Bias and Fairness in Self-Supervised Acoustic Representations for Cognitive Impairment Detection

A research team from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) has published a systematic analysis revealing significant fairness issues in AI models designed to detect cognitive impairment (CI) and depression from speech patterns. The study, titled 'Bias and Fairness in Self-Supervised Acoustic Representations for Cognitive Impairment Detection,' compared traditional acoustic features like MFCCs and eGeMAPS with contextualized embeddings from the self-supervised model Wav2Vec 2.0 (W2V2). Using the clinical DementiaBank Pitt Corpus, the researchers found that while higher-layer W2V2 embeddings outperformed baselines for CI detection, they exhibited troubling performance disparities across demographic subgroups, raising critical questions about deploying such tools in real-world clinical settings.

The technical analysis showed that for CI detection, W2V2 achieved an Unweighted Average Recall (UAR) of up to 80.6%. However, the Area Under the Curve (AUC) was significantly lower for female participants (0.769) and younger participants (0.746) compared to their counterparts. The specificity gap (Δ_spec) reached up to 18% for females and 15% for younger individuals, meaning the model was more likely to incorrectly label healthy people in these groups as impaired. Furthermore, the study found limited cross-task generalization between CI and depression classification, suggesting each condition relies on distinct acoustic markers. These findings underscore that raw performance metrics are insufficient and mandate rigorous, fairness-aware evaluation with subgroup-specific analysis before clinical adoption, especially given the demographic heterogeneity of patient populations.