Fair Cognitive Impairment Detection Through Unlearning

The article introduces a multimodal framework for detecting Mild Cognitive Impairment (MCI) from spontaneous speech, integrating cross-model fusion of speech, text, and image data with an unlearning approach using gradient reversal to mitigate bias from demographic attributes. Evaluated on the TAUKADIAL and PREPARE multilingual benchmarks, the method surpasses existing state-of-the-art models in MCI classification while significantly narrowing performance disparities across different demographic subgroups. This work is crucial for practitioners as it enhances the robustness and fairness of AI models in clinical settings, ensuring more equitable health assessments.