NeuroFM foundation model enables brain health profiling and dementia risk estimation from MRI

This cohort study evaluated NeuroFM, a foundation model for precision neuroimaging and individualized brain health estimation. The model was trained on 100,000 healthy synthetic brain volumes and evaluated on 136,361 real volumes spanning multiple cohorts. The study established a disease-naive foundation model paradigm for precision neuroimaging, meaning no diagnostic labels were used during training.

The main findings showed that NeuroFM organizes brain MRIs into population-level patterns that encode meaningful brain health differences. The model representations transferred across five neuroscience domains without adaptation. NeuroFM enabled individual-level brain health profiling, including estimation of future dementia risk years before diagnosis. However, no effect sizes, confidence intervals, or statistical significance measures were reported for these outcomes.

Safety and tolerability data were not reported. The model was trained exclusively on synthetic data, which represents a key methodological limitation. No diagnostic labels were used during training, and the study did not report effect sizes, confidence intervals, or statistical significance for outcomes. The practice relevance is that this establishes a disease-naive foundation model paradigm for precision neuroimaging, but clinical application requires validation in diagnostic settings with reported statistical measures.