Elisa Mallick, Fredrik Öhman, Nicklas Linz, Alexandra König, Michael Schöll, Silke Kern, Johannes Tröger & Ingmar Skoog
* Poster presented at the 17th Clinical Trials on Alzheimer’s Disease (CTAD), Madrid (Spain)
Abstract
Background: The Clinical Dementia Rating Scale (CDR) is arguably the most widely used instrument to detect and stage dementia, it is one of the most-established clinical trial endpoints in Alzheimer’s and is routinely used to characterize trial populations. However, it is not ideal for a clinical trial cost-effective low-burden general population screening funnel. Speech-based digital measures on the other hand can be deployed in a highly automated and scaled fashion. Therefore, we present the results of an CDR stage screening based on a digital Speech Biomarker for Cognition (SB-C) in the Swedish H70 birth cohort study.
Method: This research is based on a sample from the Swedish H70 Birth Cohort study (N=695; 542 CDR Total score = 0, 153 CDR Total score = 0.5) (Table 1). We automatically extract the SB-C score and its subscores (executive function, memory, semantic memory, processing speed) from SVF and RAVLT speech recordings using ki:elements’ proprietary speech analysis pipeline including automatic speech recognition and feature extraction. We weighted the SB-C subscores for CDR sensitivity and performed (1) inferential statistics comparing CDR groups based on the SB-C Weighted Score and (2) determined a cutoff to differentiate between CDR groups (CDR = 0 vs. CDR = 0.5). For (1) we performed a non-parametric Kruskal-Wallis test to compare SB-C Weighted Score of both CDR = 0 and CDR = 0.5 groups to check for general feasibility. For (2), we applied an optimal cut-off procedure aimed at maximizing the sum of sensitivity and specificity in discriminating between CDR groups.
Results: The Kruskal-Wallis test revealed a significant difference of the SB-C Weighted Score between the CDR groups (CDR = 0 > CDR = 0.5; χ2 = 87.68 (1), p <0.001; Figure 1). The optimal cutoff analysis showed that a cutoff of -0.53 in the SB-C Weighted Score differentiated between CDR groups with a balanced accuracy and a ROC-AUC score of 0.71 (Figure 2). Sensitivity and specificity of the classification were both 0.71 (Table 3).
Conclusion: We found that a threshold on the SB-C Weighted Score can separate CDR stages 0 and 0.5 in a representative general population of older people using a speech-based automatic read-out. This method therefore represents a promising alternative for trial population characterization especially in general population screening funnels for preclinical AD trials.
