Machine learning in dementia risk identification using routinely collected data

C. Prieto, Gillian Dobbie, Claudia Rivera Rodriguez, Daniel Wilson, Susan Yates, Sarah CullumAlzheimer 2025 · DOI

TL;DR

This paper demonstrates that machine learning models can accurately identify individuals at risk of dementia using routinely collected healthcare data. By leveraging diverse data sources, these models offer a scalable solution to improve early diagnosis, enabling timely interventions and better patient outcomes.

What Problem Does This Paper Address?

Dementia is a growing global health crisis, affecting millions and placing a significant burden on healthcare systems. A major challenge is that nearly half of dementia cases remain undiagnosed, delaying crucial interventions and increasing costs. This lack of early identification is particularly problematic in regions like New Zealand, where prevalence is high, especially among Māori and Pacific populations. The paper tackles the urgent need for efficient, scalable methods to identify individuals at risk of dementia, moving beyond traditional diagnostic pathways that can be resource-intensive and time-consuming.

Methods

The researchers employed deep learning models, trained within a standard training/validation/testing framework, to predict dementia risk. A key methodological innovation was the use of a nested one-to-one case-control design. This allowed them to analyze longitudinal data, specifically examining six distinct time windows before a confirmed dementia diagnosis (ranging from 0 days to 8 years prior). This pre-diagnostic window is crucial for identifying early warning signs. The models incorporated a wide array of routinely collected health data, including sociodemographic information, clinical variables, pharmacy records, interRAI assessments, and hospitalization data (using ICD-10 codes).

Datasets

The study utilized routinely collected health data from the Te Whatu Ora Counties Manukau population in New Zealand, focusing on individuals aged 65 and older. The data included:

  • Sociodemographic variables
  • Clinical variables (longitudinal and cross-sectional)
  • Pharmacy records (specifically antidementia drug prescriptions)
  • InterRAI assessments (dementia-related evaluations)
  • Hospitalization records (ICD-10 codes related to dementia)

The paper does not introduce a new benchmark dataset but rather leverages existing, real-world healthcare data.

Key Insights and Ideas

  • Routinely collected data is a goldmine: The study powerfully illustrates that readily available healthcare data, often collected for administrative or clinical purposes, can be repurposed for sophisticated predictive modeling. This bypasses the need for specialized, expensive data collection.
  • Pre-diagnostic signals are detectable: The ability to identify risk up to 8 years before diagnosis is a significant finding. It suggests that subtle patterns in health records can flag individuals who will later develop dementia, opening a window for proactive care.
  • Multifaceted risk factors: The models highlight that dementia risk isn't tied to a single factor. A combination of factors, including residential care needs, existing comorbidities (captured by ICD-10 codes), and even ethnicity, are strong predictors. This underscores the complex nature of dementia development.
  • Pharmacy data as an early indicator: Specific medication classes (analgesics, diuretics, antithrombotics, anti-epileptic drugs, diabetes treatments) and the frequency of certain specialist appointments (ophthalmology, cardiology) emerged as highly relevant. This suggests that changes in medication or healthcare utilization patterns can be early signals.

Key Results

The developed deep learning models achieved a notable accuracy of 80.47% (95% CI: 80.23-80.72) when using data immediately preceding diagnosis. The models demonstrated a sensitivity of 71.77% (95% CI: 71.73-71.81) and a specificity of 87.27% (95% CI: 87.24-87.30). The highest performance was observed when the models incorporated all available features. Key predictive features identified included Aged Residential Care (ARC)-related factors, comorbidities (ICD-10 codes), and ethnicity. Pharmacy-related variables such as analgesics, diuretics, antithrombotics, anti-epileptic drugs, and diabetes treatments were also highly relevant, as were ophthalmology and cardiology appointments, and delirium-related features.

Limitations and Future Work

The authors acknowledge that the models' performance was highest when using data immediately before diagnosis. This suggests that while early signals are present, the predictive power might decrease further out from diagnosis. Future work could focus on optimizing models for earlier detection windows and exploring the causal relationships behind the identified predictive features. Further validation across diverse populations and healthcare systems would also be beneficial to ensure generalizability. The study also highlights the potential for these models to enhance clinical decision-making, optimize healthcare resource allocation, and ultimately improve outcomes for individuals living with undiagnosed dementia.

Summary updated

Abstract

Abstract Background Dementia is a major global public health challenge, affecting approximately 55 million people worldwide, with an estimated 10 million new cases annually (WHO). In New Zealand, prevalence estimates based on national datasets suggest that dementia affects 3.8%–4.0% of individuals aged 60 and older. When accounting for undiagnosed cases using a capture‐recapture method, this estimate increases to 9.2% (95% CI: 8.9%–9.6%), with disproportionately higher rates among Māori and Pacific populations. However, nearly 50% of dementia cases remain undiagnosed, limiting timely interventions and increasing healthcare costs. There is an urgent need for scalable, data‐driven solutions to improve dementia identification. This study aimed to develop accurate machine learning models for dementia identification using routinely collected health data. Method Routinely collected health data from the Te Whatu Ora Counties Manukau population (aged 65+) were analysed, incorporating sociodemographic and clinical variables (both longitudinal and cross‐sectional). Dementia status was determined based on pharmacy records (antidementia drug prescriptions), interRAI assessments (dementia‐related evaluations), and hospitalization records (ICD‐10 codes related to dementia). A nested one‐to‐one case‐control design was implemented, retaining pre‐diagnostic information across six‐time windows before diagnosis (0 days, 6 months, 1 year, 3 years, 5 years, and 8 years). Deep learning models were trained using a training/validation/testing framework. Result The models achieved an accuracy of 80.47% (95% CI: 80.23–80.72) when using data immediately before diagnosis with a sensitivity of 71.77% (95% CI: 71.73–71.81) and specificity of 87.27% (95% CI: 87.24–87.30). The highest performance was achieved when all available features were included. Key predictive features included Aged Residential Care (ARC)‐related factors, comorbidities (ICD‐10 codes), and ethnicity. Among pharmacy‐related variables, analgesics, diuretics, antithrombotics, anti‐epileptic drugs, and diabetes treatments were highly relevant. The number of ophthalmology and cardiology appointments, along with delirium‐related features, also contributed significantly. Conclusion Machine learning applied to routinely collected health data offers a powerful, scalable approach to dementia case‐finding, improving early identification and facilitating timely interventions. These models have the potential to enhance clinical decision‐making, optimize healthcare resources, and ultimately improve outcomes for individuals living with undiagnosed dementia.
BibTeX
@inproceedings{prieto2025machine,
  author = {C. Prieto and Gillian Dobbie and Claudia Rivera Rodriguez and Daniel Wilson and Susan Yates and Sarah Cullum},
  title = {{Machine learning in dementia risk identification using routinely collected data}},
  booktitle = {Alzheimer},
  year = {2025},
  doi = {10.1002/alz70856\_105348},
}
Figure