Establishing computational behavioural models for the detection of early dementia from speech

Lead Research Organisation: University of Bristol
Department Name: Electrical and Electronic Engineering

Abstract

The prevalence of dementia in the UK is projected to double by 2040, trebling the current annual cost to the UK economy of £34.7 billion. Dementia is a progressive decline in cognitive function, including problems with memory, executive, and visuospatial functions. Alzheimer's Disease (AD) is the most common cause of dementia and all AD patients transition through an early phase of mild cognitive impairment (MCI), but this can be difficult to distinguish from normal cognitive ageing and is not always recognised or diagnosed. However, early diagnosis is crucial because considerable cortical damage can occur by the time major cognitive symptoms are exhibited, limiting treatment options.
Some signs of future dementia, such as gait and speech disturbances, can present years before the onset of serious cognitive decline. This presents opportunities for digital phenotyping technologies such as smartphones, wearables and in-home monitoring to help in early detection and management of AD and MCI. The SPHERE system at Bristol University is a research test bed for the health-enabled smart home of the future. It uses multiple continuous sensing platforms including video and audio, as well as accelerometer and physiological data from wearable sensors, to monitor health behaviours such as gait, navigation, activities of daily living (ADL) and social contact.
The CUBOiD project aims to combine standard clinical neuropsychological testing with mulitmodal data from SPHERE to derive ground truths on digital signals of early cognitive decline. CUBOiD is funded by the Medical Research Council and has ethical approval from the Wales REC7 committee. Despite the disruption caused by COVID-19, CUBOiD is well-suited to adaptation for online data acquisition. Indeed, an in-home mode of presentation precludes the requirement to attend clinic and could reduce patient anxiety, which can introduce variability into measured data.
The goals of this PhD research project are 1) to develop novel machine learning models that are informed by same-patient data on ADL and can detect changes in speech as diagnostic signatures of early AD and MCI; and 2) to develop novel digital interventions that may alleviate symptoms and slow disease progression. CUBOiD has already been collecting speech data using a tablet-deployed, Gogglebox-inspired talking task in which participants discuss a TV show they enjoy. Combined with standard linguistic clinical tests, this naturalistic speaking task has potential to identify how variability in speech features (e.g., prosody or linguistic compensations) may correlate with changes in ADL such as navigation or household activities in early MCI and AD. It could potentially also show whether subtle changes in speech precede changes in ADL, how this may differ between MCI and early AD- and could form a predictive diagnostic tool for MCI or AD.
Over the summer of 2020 I will conduct a literature review of linguistic anomalies seen in MCI and AD, and of digital therapies and how they relate to linguistic stimulation. Alongside analysis of existing data in year 1, a feasibility study will examine how to push the talking task assessments into an online format using the current tablet-based platform with appropriate approvals from the Sponsor (UoB) and the HRA. Normative data will then be collected from age-sex matched controls (UoB REC approval required) and compared with performance on premorbid intelligence tests (such as NART-IQ) to establish how they relate to educational level in people without dementia. In year 2, I will collect speech task data from people with a diagnosis of MCI and AD to establish correlations with ADL and ground truths on speech patterns. Alongside writing up my thesis, in year 3, I will conduct a study to establish the feasibility of interventions that could alleviate symptoms or facilitate care based on passive sampling of speech.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S023704/1 01/04/2019 30/09/2027
2275613 Studentship EP/S023704/1 23/09/2019 22/09/2023 Daniel Kumpik