A new paper shows that fixing AI bias for one demographic trait at a time does not fix it for patients who sit at the overlap of several.
Researchers built clinical prediction models on the MIMIC-Eye and MIMIC-IV ED datasets, which combine text, time series, tabular data, event logs, and images from real electronic health records. They used three pretrained clinical language models (MedBERT, Clinical BERT, and Clinical BioBERT) to learn a single unified text representation across those data types. Then they tested how well bias-mitigation techniques designed for single attributes, like race or gender alone, held up for intersectional subgroups formed by combining several attributes at once. The single-attribute fixes lost effectiveness at those intersections, so the team built a mitigation approach tuned to each subgroup specifically, and it held up across datasets, subgroups, and all three language models.
This matters because most fairness audits in healthcare AI still check one attribute at a time, asking whether a model treats men and women differently or different races differently, and call it done. That approach can miss patients who fall into multiple minority categories at once, and the multimodal nature of EHR data means bias can hide in different data types depending on the group.
A model that passes a standard bias test can still be quietly failing the patients standard bias tests were never built to notice.