Leveraging Explainable AI to Identify Novel Risk Factors in Dementia

The World Health Organization (WHO) declared dementia a global health priority in 2017. ¹ In Aotearoa New Zealand (NZ), the number of people with dementia is projected to double from 83,000 people in 2025 to almost 170,000 people by 2050, placing immense pressure on healthcare systems, families, and communities. This burden is not evenly distributed. Māori, Pacific peoples, and socioeconomically disadvantaged populations face higher risks due to intersecting health and social inequities. ² 

In 2020, the population attributable fraction (PAF) for 12 potentially modifiable risk factors was estimated at 40% globally. These factors were hypertension, obesity, diabetes, traumatic brain injury (TBI), excessive alcohol consumption, smoking, physical inactivity, air pollution, depression, hearing loss, social isolation, and less education.³ The NZ total dementia prevention potential was predicted to be 47.7% for the total population and over 50% for its Māori and Pacific populations, due to the higher prevalence of risk factors in these groups. ⁴ 

While complete risk factor elimination is not realistic, even modest reductions in population risk factor prevalence can meaningfully reduce dementia prevalence.

Prof Simpson and his team recently received funding for their research using explainable AI (XAI) to explore modifiable risk and protective factors for dementia, with a particular focus on individuals living with diabetes. The project aims to uncover insights that traditional statistical methods may overlook, while ensuring transparency and trust in the findings.

The team will leverage the Stats NZ Integrated Data Infrastructure (IDI), to combine comprehensive health and socioeconomic data to investigate dementia risk.

The Population Cohort Demographics Table serves as the backbone for linking multiple datasets via the National Health Index (NHI), including:

·       Hospital discharges

·       Pharmaceutical dispensing

·       Chronic condition registries

·       Cancer registrations

·       Maternity records

·       Mortality data

·       Mental health services

·       Non-admitted patient collections

Socioeconomic data from the 2023 Census, along with historical datasets on income, employment, education, housing, and benefits, will provide context for understanding the broader social determinants of dementia.

A tiered modelling strategy will be adopted to balance interpretability and predictive accuracy. Logistic regression will offer a baseline, interpretable assessment of individual risk factors such as age, BMI, and family history. Decision trees and related methods will enable visualisation of interactions between risk factors, revealing simple predictive patterns. More complex machine learning models will be trained to capture non-linear and higher-order relationships, with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) used to provide both global and local interpretability. Structural causal models (SCM) will support causal inference using directed acyclic graphs and instrumental variable regression. To support secure, privacy-preserving, and collaborative research, synthetic datasets based on the most important dementia risk factors will be generated, enabling reproducibility while maintaining confidentiality.

The project aims to provide evidence to inform policy interventions that improve dementia prevention, equitable healthcare delivery and enhance quality of life for people in Aotearoa. By applying XAI to complex, multifactorial conditions like dementia, the project highlights the potential of transparent, data-driven approaches to advance public health research and drive meaningful change.

Research Team: 

Professor Colin Simpson, Professor Ivy Liu, Mr Andrew Sporle, Associate Professor Yoram Barak, Dr Aliitasi Su'a-Tavila, Dr James Mbinta, Dr Alex Wang, Dr Richard Haarburger, Dr Greg Martin. 

References

[1] World Health Organization. "Global action plan on the public health response to dementia 2017–2025." Global action plan on the public health response to dementia 2017–2025. 2017.

[2] Ma'u, E., et al. Dementia economic impact report 2020/prepared for Alzheimers New Zealand: authors: Ma’u E, Cullum S, Yates S, Te Ao B, Cheung G, Burholt V, Dudley M, Krishnamurthi R, Kerse N. Wellington: [Alzheimers New Zealand], 2021., 2021.

[3] Stephan BCM, Cochrane L, Kafadar AH, et al. Population attributable fractions of modifiable risk factors for dementia: a systematic review and meta-analysis. Lancet Healthy Longev. 2024; 5(6): e406-e421

[4] Ma'u E, Cullum S, Cheung G, Livingston G, Mukadam N. Differences in the potential for dementia prevention between major ethnic groups within one country: a cross-sectional analysis of population attributable fraction of potentially modifiable risk factors in New Zealand. Lancet Regional Health Western Pacific. 2021; 13:100191.

Content Manager:

Nathan Baker, AI in Health Research Network

Nathan.baker@auckland.ac.nz