AI Improves Fair Cognitive Impairment Detection by Reducing Bias

Researchers have introduced a multimodal framework designed to improve the fairness and accuracy of Mild Cognitive Impairment (MCI) detection. This method leverages spontaneous speech data and integrates multiple data modalities, including speech, text, and images, through a cross-model fusion technique. A key innovation is the incorporation of "unlearning" via gradient reversal. This mechanism actively discourages the shared embedding layer from encoding demographic attributes that are irrelevant to MCI detection, such as sex or language. By mitigating the influence of these biased cues, the model learns more robust and generalizable representations. Evaluations on multilingual benchmarks, TAUKADIAL and PREPARE, demonstrated that this framework not only surpasses state-of-the-art baselines in MCI classification but also substantially narrows the performance gap between different patient subgroups. The study also showed that demographic unlearning contributes to more robust representations when transferring models across datasets.