Cross-Lingual Alzheimer's Disease Speech Detection: Polarity Inversion and Few-Shot Calibration Strategies

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Medical Informatics

Background:

• Speech-based screening offers a cost-effective method for early Alzheimer's disease (AD) detection.
• Deep learning models for AD detection are limited by scarce labeled speech data in low-resource languages, hindering cross-lingual transfer learning.

Purpose of the Study:

• To investigate the failure mechanism of conventional cross-lingual transfer learning in pathologic speech analysis.
• To develop a novel framework, Monte Carlo Polarity Flip Calibration (MC-PFC), for data-efficient cross-lingual medical analysis.

Main Methods:

• Analysis of disease-associated representation vectors in a self-supervised HuBERT space to identify cross-lingual polarity flips.
• Development and application of Monte Carlo Polarity Flip Calibration (MC-PFC), a few-shot framework utilizing separability-weighted ensemble voting for direction flip estimation.
• Evaluation on a held-out Chinese blind test set.

Main Results:

• A systematic mechanism, termed cross-lingual polarity flip, was identified as the cause of performance degradation in cross-lingual pathologic speech.
• MC-PFC achieved an AUC of 0.871 on a Chinese test set, closely matching the in-domain upper bound (AUC = 0.875).
• Direction calibration via MC-PFC yielded a significant AUC gain (+0.361) compared to standard distribution alignment (+0.081).

Conclusions:

• Cross-lingual polarity flips are a stable, structural phenomenon in discriminative dimensions, not merely ungeneralizable noise.
• MC-PFC establishes a data-efficient paradigm for cross-lingual medical analysis by actively modeling and calibrating cross-lingual discrepancies.
• This approach significantly enhances the clinical utility of AI for early Alzheimer's disease detection in diverse linguistic populations.