NutriBERT is a small transformer that learns what a person ate by reading their day as a sequence of food events, then turns it into a 128-number "dietary fingerprint" that beats every published dietary score (HEI-2020, AHEI, DASH, Mediterranean, EAT-Lancet) at predicting mortality and metabolic biomarkers in 63,684 NHANES adults — and proves that when you eat carries most of the diet→disease signal, not just what.
MIT