Conversational robots to support well-being and home safety in dementia care
File(s)
Author(s)
Raposo de Lima, Maria
Type
Thesis
Abstract
As the global population ages and dementia prevalence increases, there is a pressing need for accessible and scalable technologies to support well-being and safety at home. Timely detection of cognitive impairment remains a major unmet need. Conversational robots hold potential for home assistance through verbal interactions. However, real-world deployments face challenges in long-term engagement, personalised care, and clinical utility. This dissertation designed and evaluated conversational robots to support well-being, enhance home safety, and monitor cognitive health. In a year-long deployment of conversational technology in homes of people affected by dementia (N=14), this work demonstrated the feasibility of mapping behaviour by analysing activity sequences, highlighting the potential of conversational AI to provide proactive verbal support in response to safety risks. This research introduced a machine learning approach and evaluated spoken language as a biomarker for automated assessment of cognitive health. In screening cognitive impairment, the best-performing model, trained on 100 interpretable linguistic features extracted from a benchmark DementiaBank dataset (N=291), achieved a mean sensitivity of 69.4% and specificity of 83.3%. On pilot data collected in-residence (N=22), this model achieved a mean sensitivity of 70% and specificity of 52.5%. In predicting the severity of cognitive decline via Mini-Mental State Examination (MMSE) scores, results showed a mean absolute MMSE error of 3.7 and comparable performance of 3.3 on real-world data, demonstrating generalisability. Risk stratification and feature importance analysis enhanced the clinical utility of model predictions and enabled the identification of higher-risk individuals for timely interventions. This dissertation involved stakeholders with lived experience of ageing and dementia (N=111) in real-world evaluation studies conducted in the UK, India, and the USA, assessing technology acceptance, user engagement, and feasibility of home use. Additionally, this work designed a large language model-powered socially assistive robot and validated a new approach to in-residence cognitive support through human-robot verbal dialogue.
Version
Open Access
Date Issued
2024-10-03
Date Awarded
01/02/2025
License URL
Advisor
Rodriguez y Baena, Ferdinando Maria
Vaidyanathan, Ravi
Publisher Department
Department of Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)