Is it possible to vaccinate AI against bias? An exploratory study in epilepsy
Epilepsy & Behavior · 2027-01 · 186 · 111379
DOI: 10.1016/j.yebeh.2026.111379 ↗Kurz gefasst
Fachartikel in der Zeitschrift Epilepsy & Behavior, 2027-01. Verfasst von Rohan Manish Bhansali und 2 weiteren Autor:innen. Beteiligte Einrichtung: Harvard Medical School, Boston, MA, USA; Beth Israel Deaconess Medical Center, Boston, MA, USA. Die Arbeit nennt 8 Literaturangaben. Gefördert von: NINDS NIH HHS, NHLBI NIH HHS, NIA NIH HHS.
Abstract (Originalsprache)
Large language models (LLMs) may respond differently to clinical vignettes containing different patient descriptions. This exploratory proof-of-concept study examined whether an instruction to reconsider irrelevant information changed such responses in epilepsy. One diagnostic and one treatment vignette each had an affluent and a disadvantaged composite profile, differing jointly in sex and social descriptors. The reported design comprised six models, two prompt conditions, and ten repetitions per cell (480 responses). We retained the original binary reference-answer rubric and recalculated descriptive results from the reported aggregate table. Diagnostic accuracy under this rubric increased from 43/120 (35.8%) to 66/120 (55.0%), while the signed pooled accuracy gap decreased from 45.0 to 26.7 percentage points. Treatment accuracy increased from 61/120 (50.8%) to 75/120 (62.5%), while the pooled gap decreased from 25.0 to 8.3 points. Mean absolute model-level gaps decreased from 45.0 to 30.0 points for diagnosis and from 35.0 to 11.7 points for treatment. Individual models showed improvement, deterioration, or unchanged performance. Retained response collections and scoring records do not permit complete verification of all reported runs or model versions. These findings illustrate a method for examining prompt sensitivity; they do not isolate socioeconomic bias or establish a clinically effective debiasing intervention. Simple prompt-based interventions may offer a practical strategy to mitigate AI bias in epilepsy care.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Harvard Medical School, Boston, MA, USA; Beth Israel Deaconess Medical Center, Boston, MA, USA.
- Harvard Medical School, Boston, MA, USA; Beth Israel Deaconess Medical Center, Boston, MA, USA. Electronic address: daniel.goldenholz@bidmc.harvard.edu.
Förderung
- NINDS NIH HHS
- NHLBI NIH HHS
- NIA NIH HHS
- National Institutes of Health
- National Institute of Neurological Disorders and Stroke
- Amazon Web Services Inc
Zitieren
APA
Bhansali, R. M., Westover, M. B., & Goldenholz, D. M. (2027). Is it possible to vaccinate AI against bias? An exploratory study in epilepsy. Epilepsy & Behavior, 186, 111379. https://doi.org/10.1016/j.yebeh.2026.111379
Vancouver
Bhansali RM, Westover MB, Goldenholz DM. Is it possible to vaccinate AI against bias? An exploratory study in epilepsy. Epilepsy & Behavior. 2027;186:111379. doi:10.1016/j.yebeh.2026.111379
BibTeX
@article{Bhansali2027,
author = {Bhansali, Rohan Manish and Westover, M Brandon and Goldenholz, Daniel M},
title = {Is it possible to vaccinate AI against bias? An exploratory study in epilepsy},
journal = {Epilepsy & Behavior},
year = {2027},
volume = {186},
pages = {111379},
doi = {10.1016/j.yebeh.2026.111379}
}
EUVIMED