Abstract
This study examines the presence of gender bias in the human perception of Alzheimer’s Disease (AD) based on vocal cues. By analyzing speech samples from Mandarin and Greek speakers, researchers identified a significant tendency for listeners to label male speakers as having AD more frequently than female speakers. This bias is particularly pronounced in the listeners' native language and is driven by specific acoustic features like speech portion and shimmer. These findings suggest that human-labeled datasets used to train diagnostic AI may contain inherent biases that could lead to clinical inequities.
Introduction
The development of non-invasive screening tools for Alzheimer’s Disease (AD) increasingly relies on automated speech analysis. However, speech is a highly personalized and gendered medium. While biological differences in AD prevalence are documented, the role of perceptual bias—how stereotypes influence the interpretation of vocal patterns—remains under-explored. If clinical labels or training data for AI are influenced by gendered expectations (e.g., interpreting male vocal instability as pathology while dismissing similar traits in females), the resulting diagnostic tools may be fundamentally flawed. This study investigates whether such biases exist across different linguistic contexts and identifies the acoustic drivers of these perceptions.
Methodology
The researchers conducted a perception experiment involving 16 native Mandarin-speaking listeners (8 male, 8 female). The stimuli consisted of 60 audio clips: 30 in Mandarin and 30 in Greek.
Tasks and Datasets
The clips were derived from standard clinical "picture description" tasks:
- Mandarin Data: Native speakers describing the "Cookie Theft" scene, sourced from the NCMMSC2021 AD Recognition Challenge.
- Greek Data: Speakers describing a lion cub scene, sourced from the ADReSS-M challenge (ICASSP 2023).
Analysis
Listeners classified each speaker as either AD or Healthy Control (HC). The research team extracted three primary acoustic features:
- Speech Portion: Ratio of speaking time to total duration (detecting long silences).
- Jitter: Pitch instability (frequency variation).
- Shimmer: Loudness instability (amplitude variation).
A Generalized Linear Mixed-Effects Model (GLMER) was employed to analyze the data, accounting for fixed effects (gender, language, acoustics) and random effects (listener variability).
Results
The analysis revealed a clear and significant gender bias in AD perception.
As shown in the score distributions, male speakers were significantly more likely to be identified as having AD across the combined dataset (p = 0.035, Odds Ratio = 1.41). This effect was even more pronounced in the Mandarin-only subset, where the Odds Ratio for males being perceived as AD rose to 2.10.
The GLMER results (Table 1) highlight the specific acoustic drivers of these labels. A lower "Speech Portion" (increased silence) was the most robust predictor of an AD label (p < 0.001). Additionally, "Shimmer" (loudness instability) was significantly associated with AD perception (p = 0.020), whereas "Jitter" (pitch instability) did not reach statistical significance.
Discussion
The findings suggest a "Rationality Bias" where societal expectations of male speech as stable and rational make deviations (like memory-related pauses or vocal shakiness) more salient to listeners. Interestingly, this bias was stronger in the listeners' native language (Mandarin) than in Greek. This suggests that semantic comprehension amplifies social biases; when listeners understand the content, they are more likely to apply gendered cognitive stereotypes.
Furthermore, the significance of "shimmer" over "jitter" indicates that specific types of vocal prosody are culturally linked to perceptions of cognitive decline. The lack of significant difference in AD labeling likelihood between Greek and Mandarin suggests that while the accuracy of detection might vary by language familiarity, the underlying bias remains consistent.
Conclusion
This study demonstrates that human perception of AD is not gender-neutral. The tendency to over-diagnose males based on speech patterns poses a significant risk for the development of AI-driven diagnostic tools. If "ground truth" labels are biased, machine learning models will inevitably learn and scale these inequities. Future research must expand to include more diverse listener demographics and native speakers of multiple languages to further isolate cultural variables and develop robust de-biasing strategies for medical AI.
Summary updated