Fair Cognitive Impairment Detection Through Unlearning
Researchers have developed a new multimodal framework for detecting Mild Cognitive Impairment (MCI) from speech, aiming to reduce performance disparities across demographic subgroups. The system employs cross-model fusion of speech, text, and image data, combined with gradient reversal unlearning to prevent demographic attributes from influencing the shared embedding. Tested on the TAUKADIAL and PREPARE benchmarks, this method not only surpasses existing baselines in MCI classification but also significantly narrows the performance gap between different patient groups, such as by sex and language. AI
IMPACT This research could lead to more equitable AI tools for medical diagnosis, reducing bias in healthcare applications.