Anahita Kyani, PhD; Federico Parisi, PhD; Dzmitry A. Kaliukhovich, PhD; Adria Mallol-Ragolta, PhD¹ Macarena Garcia Valdecasas Colell, PhD; Charmaine Demanuele, PhD; Kristin Farrell, PhD; Nicklas Linz, PhD; Johannes Tröger, PhD; Janna Herrmann, MS; Gayle Wittenberg, PhD; Arthur Simen, MD, PhD; Fiona Elwood, PhD; Tricia Thornton-Wells, PhD; Lennert Steukers, DVM, PhD

*Poster presented at AAIC 2026

Background

REτAIN (NCT06544616) is a Phase 2b trial assessing the efficacy of a phosphorylated tau-targeted immunotherapy (JNJ-64042056), developed with AC Immune SA (Lausanne, Switzerland), in individuals with preclinical Alzheimer’s disease (AD). The prescreening stage in this trial includes a speech-based cognitive (SB-C) assessment. We examined whether the SB-C assessment score can distinguish cognitively normal individuals (MMSE ≥ 27, CDR global score = 0) from those with impairment across multiple countries and explored its potential as a scalable, standardized prescreening biomarker.

Methods

Participants completed four repetitions of the Rey Auditory Verbal Learning Test and one Semantic Fluency Test in their native language using the Mili platform (ki:elements GmbH). Speech recordings of these tests were processed to generate a single composite SB-C score per participant, with lower scores indicating greater cognitive impairment. The ability of the SB-C score to distinguish between cognitively normal and impaired participants was assessed using ROC curves. The SB-C threshold that maximized balanced accuracy was chosen to distinguish between the two cognitive categories and to calculate additional performance metrics for each country individually.

Results

As of September 2025, data from 665 participants were analyzed: 540 from the USA (English-speaking), 39 from the UK, and 86 from Japan. Participant characteristics are presented in Table 1. Classification performance for distinguishing cognitively normal from impaired individuals yielded ROC AUCs of 72.2%, 75.0%, and 67.5% for participants from the USA, UK, and Japan, respectively. Optimal SB-C thresholds, along with additional performance metrics, are reported in Table 2. 

Conclusions

Classification performance was consistent across countries, supporting the generalizability of the SB-C biomarker. Variations in optimal SB-C thresholds highlight the need for language- and country-specific calibration to ensure optimal performance while using the same standardized underlying measure. The non-US language datasets presented two key challenges: small sample sizes and class imbalance. These factors underscore the need for further validation of the SB-C biomarker across languages and geographies. Our findings demonstrate that speech-based biomarkers can effectively detect cognitive impairment in preclinical AD. This supports their use as non-invasive, scalable prescreening tools that could potentially reduce trial costs and participant burden by improving prescreening efficiency and reducing unnecessary downstream assessments.