*Poster presented at AAIC 2026

Background

REτAIN (NCT06544616) is a Phase 2b trial evaluating the efficacy, safety and immunogenicity of a phosphorylated tau-targeted active immunotherapy (JNJ-64042056), developed with AC Immune SA, in preclinical Alzheimer’s disease. While tau PET measures pathology, scalable tools for functional manifestations are needed for efficient prescreening. We evaluated whether speech-derived metrics predict elevated brain tau pathology as quantified by tau PET.

Methods

Prior to randomization, participants underwent tau PET imaging and completed a digital, speech-based cognitive assessment using the Mili platform (ki:elements GmbH). All had positive plasma pT217 levels indicating amyloid positivity, a Clinical Dementia Rating (CDR) of 0, and a Mini-Mental State Examination (MMSE) score ≥27, adjusted for education. The assessment included four Rey Auditory Verbal Learning Test repetitions and one semantic category fluency task, from which 114 speech-derived features were extracted. Participants were grouped by tau pathology status (Braak 3 ROI SUVR > 1.1). The dataset was randomly split 100 times into training (67%) and test (33%) sets. Training sets were rebalanced via minority-class oversampling and used to fit logistic regression models. Feature importance was determined through forward sequential selection with 5-fold cross-validation using ROC AUC. Rankings were averaged across partitions, and the top K features yielding the highest mean ROC AUC were selected. A final logistic regression model with these features was evaluated across all test sets to estimate generalization performance.

Results

Preliminary data from September 2025 included 136 US-based, English-speaking preclinical AD participants (Table 1). Training sets achieved mean ROC AUC >73% using 4–15 informative features, with peak performance of 73.9% (SD: 4.0%) at K = 10 (Figure 1). Importantly, performance remained stable with fewer features: on test sets, four features yielded 74.1% (SD: 6.9%) (Figure 2), comparable to 74.8% (SD: 6.8%) with ten. These findings indicate speech-based biomarkers can accurately predict tau PET status with a compact feature set, enabling scalable, non-invasive screening.

Conclusions

Speech-derived cognitive measures show promise for identifying elevated tau pathology in individuals with preclinical Alzheimer’s disease. Scalability, minimal burden, and cost efficiency of these measures justify and enable their seamless integration into multimodal screening strategies, helping to accelerate participant recruitment in clinical trials.