Developing a foundation model in in-home monitoring data for healthcare applications
File(s)
Author(s)
Cui, Jin
Type
Thesis
Abstract
We propose a foundation model for in-home monitoring data to identify behavioral patterns and verify their clinical relevance in people living with dementia (PLWD). This method combines self-supervised representation learning with clinical outcome-oriented validation, focusing on agitation, urinary tract infections (UTIs), and cognitive decline. We encode multimodal patient data (sensors, activity patterns, EHRs) into structured latent representations using a pre-trained language model and a PageRank-based state transition model. Using Retrieval Augmented Generation (RAG), we synthesize privacy-preserving virtual cohorts statistically consistent with real data.
Validation indicates that this model outperforms traditional methods in predicting agitation, UTIs, and cognitive scores. Specifically, the ADAS-Cog prediction MAE was reduced to 10.46 (95% CI: 8.31-12.63), compared to 11.41 for the baseline. For UTI prediction, the model achieved 0.795 accuracy on real data and 0.906 on synthetic data. Sensitivity and PPV on synthetic data reached 0.909 and 0.957, respectively, surpassing real data metrics. The generated virtual cohort demonstrated high fidelity, achieving a Jensen-Shannon Distance of 0.2623 and Frechet Inception Distance (FID) of 0.0212. While agitation classification remained challenging (accuracy 52.3%), highlighting the need for prior knowledge, the overall two-stage coding architecture successfully extracted deep behavioral patterns. This work advances personalized dementia care by providing clinically instructive insights and secure, large-scale health monitoring through high-fidelity virtual cohort generation.
Validation indicates that this model outperforms traditional methods in predicting agitation, UTIs, and cognitive scores. Specifically, the ADAS-Cog prediction MAE was reduced to 10.46 (95% CI: 8.31-12.63), compared to 11.41 for the baseline. For UTI prediction, the model achieved 0.795 accuracy on real data and 0.906 on synthetic data. Sensitivity and PPV on synthetic data reached 0.909 and 0.957, respectively, surpassing real data metrics. The generated virtual cohort demonstrated high fidelity, achieving a Jensen-Shannon Distance of 0.2623 and Frechet Inception Distance (FID) of 0.0212. While agitation classification remained challenging (accuracy 52.3%), highlighting the need for prior knowledge, the overall two-stage coding architecture successfully extracted deep behavioral patterns. This work advances personalized dementia care by providing clinically instructive insights and secure, large-scale health monitoring through high-fidelity virtual cohort generation.
Version
Open Access
Date Issued
2025-09-15
Date Awarded
01/02/2026
License URL
Advisor
Barnaghi, Payam
Scott, Gregory
Publisher Department
Department of Brain Sciences
Publisher Institution
Imperial College London
Qualification Level
Masters
Qualification Name
Master of Philosophy (MPhil)