Vijay Vankadaru

Vijay Vankadaru

MS Data Science, UC Berkeley
Mechanistic interpretability of LLMs and VLMs
Machine Learning Engineer, DASION

Potential for Age-Based Bias in AI-Informed Neurosurgical Decision-Making: A Sociodemographic Subgroup Assessment Approach for Open-Source Glioma Molecular Subtype Classifiers

In reviewUCSF, with the Departments of Neurological Surgery and Radiology

E. S. Danthinne, V. Vankadaru, H. Miao, P. Nedelec, A. Rauschecker, R. Morshed, A. Molinaro, M. Sushil

With Elisa Danthinne, Hongbin Miao, Pierre Nedelec, Andreas Rauschecker, Ramin Morshed, Annette Molinaro, and Dr. Madhumita Sushil at UCSF.

We externally validated three published, open-weight models that predict glioma molecular subtype (IDH mutation, 1p/19q codeletion, grade) from preoperative MRI on an unseen institutional cohort. To get subgroup sample sizes large enough to test anything, we used an LLM to label 9,377 pathology reports (majority vote over three runs, F1 0.96 to 0.97 against human labels), expanding the labeled cohort from 442 to 3,380 patients. All three models underperformed their published numbers, and errors concentrated by age: specificity for IDH-wildtype patients under 40 was 54%, against 80% for patients 55 and older.

Cite as: Danthinne, E. S., Vankadaru, V., et al. Potential for Age-Based Bias in AI-Informed Neurosurgical Decision-Making. In review.