Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking

2025-06-05

AI Summary

Speaker gender introduces confounding bias in Alzheimer’s disease (AD) detection from speech transcripts, compromising model fairness and generalizability. Method: We propose an unsupervised, structured debiasing approach that identifies and masks gender-predictive parameters in pretrained language models (e.g., RoBERTa) without requiring gender labels. Our method introduces an extended confounder filter and a dual-filter mechanism, integrating gradient attribution with parameter importance scoring to achieve modular, layer-wise parameter decoupling. Contribution/Results: Evaluated across multiple cognitive impairment speech datasets, the method significantly reduces gender confounding (p < 0.01), improves cross-gender robustness by 23%, and incurs only a marginal 1.4% drop in AD classification accuracy—demonstrating a favorable trade-off among fairness, generalizability, and diagnostic performance.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Bias, Fairness & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Algorithmic accountability and transparency on the web

Abstract

Deep transformer models have been used to detect linguistic anomalies in patient transcripts for early Alzheimer's disease (AD) screening. While pre-trained neural language models (LMs) fine-tuned on AD transcripts perform well, little research has explored the effects of the gender of the speakers represented by these transcripts. This work addresses gender confounding in dementia detection and proposes two methods: the $ extit{Extended Confounding Filter}$ and the $ extit{Dual Filter}$, which isolate and ablate weights associated with gender. We evaluate these methods on dementia datasets with first-person narratives from patients with cognitive impairment and healthy controls. Our results show transformer models tend to overfit to training data distributions. Disrupting gender-related weights results in a deconfounded dementia classifier, with the trade-off of slightly reduced dementia detection performance.

Problem

• Addressing gender confounding in dementia detection models
• Isolating and ablating gender-related weights in classifiers
• Reducing overfitting in transformer-based dementia screening

Innovation

• Weight masking to mitigate gender confounding
• Extended Confounding Filter for weight isolation
• Dual Filter to ablate gender-related weights