Mobile Health Diagnosis

Screening, monitoring, and progression forecasting on phones and wearables

This line of work targets diagnosis and longitudinal health monitoring in ubiquitous settings: screening, risk stratification, and forecasting of disease course from data collected outside the clinic. We combine mobile and wearable sensing with machine learning so that timely signals can support early intervention. Audio is one important modality—for example cough, breath, and voice—for respiratory, affect-related, and mental health contexts, together with other physiological or behavioural cues where appropriate. The emphasis is on end-to-end clinical and public-health questions in mobile health, rather than generic modelling of structured time series or tabular records; those methodological themes are developed separately under structured health data modelling.

Project members: Hongyu Jin, Ye Bai

Speech and Voice Intelligence

AI for speech perception, generation, and adaptation in conversational systems

This project focuses on building speech and voice intelligence across the full pipeline of perception and generation: automatic speech recognition, text-to-speech, speaker and style adaptation, and conversational interaction. We are particularly interested in modelling ambiguity and uncertainty in speech signals, and in developing robust, efficient, and trustworthy speech foundation models that can generalise across domains, speakers, and affective states. By combining speech processing with large language models and multimodal signals, we aim to enable next-generation conversational AI systems that listen, speak, and adapt in a human-centred way for health and everyday applications.

Project members: Yang Xiao, Siyi Wang, Jiaheng Dong

Structured Data Modelling for Health

Learning from structured health data with a focus on time series

Many health data sources, such as ECGs, EEGs, acoustic signals, electronic health records, and derived clinical features, are naturally structured and often longitudinal. This project investigates representation learning for such structured health data, with a particular emphasis on health time series. We study self-supervised and test-time adaptation methods that can handle distribution shifts (e.g., missingness, motion artefacts, population differences) and aim to build models that are robust, data-efficient, and clinically meaningful. The goal is to improve downstream tasks such as risk prediction, disease trajectory modelling, and personalised decision support.

Project members: Jie Huang

Earable Sensing

Multimodal sensing for continuous monitoring of vital signs

This project aims to explore the potential of earable devices for continuous monitoring of vital signs. By leveraging multimodal sensing technologies integrated into earable devices, we seek to develop non-invasive methods for tracking various physiological parameters. The research focuses on overcoming challenges related to signal quality, power efficiency, and user comfort, while ensuring accurate and reliable health monitoring in everyday settings.

Project members: You Zuo