Disentangling Acoustic Cues in Alzheimer's Pathology and Perception: The Roles of Language and Gender

AI models for detecting Alzheimer's Disease through speech perform inconsistently across demographic groups: The study found that pathology detection models failed to exceed chance performance for Greek and male speakers, while succeeding for Mandarin and female speakers—a critical failure mode that would be invisible without demographic-specific testing.

What AI "learns" to detect may differ from what human listeners actually perceive: The research revealed that diagnostic AI and human perceptual judgments aligned significantly for Mandarin and female speakers but diverged for Greek and male speakers, suggesting the acoustic cues driving AI decisions don't always match clinically relevant human observations.

Standard explainability tools can hide demographic disparities: Using SHAP interpretability analysis, the researchers found that global explanations of AI decision-making masked important differences in how the models performed across language and gender groups—a reminder that transparency tools alone don't guarantee fairness.

Language and gender significantly shape both disease markers and perception: The study demonstrates that acoustic biomarkers of Alzheimer's and how listeners perceive speech vary meaningfully by language and gender, meaning one-size-fits-all AI models may not work equitably across populations.

Population-specific auditing is essential before deploying clinical AI: The authors argue that responsible deployment of speech-based diagnostic AI requires testing performance separately across demographic groups, not just overall performance metrics, to prevent inequitable outcomes in clinical settings.