Enhanced Machine Learning Classification of Cognitive Impairment with Multimodal Digital Biomarkers

Background Linus Health’s DCTclock™ is an FDA‐listed cognitive assessment solution capable of detecting early signs of cognitive decline. DCTclock digitizes the long‐standing pen‐and‐paper Clock Drawing Test (CDT) and captures a wide variety of cognitive and motor metrics during the entire drawing process. Multiple studies have validated the DCTclock’s ability to detect early cognitive impairment and improve on existing assessment methods (e.g., Souillard‐Mandar et al. 2015 & 2021, Rentz et al. 2021). Linus subsequently developed the Digital Clock and Recall (DCR™) assessment, which incorporates voice and delayed verbal recall alongside the DCTclock. The analysis below evaluated the added benefit of the DCR and the relationship between voice analytics and diagnosis. Method The DCR is being administered as part of the BioHermes‐Gap study. Data from 495 participants were used to correlate DCR or DCT with MMSE. A repeated‐measures ANOVA model was used to analyze speech and determine within‐subject differences between sessions (immediate repeat and delayed recall) on 40+ acoustic features extracted from 3‐word recall tasks of 664 older adults. Results were used to build Multilayer Perceptron Neural Networks (ANN) to assess the accuracy of combined acoustic features to classify cognitive status. Result DCR is more highly correlated with MMSE than the DCTclock (r=0.43 vs. 0.38). ANN results indicate that combining acoustic features with cognitive assessments achieves greater classification accuracy for healthy (AUC=0.95), MCI (0.93), and AD (0.97) groups than models including either acoustic features or cognitive assessments alone. Conclusion A multimodal machine learning investigation of cognitive state, DCR, which includes DCTclock and voice/speech features from delayed recall, increases the accuracy (sensitivity and specificity) of cognitive status classification.