Speech-Based Multimodal LLMs for Cognitive Impairment Detection — Targeting Generalization Across Populations

arXiv 2607.21496·low signal

Yingchao Huang, Xin Wang and Yuhan Su apply speech-based multimodal LLMs to cognitive impairment detection, explicitly targeting generalizable performance rather than dataset-specific accuracy. Speech biomarkers for cognitive decline have historically failed to transfer across languages, recording conditions, and demographics — the generalization framing is the contribution. For health-AI builders, the transferable lesson is that in-distribution accuracy on a single clinical corpus predicts almost nothing about deployment.