Picture the room first. The walls are built to keep the planet's own magnetism out. Inside sits a helmet-shaped scanner with 306 sensors, and under it a person with their eyes closed, asked to do nothing at all for a few minutes. No injection, no radiation, no noise, no sliding into a tube. The machine is listening for the faint magnetic fields thrown off by neurons as they fire, resolved millisecond by millisecond. The hope attached to this picture is a big one: that the brain's electrical conversation falters before memory does, and that a machine learning model could hear the falter. A scoping review published on 7 April 2026 in Frontiers in Dementia, led by Electra Chatzidimitriou at the Aristotle University of Thessaloniki with a co-author at the UCSF Memory and Aging Center, set out to count how much evidence sits under that hope. The answer is fourteen studies.
Fourteen, from a PubMed-MEDLINE search covering January 2015 to October 2025. The populations ranged from healthy controls through subjective cognitive decline and mild cognitive impairment to Alzheimer's disease and other dementias. Most used 306-channel whole-head systems. The algorithms ran from support vector machines and random forests to convolutional neural networks. Reported classification accuracies ran from moderate, around 60 percent, to high, with several studies above 80 percent at separating diagnostic categories or predicting progression from MCI to Alzheimer's. The features that kept recurring were frequency-specific oscillatory activity, functional connectivity, and large-scale network dynamics; network-based embeddings produced accuracies between 61 and 82 percent. One deep learning study cited in the review, an ensemble of two-dimensional convolutional networks, reached F1 scores as high as 0.92 for MCI classification. Combining MEG with structural MRI improved performance.
Those are striking numbers, and the review is unusually candid about why they should not be read as a test score. A scoping review maps a literature; it does not pool results or rank them by quality. Its own summary of the field names the problems plainly: heterogeneity in methods, limited longitudinal data, small sample sizes, overreliance on resting-state recordings, and a predominant focus on group-level analyses. Each of those cuts in the same direction. A model that separates two groups of people who already carry different diagnoses, inside the single dataset it was trained on, has not yet shown it can tell one new person what is coming.
The closest thing to a reality check comes from the largest shared MEG resource in this area. The BioFIND dataset, described by Delshad Vaghari and colleagues in a dataset paper posted in May 2021, pooled 324 people, 158 with clinically diagnosed MCI and 166 controls, recorded with eyes closed on 306-channel scanners at two sites, Cambridge and Madrid. In a 2022 NeuroImage analysis using 163 MCI cases and 144 controls from that dataset, Vaghari, Ehsanollah Kabir and Richard Henson reported that MEG alone classified MCI versus controls with 68 percent accuracy, worse than structural MRI alone at 71 percent. Combining the two helped: 74 percent when the modality-specific feature kernels were merged, and 77 percent when the predictions of separate classifiers were combined. Worth noting too, the Cambridge MCI participants were people referred from memory clinics with symptoms, not people picked up by screening while still cognitively well. Much of this field is classifying people who already know something is wrong.
So the honest answer to the question in the headline is: not yet, and not from these fourteen papers. What exists is a plausible signal, measured by a method the review describes as silent, non-invasive and well tolerated, which matters because it can be repeated over time without cost to the person being scanned. What does not exist is a validated presymptomatic test. The review's prescription is the unglamorous kind: standardized preprocessing pipelines, standardized task protocols across centres, larger multicentre longitudinal cohorts, explainable models whose reasoning a clinician can inspect, and a move toward newer sensor technology, optically pumped magnetometers, that could eventually take recordings outside the shielded room.
For a family reading this with a recent diagnosis or a worrying forgetfulness, nothing here is a test you can ask for, and nothing here should change a conversation already underway with a neurologist. MEG scanners remain rare, expensive and concentrated in research centres. What the review offers is a map of where a possible future tool currently stands: interesting, early, and roughly fourteen studies deep.