from the eICU database, we demonstrate the frameworks ability to detect systematic biases. For age-based anal-ysis, we identify a 5.1 percentage point mortality disparity with differential predicted treatment effects (3.2pp younger vs. 1.8pp older patients). For race-based analysis, severity-adjusted outcome disparities (average 2.2pp, reaching 5.6pp at high severity) suggest potential differences in care quality or algorithmic recommendations despite similar aggregate outcomes. We quantify how different fairness metrics (demographic parity, equalized odds, calibration) reveal distinct bias patterns, providing guidance for bias auditing in clinical AI systems. This framework enables healthcare systems to identify and address algorithmic bias before deployment, supporting more equitable clinical decision support.

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.