Noise-cancellation variability as a source of longitudinal drift in speech-based digital endpoints for Alzheimer's disease trials
Year: 2026
Type: conference-abstract
Abstract: Poster P446, Clinical Trials on Alzheimer's Disease (CTAD 2026), Boston, MA, 16-19 November 2026. An audio processing update on a participant's phone can change speech recordings enough to look like cognitive decline or improvement. In Cephalgo's own checks, a minor iOS release changed audio processing in measurement mode on both iPhones checked, and Android and web browser updates also changed audio processing settings. Findings. Three published speech datasets were used. In the hardest comparison, Alzheimer's disease versus MCI in one dataset, one simulated audio processing update moved 32% of people to the other group with standard handcrafted speech features, and 27% with Cephalgo's Multi-stream model, although their condition had not changed. Multi-stream, which combines linguistic, acoustic and self-supervised analysis, stayed more stable across recording conditions than standard handcrafted features; after the worst audio processing update, its sensitivity and specificity were 17 to 19 percentage points higher. In simulated year-long trials with a known decline added, one audio processing update made speech timing features decline up to 39% faster or slower than the true decline. Recalibration from the audio removed 95% to 97% of the false trend, so an audio processing update no longer looked like decline or improvement. For sponsors running speech endpoints. Record the mobile device model and software version with every recording. Ask the vendor how its speech measures behave when audio processing changes. Test measures on the same audio processed in different ways. Before unblinding, state in the analysis plan how readings from before and after a change are compared. Full poster on request: [email protected]. More at https://www.cephalgo.com
Language: English
Subfield: Psychiatry and Mental health
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7235076567
