SAFEVID: Spasms Analysis using Federated Learning from Videos Across Multiple TREs

Neurological disorders affect millions of people in the UK and are a major cause of disability and mortality. Early diagnosis is particularly challenging for infants and children, where delays in recognising symptoms can significantly affect outcomes.

The SAFEVID project is developing machine learning models that analyse videos of infant movement to identify early signs of neurological conditions, such as infantile epileptic spasms and cerebral palsy. However, training reliable models requires access to diverse datasets held across different institutions.

Using federated learning, the project will enable models to learn from datasets stored within multiple Trusted Research Environments without transferring sensitive video data. The work will evaluate various federated approaches and assess their impact on model performance and generalisability.

Patient and public involvement is integral to the project. Workshops and engagement activities will gather feedback on the technology and its potential uses, helping guide responsible innovation and ensuring transparency in how sensitive video data is used.

By the end of the project, SAFEVID will:

• Evaluate federated learning across TREs hosted on different cloud platforms
• Assess model performance across distinct patient cohorts
• Compare federated learning tools within the DARE UK ecosystem
• Develop a prototype machine learning model for detecting infantile epileptic spasms
• Produce guidance for implementing federated AI research approaches

Project information

Lead organisation: University of Glasgow
Principal investigator: Edmond Shu Lim Ho
Project duration: 12 months
Project partners: Health Informatics Centre, West of Scotland Innovation Hub
Funding provided: £451,368
Primary contact email: Shu-Lim.Ho@glasgow.ac.uk