This is a project done while interning at Winterlight Labs. We identify the problem of age being confounded into dementia detection with linguistic features, propose to use fair representation learning to address it, and propose to evaluate with a modified equalized odd score.
On two datasets, DementiaBank and Famous People, our best methods outperform traditional statistical adjustments (residualization and inverse probability weighting), and are comparable to the theoretical upper bound.
Presentation slides
Full paper is available at arxiv.
Media coverage by Medical XPress.
Citation
@article{age-indep,
archivePrefix = {arXiv},
arxivId = {1807.07217},
author = {Zhu, Zining and Novikova, Jekaterina and Rudzicz, Frank},
eprint = {1807.07217},
title = { {Deconfounding age effects with fair representation learning when assessing dementia} },
url = {https://arxiv.org/pdf/1807.07217.pdf},
year = {2018}
}