For thirty years, a fluorodeoxyglucose PET scan has been read a little like a weather map: the parietal lobes look cool, the temporal lobes look cool, the pattern is consistent with Alzheimer's. Hot spots and cold spots, region by region. What a team at the Indiana University School of Medicine has done, with colleagues at the Jackson Laboratory for Genomic Medicine, Marian University and Purdue, is to stop reading the map and start reading the grammar: not how much glucose any one region burns, but how tightly each region's burn rate moves in step with every other region's, across a whole group of people. Their paper, published in Alzheimer's & Dementia on April 28, 2026, argues that this coordination has its own stage-by-stage signature.
The study is a retrospective reanalysis of images that already existed. From phases 2 and 3 of the Alzheimer's Disease Neuroimaging Initiative, the authors pulled 431 people who had an 18F-FDG PET scan and cognitive testing (ADAS-Cog, CDR and MoCA) within 180 days of it: 98 cognitively normal, 76 with early mild cognitive impairment, 129 with MCI, 42 with late MCI and 86 with Alzheimer's disease. Ages ran from 55 to 95, averaging 73.6 years; 55 percent were men. Each scan was warped into a common atlas space and reduced to mean glucose uptake in 59 brain regions, and for each of the five clinical groups the authors built a 59 by 59 matrix of how strongly those regional values covaried across the people in that group. Then they ran graph theory over the result: network density, clustering, nodal strength, entropy, and the shortest paths between regions. The approach is an extension of the same laboratory's earlier work applying metabolic covariance networks to a mouse model of Alzheimer's, published in the same journal in 2023.
The numbers that come out are not about any individual brain, and that distinction matters enormously; but as group portraits they are striking. Network density traced what the authors call a W shape across the continuum in both sexes, dropping at the transition stages, early and late MCI, and recovering between them. Path lengths, a measure of how efficiently signal-carrying regions sit relative to one another, shortened from cognitively normal through MCI and then increased significantly in Alzheimer's. In the Alzheimer's group, whole-brain entropy, connection strength and clustering coefficients all rose. The authors read this as reorganisation under strain: the network rewiring itself to hold something together, then losing efficiency once the disease is established. That reading is an interpretation of correlation structure, not a measurement of compensation.
The sex split is the part likeliest to unsettle a clinician. Using a data-driven consensus clustering method with 10,000 permutations, the team counted how many distinct modules each group's metabolic network broke into. In men the count fell as the disease advanced: 10 modules in the cognitively normal group, 6 in early MCI, 8 in MCI, 6 in late MCI, and 3 in Alzheimer's. In women it climbed: 5, then 14, 13, 7 and 15. Same disease label, same scanner protocol, opposite trajectory in how the brain's energy economy divided itself up. A new method the authors built, region set enrichment analysis, mapped those modules onto function and found, in both sexes, a decline in regions tied to hearing, cognition, memory and language alongside an increase in motor regions. In a subgroup of 116 people with blood gene expression data available (77 cognitively normal, 18 MCI, 21 Alzheimer's), transcriptomic modules identified by network analysis lined up with the imaging picture.
For a family reading this, the useful translation is narrow and worth being precise about. Nobody's scan was scored with this method to give them a stage; the networks only exist at the level of a group, assembled from one image per person, so the method as run here cannot say anything about an individual. Nobody was followed over time, so the apparent arc from normal through MCI to Alzheimer's is stitched from different people at different points, not from watching anyone travel it. The ADNI volunteers are a research cohort, older, heavily screened and not representative of everyone who gets this diagnosis. And metabolic covariance is correlation: two regions moving together across a group does not mean one is driving the other, and the compensatory story the authors offer is a hypothesis their data are consistent with, not a mechanism they measured.
What it does buy is a change of unit. If the signal of disease stage lives in the relationships between regions rather than in any single one, then the thing to measure is a pattern, and patterns can in principle be tracked, compared between men and women, and used to ask whether a drug is changing the system rather than one spot in it. The authors themselves frame the next step as integrating metabolic, functional and structural data. That work has not been done yet.