PhD Studentship: Computer vision to predict falls from everyday movement in older adults

Falls are the most common cause of injury-related hospital admission and death among older adults, yet fall prevention has made little measurable progress for two decades. Around a third of people aged 65 and over fall each year, and falls result in around 210,000 emergency hospital admissions a year in England. Current risk assessment relies on brief clinical tests, carried out once, in a clinic. These tests reveal little about how a person moves through an ordinary day, which is when most falls happen. Our recent research on real world falls recorded with wearable sensors has shown what goes wrong. Most falls happen indoors, during everyday actions such as standing up, sitting down, walking and turning, and 82% involve interaction with furniture or other objects.

We recently identified the common movements most often present before a fall, yet people rarely report them. In addition, in people with neurological conditions, movement becomes measurably abnormal in the minutes before a fall. We then recreated these high-risk situations in the STEPPING OUT Home Scene, a simulated home at the University's VSimulators facility, within Professor Sallie Lamb's ERC-funded programme. Older adults stand up, sit down, turn, move around furniture, open doors, change surfaces and respond to slips and trips while multi camera video, motion capture and wearable sensors record every movement. Participants' clinical and falls histories are known, fallers with non-fallers.

Data collection is ongoing, and by the start of the studentship, data from around 225 participants will have been collected. This PhD asks whether a simple camera can see what an experienced clinician sees. The aims are to: (1) derive measures of stability and movement quality from markerless pose estimation, validated against motion capture; (2) find the movement patterns and markers of instability that distinguish people who fall from those who do not, including near-falls and the moments before a loss of balance; (3) test how well these measures predict falls and how early change can be detected; and (4) make the methods work on low-cost devices in homes, care homes and hospitals, with processing on the device. 

The student will use pose estimation, deep learning for movement sequences and self supervised learning, with interpretability as a priority so that clinicians can understand and act on the results. Deliverables include a validated pipeline, open benchmarks for stability assessment from video, evidence on how well falls can be predicted from ordinary video, and a tested low-cost prototype. The studentship offers training across computer vision, data science, biomechanics and clinical falls research, with a supportive supervisory team spanning all of these areas.

Advert information