Blood Tests for a Sharp Old Brain? A Korean Study Says Maybe, But the Signal Is Thin
Researchers at Ewha Womans University in Seoul drew blood from 81 healthy people aged 65 and over, 39 of whom performed on cognitive tests like adults in their 40s (“SuperAgers”) and 42 of whom performed as expected for their age. They measured 55 blood markers, from routine cholesterol and glucose panels to specialist assays for inflammation and lipid transport proteins. Conventional statistics found only four markers that differed between the groups, and none of those would survive correction for multiple testing. The team then applied machine learning with feature selection, synthetic data generation using a large language model, and SHAP interpretability, arriving at a 15-marker panel that classified SuperAger status with 76 percent accuracy and an AUC of 0.739 on a held-out set of 33 people. Notably, Alzheimer’s proteins and classic inflammatory cytokines showed no group difference at all. The metabolic markers did most of the work.
The interesting question in cognitive aging is not why some people decline. It is why a small minority do not. SuperAgers, people past 65 whose memory and executive function match adults thirty years younger, have been studied mostly through expensive brain scans that show thicker cortex and preserved connectivity. What nobody had done was ask a simpler question: does a SuperAger’s blood look different from everyone else’s?
This Korean team asked it. They recruited 81 healthy elderly volunteers from a Seoul hospital, sorted them by performance on a standard neuropsychological battery, and ran 55 blood measurements on each. The design was deliberately practical. If you could spot cognitive resilience from a routine health-checkup panel, you would have a cheap screening tool that works at population scale, unlike MRI.
The headline result is a negative one, and it is the most informative part of the paper. Tau protein, amyloid precursor protein, IL-6 and TNF-alpha, the biomarkers that dominate dementia research, showed no difference between the two groups. Neither did LDL, triglycerides, CRP, kidney function, or most of the panel. Whatever distinguishes a SuperAger is not visible through the lens the field has been using.
What did emerge, faintly, was metabolic. SuperAgers had lower fasting glucose (98 versus 106 mg/dL), lower HbA1c (5.75 versus 6.00 percent), higher HDL cholesterol (59 versus 56 mg/dL), and marginally lower oxidised LDL. Each of these differences is small, and each carries a p-value that the authors obtained using one-tailed tests without correcting for 55 comparisons. Taken individually, none is convincing.
The authors’ argument is that the individual markers are the wrong unit of analysis. Cognitive resilience, they suggest, is a pattern across many markers rather than any single one. To test that, they trained a gradient boosting classifier and found the combination did carry real information: roughly 76 percent classification accuracy where guessing the larger group would give 52 percent.
The catch is that they got there partly by inflating a 49-person training set with 70 synthetic patients generated by a language model. Synthetic data can stabilise a model. It cannot create information that was not in the original 49 people. The paper is a well-motivated proof of concept, not evidence that a blood panel predicts cognitive reserve. [Confidence: High]
Insights
The differences that showed up are real in direction but small in magnitude. Statisticians measure this with Cohen’s d, which asks how far apart two groups are relative to how spread out people are within each group. A d of 0.4 means the two bell curves overlap by about 84 percent. Practically, if you picked one SuperAger and one typical ager at random, the SuperAger would have the better fasting glucose only about 61 percent of the time, versus 50 percent by coin flip.
Every blood difference in this study sits between d = 0.25 and d = 0.41. That is the magnitude of a modest population trend, not a personal predictor.
What the data weakly supports, and what is already supported by much stronger evidence elsewhere: keeping fasting glucose in the high 80s to low 90s rather than the low 100s, keeping HbA1c below roughly 5.8, and keeping HDL above 54 mg/dL rather than merely above the standard clinical cutoff of 40 to 50. The recurring theme is that clinically normal is not the same as optimal, and the SuperAgers clustered at the favorable end of normal rather than outside it.
What this study does not support: any claim that AGE, RAGE, CD36, vimentin or leptin levels tell you anything about your brain. Those markers made the panel on algorithmic ranking alone and showed no group difference.
Context and Source
-
Open Access Paper: Machine learning based prediction of cognitive metrics using major biomarkers in SuperAgers
-
Authors: Hyo-Bin Lee, So-Yeon Kwon, Ji-Hae Park, Bori Kim, Geon-Ha Kim, Jang-Hwan Choi, Young Mi Park
-
Institution: Ewha Womans University (Departments of Computational Medicine, Molecular Medicine, and Artificial Intelligence) and Ewha Womans University Mokdong Hospital, Seoul, South Korea
-
Journal: Scientific Reports, volume 15, article 18735
-
Impact evaluation: The impact score of this journal is 3.9 (2024 Journal Citation Reports two-year JIF; aggregator sources report a 2025 figure near 4.9, which I have not been able to confirm against the primary JCR record), evaluated against a typical high-end range of 0 to 60+ for top general science, therefore this is a Low-to-Medium impact journal. Scientific Reports is a high-volume, soundness-only megajournal: peer review assesses technical validity rather than novelty or importance, so publication here carries little signal about a finding’s significance.