Context is all you need: A causal framework for governing bias in clinical AI
5 affiliations
- 1BRIDGE GenAI Lab, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, Massachusetts, United States of America
- 2Department of Neurology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, Massachusetts, United States of America
- 3The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, New York, New York, United States of America
- 4The Hasso Plattner Institute for Digital Health at Mount Sinai, Mount Sinai Health System, New York, New York, United States of America
- 5Department of Radiology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, Massachusetts, United States of America
Abstract
Clinical AI systems can produce different recommendations when a patient’s race, income, housing status, or gender identity changes, but a differential alone does not establish unjustified bias. What separates a clinically justified differential from an unjustified one is not its size or direction but its origin: whether the feature acts through a clinically justified pathway, or marks how patients of that group have been treated. That distinction is causal, and a model trained only to detect statistical association records that a feature carries information but never why. Existing evidence does not establish reliable recovery of this adjudication inside that model; we therefore argue for explicit causal and normative supervision, whether supplied by an external layer, built into training, or both. We set out a six-criterion rubric, four approaches to implementing it, and a two-sided evaluation reporting preserved clinical sensitivity alongside reduced bias.
Related
- The Principle of Natural Selection and its Governing Interactions of Reproduction and Speciation: A Formal Framework for Reproductive Architecture and Ancestry
- Intercellular communication and multi-organ axis interactions in the pathogenesis of MASLD: a signal-class synthesis graded by disease context and causal status
- miRAssist: a context-aware, evidence integration framework for interpretable miRNA-target prioritization
- Forest-weighted causal correlation: A scalable framework for causal inference in ecological time series
- Do We Need to Assess Quality-of-Life with Both the St George’s Respiratory Questionnaire and the EuroQol 5-Dimension Questionnaire in a Clinical Study with an Economic Component: Insights from the REVOLENS Study in Severe Emphysema
BibTeX
@article{gorenshtein2026context,
title = {{Context is all you need: A causal framework for governing bias in clinical AI}},
author = {Alon Gorenshtein and Mahmud Omar and Yiftach Barash and Girish N. Nadkarni and Eyal Klang},
journal = {PLOS Digital Health},
year = {2026},
month = nov,
doi = {10.1371/journal.pdig.0001819},
url = {https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001819},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Context is all you need: A causal framework for governing bias in clinical AI
AU - Alon Gorenshtein
AU - Mahmud Omar
AU - Yiftach Barash
AU - Girish N. Nadkarni
AU - Eyal Klang
PY - 2026
DA - 2026/11/06
JO - PLOS Digital Health
DO - 10.1371/journal.pdig.0001819
UR - https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001819
AB - Clinical AI systems can produce different recommendations when a patient’s race, income, housing status, or gender identity changes, but a differential alone does not establish unjustified bias. What separates a clinically justified differential from an unjustified one is not its size or direction but its origin: whether the feature acts through a clinically justified pathway, or marks how patients of that group have been treated. That distinction is causal, and a model trained only to detect statistical association records that a feature carries information but never why. Existing evidence does not establish reliable recovery of this adjudication inside that model; we therefore argue for explicit causal and normative supervision, whether supplied by an external layer, built into training, or both. We set out a six-criterion rubric, four approaches to implementing it, and a two-sided evaluation reporting preserved clinical sensitivity alongside reduced bias.
SN - 2767-3170
ER -