Zampeta-Sofia Alexopoulou, Louisa Schwed, Johannes Tröger, Nicklas Linz, Qingyue Li, Stefanie Köhler, Josef Priller, Annika Spottke, Gabor C. Petzold, Jens Wiltfang, Frank Jessen, Anja Schneider, Emrah Düzel, Michael Wagner, Christoph Laske, Frederic Brosseron, Stefan J. Teipel, Valeria Manera, Alexandra König
*Poster presented at AAIC 2026
Background: Cerebrospinal fluid (CSF) biomarkers provide evidence of Alzheimer’s disease (AD) pathology, yet their relationship to cognitive change captured by digital tools remains unclear. Speech-based cognitive measures are sensitive to early impairment; whether they capture distinct longitudinal trajectories in biomarker-positive versus biomarker-negative individuals is unknown. Here, we examined whether a speech-based global cognitive composite (SB-C) and domain-specific scores (memory/processing speed/executive function) capture longitudinal change as a function of CSF biomarker status.
Methods: Within PROSPECT-AD, data were drawn from the German DELCODE/DESCRIBE cohorts with available CSF biomarkers (Aβ42, Aβ42/40, P-tau181, T-tau, Aβ42/P-tau181) (N = 12 healthy, N = 51 subjective cognitive decline, N = 33 mild cognitive impairment (MCI)). Participants completed an automated remote phone-based speech assessment every three months over 15 months. SB-C scores were derived using a validated pipeline (ki:elements). CSF was collected prior to speech; all biomarker-positive participants were included, while biomarker-negative participants were restricted to a CSF-to-baseline-speech gap ≤2.5 years. Biomarker positivity was determined using established cut-offs. Longitudinal SB-C trajectories were examined using linear mixed-effects models per biomarker with time and biomarker status as primary predictors. Models included age, sex, education, baseline diagnosis, CSF-to-speech time-gap and participant-specific random effects. Estimated slopes and least-squares mean changes from baseline were derived.
Results: Across biomarkers, baseline SB-C didn’t differ between biomarker-positive vs
biomarker-negative participants. MCI diagnosis was consistently associated with lower SB-C (p<0.001). Longitudinally, significant time effects were observed for all models. Overall, biomarker-negative participants significantly improved over time, compared to no significant change in biomarker-positive ones (Figs. 1-3). SB-C processing speed consistently demonstrated significant improvement in biomarker-negative groups. Analysis on Aβ42/40 yielded marginal effects.
Conclusion: Despite modest sample size and temporal separation between CSF collection and speech, SB-C captured longitudinal patterns by CSF biomarker status. Improvement confined to biomarker-negative individuals, potentially reflecting preserved learning/practice-related gains. Findings highlight the potential of speech-based markers in complementing biomarker-informed stratification and improving longitudinal monitoring.