Screening for MCI in the Swedish H70 Birth Cohort Study using digital automatic speech biomarker tests for cognition and a machine Learning classifier

Johan Skoog2, Elisa Mallick1, Johannes Tröger1, Timothy Hadarsson Bodin2, Nicklas Linz1, Mario Mina1 & Ingmar Skoog2

1) ki:elements, 66121, Saarbrücken, Germany. 2) Institute of Neuroscience and Physiology, Sahlgrenska Academy, University of Gothenburg, 405 30, Gothenburg, Sweden.

Poster presented at AAIC 2022.

Background: Even if classic neuropsychological tests often have excellent psychometric properties to detect Mild Cognitive Impairment (MCI), they are not suitable for cost-effective low-burden screening at scale. Speech-based digital biomarkers can be deployed in a highly automated fashion. We present the results of an MCI screening algorithm based on a digital Speech Biomarker for Cognition (SB-C) in the Swedish H70 birth cohort study. 

Method: A sample from the Swedish H70 Birth Cohort study (N=404; 356 cognitively healthy (HC), 48 MCI). 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 performed (1) inferential statistics comparing MCI and HC group based on the biomarker scores and (2) built a machine learning model to screen for MCI. For (1) we performed a non-parametric Kruskal-Wallis test to compare SB-C scores of both HC and MCI groups to check for general feasibility. For (2), we trained a support vector machine model with class weights and leave-one-out cross validation to classify between MCI and HC using the SB-C scores as input (overall score and the subscores). 

Results: There was a group difference for the SB-C aggregated cognition score between the groups (HC > MCI; χ2 = 45.9 (1), p <0.001; Figure 1), and also for the subscores (Table 2). To classify between MCI and HC, using a feature selection method, the best model was found for all the five biomarker scores selected with an Area Under Curve of 0.77 (Figure 2), a specificity of 0.77 and a sensitivity of 0.76 (Table 3).

Conclusion: We found that a machine learning-based screening algorithm based on the SB-C can detect probable MCI patients in representative population sample of older people using a speech biomarker read-out.

Share this article