Cognitively normal older adults are frequently treated as a homogeneous population in preventive and disease‑modifying trials for dementia. However, longitudinal cognitive outcomes vary substantially among individuals who are clinically categorized as cognitively normal. The source study applied a data‑driven approach to determine whether multimodal biological measures can stratify cognitively normal older adults into subgroups with distinct neurobiological profiles and differing risk for future cognitive decline.
The analysis used data from a large prospective cohort of N=1,136 cognitively normal older adults. Input features encompassed multiple domains: cortical thickness derived from MRI, plasma biomarkers related to Alzheimer’s disease, plasma inflammatory biomarkers, and vascular measures. The authors used openly available datasets (Health & Aging Brain Study – Health Disparities [HABS‑HD] and the Alzheimer’s Disease Neuroimaging Initiative [ADNI]) and stated that appropriate ethical approvals and participant consent were in place.
A deep learning clustering approach, Deep Embedded Clustering, was employed to integrate the multimodal biomarker data and partition individuals into biologically meaningful groups. The method jointly learns a low‑dimensional embedding and cluster assignments from the input features, enabling discovery of latent subgroup structure in the multimodal dataset.
The algorithm derived three biologically distinct subgroups organized along a gradient of neurobiological burden: low, intermediate, and high. Two modalities mostly drove cluster assignment: cortical thinning from MRI and inflammatory burden measured in plasma. The resulting clusters therefore reflect combined variation across neuroanatomy and systemic inflammation, with additional contributions from plasma AD biomarkers and vascular measures.
Although all participants were classified as cognitively normal at baseline, subgroup membership corresponded with measurable differences in cognitive performance and medical comorbidity. The High‑Burden subgroup exhibited significantly worse memory and executive function relative to the other subgroups. This group also had elevated rates of cardiometabolic comorbidity, indicating greater vascular and metabolic risk factors co‑occurring with the neurobiological burden identified by clustering.
The High‑Burden subgroup experienced markedly higher rates of conversion to mild cognitive impairment (MCI) or dementia within two years compared with the other subgroups. This finding indicates that the multimodal biological profile identified among clinically normal individuals is associated with near‑term risk for cognitive decline.
The subgroup results were validated in an independent external sample, demonstrating reproducibility of the clustering pattern across datasets. The source notes use of openly available data resources (HABS‑HD and ADNI) for the primary analysis and validation; precise validation sample characteristics and detailed validation metrics were not reported in the source article.
The study provides a data‑driven stratification framework that can identify cognitively normal individuals who nevertheless carry a higher biological burden and elevated short‑term risk of progression to MCI or dementia. Such stratification could be used to enrich enrollment in disease‑modifying or preventive clinical trials, targeting participants with a higher likelihood of near‑term decline and potentially increasing trial efficiency.
The source article reports the clustering approach, primary input modalities, subgroup characteristics, and that results were validated externally. The report does not provide detailed numeric validation metrics, full lists of which specific plasma biomarkers most strongly influenced clustering beyond the general statement that inflammatory burden and cortical thinning were primary drivers, or granular subgroup sample sizes. Those specific details were not reported in the source and therefore are not included here.
Using Deep Embedded Clustering on multimodal biomarker profiles from 1,136 cognitively normal older adults, researchers identified three biological subtypes aligned on a low–high neurobiological burden gradient. Cortical thinning and inflammatory burden primarily defined subgroup membership. The High‑Burden subgroup showed worse baseline cognition, greater cardiometabolic comorbidity, and substantially higher conversion to MCI or dementia within two years. Findings were validated in an external dataset. The authors suggest this framework can help enrich preventive and disease‑modifying trials by selecting cognitively normal individuals at higher risk of future cognitive decline.