AI predicts brain age, flags bias

Beth Israel Deaconess Medical Center researchers reported in March that an artificial-intelligence model trained on sleep EEG recordings could estimate a person’s “brain age” and that larger gaps between estimated and chronological age tracked with higher dementia risk. In a separate 2026 paper, Johns Hopkins University researchers working with the U.S. Food and Drug Administration described a tool meant to find hidden bias in medical AI training data before those systems are deployed. Together, the two projects show where medical AI is advancing fastest: pattern detection in large datasets, not autonomous clinical judgment. The studies also point to the same bottleneck — what a model learns, and whether clinicians can trust how it got there. ### How did the sleep study estimate “brain age”? JAMA Network Open published the brain-age study online on March 19, 2026 as an individual participant data meta-analysis spanning five community-based longitudinal cohorts. The authors said the model used sleep electroencephalography, or EEG, to derive a Brain Age Index measuring the gap between EEG-estimated brain age and a person’s actual age. (jamanetwork.com) PubMed’s summary of the paper said the pooled analysis included community-dwelling adults and examined whether a higher sleep EEG-based Brain Age Index was associated with incident dementia. The authors said the signal comes from sleep EEG microstructure rather than symptoms reported in clinic visits. ### What was the key dementia-risk finding? (jamanetwork.com) The JAMA paper found that every 10-year increase in the Brain Age Index was associated with a 39% higher risk of incident dementia. The association held after adjustment for factors including age, sex, education, lifestyle factors and overall health, according to coverage that summarized the study and the paper’s abstracted findings. (pubmed.ncbi.nlm.nih.gov) JAMA Network Open also published an invited commentary saying the work introduced a “compelling candidate” digital marker for early detection of dementia in community settings. The commentary said more work is still needed to test predictive value, generalizability and clinical use. ### What problem is the bias-detection tool trying to solve? (jamanetwork.com) Johns Hopkins University said on July 22, 2026 that its researchers, working with the FDA, built a tool to uncover hidden issues in the large datasets used to train medical AI. The university said the method searches for subtle patterns that could lead models to wrong conclusions even when headline performance looks strong. (jamanetwork.com) Nature Digital Medicine published the paper describing the framework as a “generalized and modality-agnostic” approach to detecting dataset bias. The paper said medical AI can amplify latent bias when training data are not representative, and that those problems can remain hidden during standard testing. ### Why can a model look accurate and still be biased? (hub.jhu.edu) Johns Hopkins said the tool examines how prediction labels relate not only to patient attributes such as age or sex, but also to environmental and acquisition factors such as clinical site and imaging protocol. That matters because a model may appear highly predictive while relying on shortcuts tied to where or how data were collected rather than to disease itself. (nature.com) The Jerusalem Post, citing the researchers, reported that the project was designed to expose “problematic signals” that conventional performance metrics can miss. One researcher, Andreas C. Unberath, was quoted saying developers often optimize for prediction without full control over which signals a model uses. ### Does either study put AI in charge of decisions? (hub.jhu.edu) KESQ, summarizing broader evidence on medical AI, reported this week that AI systems can perform strongly on diagnosis while physicians remain better at weighing treatment options. That distinction fits both new studies: one identifies a possible screening marker, and the other audits training data, but neither replaces consent, explanation or treatment decisions by clinicians. (jpost.com) March 19, 2026 is the publication date for the dementia-risk paper in JAMA Network Open, while the bias-detection framework appeared in Nature Digital Medicine in 2026 and was described by Johns Hopkins on July 22. The next steps named by the researchers are validation in broader clinical settings for the EEG marker and use of the auditing tool by developers and regulators reviewing medical AI datasets. (jamanetwork.com) (hub.jhu.edu)