The Voice of Equity: A Systematic Evaluation of Bias Mitigation Techniques for Speech-Based Cognitive Impairment Detection Across Architectures and Demographics

High-performing AI systems can still harbor significant fairness problems: Both models achieved strong overall accuracy (F1 scores around 71%), yet showed substantial disparities across demographic groups—demonstrating that traditional performance metrics alone are insufficient measures of responsible AI in healthcare applications.

Older adults and non-English speakers face systematic disadvantages: Adults aged 80 and older showed notably lower sensitivity (46-50%) compared to younger groups, and Spanish speakers demonstrated reduced detection rates versus English speakers, indicating that bias mitigation cannot be one-size-fits-all.

Different AI architectures respond differently to bias fixes: The same mitigation techniques (like oversampling) worked effectively for one model architecture but not the other, suggesting that responsible AI requires testing fairness interventions across multiple system designs rather than assuming solutions are universally applicable.

Multiple bias mitigation approaches should be compared, not just one: This study uniquely evaluated pre-processing, in-processing, and post-processing techniques together, revealing that organizations need comprehensive testing strategies rather than implementing a single "fix" and assuming fairness is achieved.

Architectural design choices fundamentally shape who benefits and who is harmed: The research demonstrates that how engineers build AI systems—not just the data used—determines fairness outcomes, emphasizing that responsible AI requires intentional design decisions from the ground up, particularly for healthcare screening tools affecting vulnerable populations.