The limits of debiased clinical language models for cross-hospital generalization
Angyang Guo, Wenna Chen, Yan Zhang, Mingxia Ji, Bin Fang, Hongbin Huang · 2026
Read The limits of debiased clinical language models for cross-hospital generalization by Angyang Guo, Wenna Chen, Yan Zhang, Mingxia Ji, Bin Fang, Hongbin Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-13522466.txt.
Description
Angyang Guo and colleagues showed that fairness fixes for clinical language models can fail when models move to a new hospital. Published in 2026, the study tuned a clinical BERT model on one intensive care database and tested it on another, predicting ICU mortality. Fairness work aims for equal performance across patient groups, yet most tests stay inside one hospital.
Fairness-only training looked best at home but worst away, with its fairness gap growing nearly fivefold across sites. The team traced the failure to hospital-specific writing habits leaking into the model, which interacted with patient-group signals in new settings. A leakage-aware training recipe cut the outside gap sharply while lifting accuracy. The authors urge domain-aware fairness audits before any cross-hospital rollout.
Catalog record
| CREATOR | Angyang Guo, Wenna Chen, Yan Zhang, Mingxia Ji, Bin Fang, Hongbin Huang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |