Subject Independent Dementia Risk Prediction Models Using Paralinguistic and Memory Test Features with Feature Warping

Clinical and biological changes during the prodromal stages of dementia are both complicated and expensive. A biomarker for cognitive reserve exposure would be highly useful as a dementia risk predictor, but has eluded researchers. Speech, which exhibits disfluencies due to dementia, is a good candidate as it is easy to collect and non-invasive. However, previous studies have only looked at the impact of dementia on speech after diagnosis. Here we extend our previous work that showed paralinguistic features extracted from audio recordings of older participants completing the LOGOS episodic memory test can be used to discriminate between high vs low cognitive reserve, hence low vs high risk of dementia. Specifically, we use the clinically validated Lifetime of Experiences Questionnaire (LEQ) to refine our ground truth estimate of cognitive reserve instead of an abridged version. Also, we improve the generalizability of our system by using feature warping to normalize across speakers. Our k-nearest neighbours (KNN) based classifier achieved an accuracy of 84% when trained with paralinguistic features alone and 91% with paralinguistic and episodic memory features.Clinical Relevance- This establishes efficacy of using speech from older participants completing the LOGOS episodic memory test to estimate risk of dementia.