Privacy-Preserving AI for Mental Wellbeing Monitoring

: From Passive Sensing to Online Engagement, through a User-Centred Lens

Abstract

Mental wellbeing is a fundamental component of human health, yet traditional approaches
to monitoring it remain episodic, burdensome, and poorly suited for detecting day-to-day
fluctuations. The rise of smartphones, wearables, and online platforms enables continuous
collection of behavioural and physiological signals, ranging from activity, sleep, and location
patterns to app usage and social media interactions. Together, these multimodal data streams
provide rich opportunities for early and personalised mental health support, as they capture
both physical and digital aspects of daily life. However, realising this potential is hindered by
significant challenges, including privacy risks from centralised data processing, limited user trust,
and the difficulty of adapting AI models to heterogeneous, multimodal data. This Thesis aims
to advance privacy-friendly AI methods for mental wellbeing prediction that balance accuracy,
feasibility, and privacy. It makes four main contributions. First, it evaluates federated learning
across multiple passive sensing datasets, demonstrating that privacy-preserving training can
achieve competitive performance to centralised approaches under realistic conditions such as client
dropout and continual learning. Second, it explores user perceptions of federated learning through
qualitative studies, identifying key attitudes, trade-offs, and trust factors that shape adoption
of privacy-enhanced self-tracking technologies. Third, it presents a novel longitudinal dataset
that integrates multimodal passive sensing with social media activity, enabling more holistic
modelling of mental wellbeing. Finally, it investigates on-device large language models (LLMs)
for real-time multimodal wellbeing prediction, benchmarking prompting, retrieval-augmented
generation, and fine-tuning strategies across model scales, and showing that compact fine-tuned
models can achieve strong accuracy while remaining efficient enough to run directly on personal
devices. This makes it possible to provide responsive, real-time support while ensuring that
sensitive data never leaves the user’s device, thereby extending privacy preservation to the point
of inference. Collectively, these contributions bridge technical innovation with human-centred
design, advancing the development of mental health technologies that are private, trustworthy,
and effective in real-world contexts.

Date of Award	13 Jan 2026
Original language	English
Awarding Institution	University of Bristol
Sponsors	Engineering and Physical Sciences Research Council
Supervisor	Ryan McConville (Supervisor) & Roberta Bernardi (Supervisor)