Alicia Lu

PhD student, National Centre for Healthy Ageing, Peninsula Health

Supervisors: Professor Velandai Srikanth, Dr Taya Collyer, Associate Professor Chris Moran, Associate Professor Richard Beare

Alicia Lu is using thousands of real Australian patients’ hospital records to train an artificial intelligence model to recognise the early warning signs of dementia.

John is 90, and lives alone. One day, he is brought to hospital after a car accident, where ends up being diagnosed with dementia. He had been driving aimlessly for hours, unable to remember where he was going or how to get home.

John spends months in a locked dementia ward, waiting for someone to decide where he goes, because it’s no longer safe for him to live alone. His story isn’t unique: as a geriatric medicine specialist, I see this far too often. Across the world, three-quarters of people with dementia remain undiagnosed, and many often present to hospital in crisis.

Dementia doesn’t develop overnight. In the years before John’s accident, clues were already emerging in his medical records, scattered across hospital visits with no one joining the dots: a near-miss with medication, repeated falls because he could never remember to use his walking stick.

Every hospital visit leaves behind a trail of data. Some of this is structured and organised, like blood tests. But some of the most important clues are buried in clinical notes. For example, years earlier, a nurse might have written, “Patient seemed confused about what year it was”. These clues have sat inside medical records for a long time, but have been too difficult to extract and analyse at scale.

These clues are often missed by clinicians who are managing acute conditions in a busy hospital environment. Dementia recognition currently hinges on someone recognising the symptoms and flagging them with a doctor to start the diagnostic pathway. It’s not uncommon to see people brought into hospital by police or ambulance after being found wandering and unable to get home, or by distant family members who have only just discovered the extent of their decline after finding unlivable conditions at home.

Earlier recognition means earlier support, and fewer prolonged admissions that put huge pressure on our hospital systems. This allows time for accessing community supports through My Aged Care or private services, completing legal paperwork such as wills and powers of attorney, and having conversations about future care, downsizing, or residential aged care. There are also some emerging interventions that may be able to slow cognitive decline, such as lifestyle modifications or the new monoclonal antibodies for Alzheimer's disease.

My model combines machine learning techniques on structured hospital data, with natural language processing methods (such as LLMs) on clinical notes, to see if we can generate individualised dementia risk scores for older people coming through hospitals..

Even the best model is useless if people don't use it. There is a huge implementation gap for many AI tools in health care, as the governance, workflows, and clinical processes needed to support them have not been fully established. That’s why I am also speaking with doctors, people with dementia, and carers, to understand what it takes for them to trust a tool like this. While they can see the benefits of a tool like this, hospital clinicians strongly emphasise that a dementia risk score is of limited use if there aren't robust pathways and processes around who follows it up, who relays it to the affected patient/individual, and how reliable the tool is.

To develop my models, I’m using datasets from the NCHA Healthy Ageing Data Platform. It is an incredible asset to have in Victoria: the alternatives would have been to use cohort study data, which don't necessarily reflect real-world clinical practice or the data that are actually available to clinicians, or to use overseas electronic health record datasets from countries like the USA or UK, which raises questions about how well those models translate to Australian healthcare. The other option would have been to manually collect or extract data, but that would have dramatically limited both the sample size and scale of the project.

I think that's probably one of the greatest lessons a supervisor can give a PhD student, not just helping them complete a project, but helping them become an independent researcher who is ready for whatever comes next after thesis completion

My primary supervisor, Srikanth, has been extremely supportive and encouraging of my PhD progress, striking a really good balance between giving me the autonomy to pursue my own ideas while providing guidance whenever I need it. That's probably one of the greatest lessons a supervisor can give a PhD student, not just helping them complete a project, but helping them become an independent researcher who is ready for whatever comes next after thesis completion.

I love that geriatric medicine combines immense medical complexity with a genuinely person-centred approach. Our patients often have multiple interacting conditions, but despite that complexity, every decision comes back to the person's goals, values, and quality of life. That philosophy of care, together with the people within the specialty (both the staff and the wonderful older patients), were key reasons why I decided to pursue geriatric medicine.