Researchers say an AI that listens to how you talk - not just what you say - can detect Alzheimer's disease with 90% accuracy.
The system, detailed in a new preprint, transcribes spontaneous speech and builds three separate graph representations of it: one for meaning, one for grammatical structure, and one for how often words appear near each other. That third graph uses a statistical measure called Pointwise Mutual Information to flag when someone's narrative logic drifts from normal patterns. A gating mechanism then weighs all three views against each other depending on the patient's profile, since Alzheimer's symptoms vary widely between individuals. The team tested it on ADReSSo, a public benchmark dataset for speech-based cognitive assessment.
Most prior work in this space treats speech as a flat sequence of words, missing the structural breakdowns that often show up early in cognitive decline. By modeling language as overlapping graphs and accounting for patient-to-patient variation, this approach targets exactly the signals that simpler classifiers tend to smooth over. A non-invasive, speech-based screen that actually handles clinical diversity would be a meaningful step toward tools usable outside a specialist's office.
The code is public on GitHub, which is more than most clinical AI papers offer - though a preprint hitting a benchmark is still a long way from something a neurologist would trust in a waiting room.