Alzheimer's

Poster

Ensuring Demographic Equity: Evaluating the SB-C Speech-Based Cognitive Assessment Across Ethnic Groups in US

Johannes Tröger, Louisa Schwed, Simona Schäfer, Alison Buick, Nicklas Linz, Alexandra König

*Poster presented at CTAD 2026

Introduction: To meet regulatory emphasis on diverse clinical trial populations, speech-based cognitive assessments using automatic speech recognition (ASR) must ensure equitable demographic performance. Systematic ASR errors in underrepresented groups can bias cognitive performance estimates and compromise digital endpoints. This study evaluated ASR performance differences across ethnic groups among U.S. participants completing audio tasks in the ki:elements Speech-Based Cognitive Assessment (SB-C).

Methods: 99 participants with 25 self-identifying as African American (16 female; age 56.96±5.24 years; education 15.48±2.62 years), 29 Hispanic American (14 female; age 47.07±8.41 years; education 15.72±2.12 years), and 45 White American (23 female; age 55.67±5.03 years; education 15.47±2.22 years) were recruited in the United States through the Prolific online platform and completed remote assessments utilizing a Bring Your Own Device (BYOD) model. The SB-C leverages two speech-based tasks: semantic verbal fluency (SVF) and verbal learning (RAVLT). Human-rater transcripts served as ground truth and were compared with transcripts generated by ki:elements’ proprietary speech processing pipeline SIGMA. Word Error Rate (WER) was calculated for each task and overall SB-C scores. Pairwise differences in WER and subsequent SB-C scores between ethnic groups were evaluated using Mann–Whitney U tests.

Results: Mean WERs for the SVF task were 0.33%±1.25, 0.39%±1.65, and 2.41%±10.45 for African American, Hispanic American, and White American participants, respectively. Mean WERs for the VLT task were 1.74%±3.30, 1.15%±2.45, and 0.73%±2.12, respectively. No statistically significant differences in WER were observed between ethnic groups for either task. As a consequence also the SB-C values calculated from the ASR-based speech processing did not show any significant differences between all three groups.

Conclusion: ki:elements’ speech processing and ASR of the SB-C demonstrated low transcription error rates across ethnic groups, with no evidence of systematic performance degradation in African American or Hispanic American participants. These findings suggest that ASR transcription accuracy is unlikely to contribute to systematic underestimation of cognitive performance in underrepresented populations and support the use of ASR-enabled speech-based cognitive endpoints in increasingly diverse clinical trial populations. The SB-C assessment, powered by the SIGMA pipeline, offers a scientifically validated, equitable digital measure capable of supporting inclusive and decentralized global clinical trials without introducing demographic bias.

Related Articles