Alzheimer's
Poster
Tracking Cognitive Change in Mild Cognitive Impairment Using Repeated Speech-Based Digital Cognitive Assessments
Comparison with the Clinical Dementia Rating Scale
Johannes Tröger, Elisa Malick, Alexandra König, Alison Buick, Mathia Bähr Emrah Düzel, Björn Falkenburger, Gabor Petzold, Josef Pillar, Annika Spottke, Stefan Teipel, Michael Wagner, Jens Wiltfang, Inga Zerr, Stefanie Köhler
*Poster presented at CTAD 2026
Introduction: Digital cognitive assessments, particularly speech-based biomarkers (SB-C), offer a low-burden, scalable, and cost-effective alternative to traditional scales like the Clinical Dementia Rating (CDR) for frequent remote monitoring. By increasing assessment frequency, these tools may better detect subtle longitudinal cognitive decline in mild cognitive impairment (MCI) over 6–12 months. This study evaluated whether SB-C can capture such changes more sensitively than the CDR.
Methods: Participants were recruited from the DESCRIBE/DELCODE-AD registry, a multicenter memory clinic cohort. The longitudinal analysis included 31 individuals with mild cognitive impairment (MCI; 13 female; mean age 73.3 ± 7.4 years; education 13.8 ± 2.3 years) who completed speech-based digital cognitive (SB-C) assessments every three months over 15 months. Composite and domain-specific scores (executive function, memory, processing speed) were evaluated using Wilcoxon signed-rank tests and Cohen’s d. To account for learning effects, linear mixed-effects models were fitted using data from additional cognitively healthy participants (n = 80; 46 female; age 72.3 ± 7.1 years; education 15.0 ± 2.8 years) to establish normative trajectories. These models generated expected SB-C trajectories for the MCI group, and residuals representing deviations from expected learning effects were analyzed longitudinally.
Results: In the DESCRIBE/ DELCODE-AD cohort, SB-C processing speed declined significantly between baseline and 12 months (V = 129, p = 0.019, d = -0.72), while the global cognition score showed a moderate effect size despite not reaching statistical significance (V = 170, p = 0.13, d = -0.40). No meaningful longitudinal change was detected for the CDR total score (V = 1, p = 1.00, d = 0.07) or the CDR Sum of Boxes (V = 91.5, p = 0.48, d = 0.06). At the shorter 6-month interval in DESCRIBE/ DELCODE-AD, executive function (V = 128, p = 0.018, d = -0.34) and processing speed (V = 139, p = 0.032, d = -0.54) already demonstrated detectable decline.
Longitudinal modeling further demonstrated that individuals with MCI failed to exhibit the expected practice-related improvement observed in cognitively healthy participants. Residual analyses showed progressive divergence from normative learning trajectories over repeated assessments, indicating that absence of expected improvement itself may represent a sensitive marker of cognitive decline.
Conclusion: Speech-based digital cognitive assessment detected measurable decline in MCI over 6 and 12 months, outperforming traditional CDR measures. Executive function and processing speed were the most sensitive domains, and modeling practice effect deviations further enhanced detection. These remotely deployable biomarkers offer scalable endpoints for early Alzheimer’s and MCI trials, though interpretation must account for etiological heterogeneity, as restricting to biomarker-confirmed subsets would reduce statistical power.
