With an ageing population, developing new ways to efficiently care for people living with dementia is a top priority. Passive remote healthcare monitoring, in combination with machine learning methods, can enable people living with dementia to be more independent and remain in their familiar home environments for longer, presenting an invaluable opportunity to improve their quality of life. These methods can ensure care is targeted, resources are well utilised, and unplanned hospital admissions are reduced.

This thesis focusses on improving methods for remote healthcare monitoring by providing technical contributions to machine learning and evaluating methods on a real-world case study. The Minder study has deployed passive infrared sensors and bed mats within the homes of ~160 people living with dementia, who have collectively generated almost 90000 days of activity and nocturnal physiology data. In this thesis, we leverage the Minder study to inspire, develop, and evaluate clinically relevant machine learning systems for remote healthcare monitoring, with the goal of improving dementia care.

This work comprises of four main studies: The design and deployment of a risk prediction model for urinary tract infections, which are a significant cause of unplanned hospital admissions for people with dementia; the investigation of settings where data is generated by multiple sources which have differing data distributions or noise levels, and the implications for model training; the evaluation of large language models as knowledge bases within a prior elicitation framework, with the goal of learning predictive models with less data; and the use of language models as sentence encoders to perform representation learning, allowing for clustering, anomaly detection, vector searching, and similarity scoring of daily remote monitoring data.

We conclude by highlighting how the insights provided in this thesis translate to the improved delivery of care for people living with dementia.