Fair Cognitive Impairment Detection Through Unlearning

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Health & Wellbeing — Artificial Intelligence & Machine Learning, Medical Devices & Health Technology, Health & Medical Research · Depth: Expert, extended

Summary

Mild Cognitive Impairment (MCI) detection from spontaneous speech is a promising scalable screening method, but models often exhibit performance gaps across demographic subgroups due to spurious correlations. A new multimodal framework, FMD, addresses this by combining cross-model fusion (speech, text, and image) with unlearning via gradient reversal. This discourages the shared embedding from encoding task-irrelevant demographic attributes. Evaluated on the multilingual benchmarks TAUKADIAL (387 samples) and PREPARE (1644 samples), FMD achieved higher overall F1 scores than existing multilingual and multimodal baselines, with FMD^Lang reaching 92.6 F1 on TAUKADIAL and FMD^Sex reaching 60.1 F1 on PREPARE. It also substantially reduced performance gaps across sex and language subgroups, demonstrating improved generalization and robustness.

Key takeaway

For AI Scientists developing clinical diagnostic tools, you should integrate bias mitigation techniques early in your model design. FMD's approach of combining cross-modal fusion with gradient reversal unlearning demonstrates how to improve both overall accuracy and fairness across patient subgroups. Consider applying similar unlearning mechanisms to prevent models from relying on spurious demographic correlations, ensuring more equitable and robust diagnostic outcomes. This will enhance model generalizability and trust in diverse populations.

Key insights

Multimodal fusion and gradient reversal unlearning can significantly reduce demographic bias in MCI detection models while improving overall performance.

Principles

• Demographic biases in clinical AI models reduce reliability.
• Cross-modal fusion enhances feature interaction.
• Gradient reversal can unlearn task-irrelevant attributes.

Method

FMD combines unimodal encoders, a cross-attention layer for multimodal fusion, and a gradient reversal module with an auxiliary demographic classifier to unlearn demographic features from shared representations.

In practice

• Implement cross-attention for multimodal data fusion.
• Use gradient reversal for demographic bias mitigation.
• Evaluate models on worst-group performance and subgroup gaps.

Topics

• Mild Cognitive Impairment
• Bias Mitigation
• Multimodal Fusion
• Gradient Reversal Unlearning
• Speech-based Assessment
• Clinical AI

Code references

• CLU-UML/Fair-MCI-Detection

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• Fair Cognitive Impairment Detection Through Unlearning — Demographic unlearning via gradient reversal in multimodal fusion improves fair and robust MCI detection from speech.
• Can we trust AI to detect healthy multilingual English speakers among the cognitively impaired cohort in the UK? An investigation using real-world conversational speech — AI models for cognitive decline detection exhibit bias against multilingual English speakers, particularly those with specific accents.
• Fusion Learning from Dynamic Functional Connectivity: Combining the Amplitude and Phase of fMRI Signals to Identify Brain Disorders — Combining fMRI signal amplitude and phase via fusion learning improves brain disorder detection.
• Improving clinical interpretability of linear neuroimaging models through feature whitening — A novel whitening method improves neuroimaging model interpretability by decorrelating anatomically informed brain regions.
• Brain-Atlas-Guided Generative Counterfactual Attention for Explainable Cognitive Decline Diagnosis Using Multimodal Connectomes — GCAN uses counterfactual generation and brain atlas guidance for explainable, multimodal connectome-based cognitive decline diagnosis.
• Multilingual Detection of Alzheimer's Disease from Speech: A Cross-Linguistic Transfer Learning Approach — Cross-language transfer learning effectively detects Alzheimer's Disease from speech across multiple languages, achieving strong generalization and rapid inference.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.