Disease-modifying AD therapies produce small effect sizes against slow-progressing endpoints. The risk of a Phase 2 failing because measurement noise across sites and time drowned the signal is real.
Our framework is a complementary scientific approach for AD-focused programs: classical interpretable features preserved for face validity, deep-learning embeddings added in parallel for sensitivity to subtle change, with harmonized measurement across languages and study duration.
Current evidence is retrospective; prospective partner validation is the next step.
The cognitive composites used as primary endpoints — PACC5, ADCOMS, CDR-SB — are by design slow-moving in the early disease stages where disease-modifying therapies are tested. Trials are long, expensive, and prone to false negatives when the underlying signal is subtle.
Most published speech-biomarker work has been optimized for discrimination — distinguishing AD from healthy controls — a fundamentally easier statistical problem than detecting small longitudinal change under treatment. Three architectural choices matter most for the latter: representation richness, multilingual consistency, and stability over study duration.
Each choice addresses a specific failure mode in long, multi-site disease-modifying AD trials.
Hand-engineered features distinguish clinical groups well, but compress out the within-subject change a treatment effect produces. We keep classical features for face validity and add deep-learning embeddings in parallel for sensitivity to longitudinal change — complementary, not a replacement.
When each language is bolted onto an English-led model, site-to-site variation adds noise to your treatment-effect estimate. Our framework uses a single shared architecture across the major European trial languages, with construct definitions harmonized across sites — so site variability is clinical heterogeneity, not measurement noise.
Over an 18-month trial, OS updates silently change the on-device noise-cancellation applied to recordings. For a pure deep-learning biomarker, this introduces drift confounded with the treatment effect. Our dual-pipeline architecture is designed to remain stable across these preprocessing changes.
See our evaluationAcross four datasets and five tasks — including retrospective decline classification on ADReSSo — the framework discriminates clinical groups and classifies decline. Prospective validation in trial-relevant populations is the next step.
Current evidence is retrospective on public and partner-curated benchmarks. Prospective validation in trial-relevant cohorts is the active next step.
Adrian Ortiz, Hanyu Peng · Cephalgo SAS
Evaluation on 44 AD/MCI cases (mean age 80.3) + 44 age-matched healthy controls (mean age 80.5). Fully interpretable speech features, traceable to specific acoustic and linguistic constructs.
Four high-value entry points for AD-focused programs.
Among amyloid-positive candidates, identify those most likely to show functional progression over the trial timeline.
High-frequency speech assessment captures trajectory information that infrequent composite-based visits cannot. The sensitivity, multilingual, and stability properties above are all most directly relevant here.
Prosodic and acoustic features track agitation, apathy, and affective state — directly relevant to AD-related neuropsychiatric symptom trials.
We do not require new data collection to begin. If your AD program has previously collected speech data, we can apply our framework to it under a research collaboration.
On legacy speech data from completed or ongoing AD trials.
In active protocols, with no impact on primary endpoint design.
On partner-curated cohorts, with shared design and reporting.
Tell us about the studies you're running and the questions you'd like speech to help answer.
Get in touch →