The early warning signs of dementia can be so subtle that they’re easy to miss—forgetting an appointment or losing track of time. It's no wonder that physicians sometimes struggle to diagnose it accurately. Diagnosis can be influenced by subjective factors and cognitive biases, leading to inconsistencies as different doctors might interpret the same symptoms in varying ways.
Clinical biomarkers add another layer of complexity, with different subtypes of dementia linked to various physiological changes, such as beta-amyloid plaques, neurofibrillary tangles (NFT), or brain blood vessel damage.
Dennis Wang is a Senior Principal Investigator at A*STAR Institute for Human Development and Potential (A*STAR IHDP), formerly Singapore Institute for Clinical Sciences (SICS), and Bioinformatics Institute (BII), believes machine learning (ML) can transform this landscape. “ML reduces redundancy, handles large and complex datasets and is more adaptable and scalable than clinicians’ assessments,” said Wang, highlighting how ML can make dementia diagnoses faster, more efficient and more cost-effective.
In their exploration of ML-based dementia diagnostics, Wang and researchers from the University of Sheffield, Cambridge Public Health and Newcastle University in the UK, used data from the Cognitive Function and Ageing Study (CFAS), involving 186 brain samples. They developed several ML methods to rank over 30 neurological features, focusing on Alzheimer’s and NFT markers among other dementia-related pathologies.
Out of 34 neuropathological features, 22 were identified as crucial for dementia classification. The most significant included Braak NFT stage, beta-amyloid and cerebral amyloid angiopathy. The study found that some dementia cases were misclassified due to overfitting, which happens when too many features are used, and the presence of unusual brain changes.
By highlighting the most important features and removing redundant ones, the researchers improved both the model’s accuracy and the design of future models. They also combined common and uncommon brain features to reduce misclassification.
“There’s always a risk of hidden confounders when learning from elderly subjects,” commented Wang. “The study cohort we worked with had recorded other lifestyle and behavioural information, but we’ve yet to incorporate these into our models.”
Their study reported the best-performing ML model to achieve 79 percent sensitivity, 69 percent specificity and 75 percent precision in classifying dementia. “Our model can accurately identify and rank critical features associated with dementia, potentially reducing subjectivity and variability in computer-aided dementia diagnosis and helping to track disease progression,” said Wang.
With their focus on early factors that influence health outcomes, the researchers are now aiming to identify biomarkers that appear earlier in the disease process—such as molecular changes in blood samples or early-life stress in Singaporean cohorts of mothers and children—as risk factors for dementia.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute for Human Development and Potential (A*STAR IHDP) and Bioinformatics Institute (BII).
