Kan AI opdage visse sygdomme ved at se på billeder af ansigter ?

Nuværende AI-systemer kan udtrække suggestive signaler fra ansigtsfotografier – ændringer i tekstur, asymmetri, pigmentering og subtil hævelse – der korrelerer med visse metaboliske, hjerte- og endokrine lidelser, men disse signaler er ikke sygdomsspecifikke og overlapper ofte med normal variation eller andre tilstande. Forskningsgrupper har rapporteret moderate nøjagtigheder (ofte 60–80 % AUC) for at opdage sygdomme som diabetes, kronisk nyresygdom eller koronar arteriesygdom, hvilket bygger på store datasæt og dyb læring-modeller trænet på titusinder af mærkede billeder. Da ansigtsbiomarkører er indirekte og påvirkes af alder, køn, belysning og etnicitet, forbliver teknologien eksperimentel og er ikke godkendt til klinisk diagnose. Den anvendes i øjeblikket hovedsageligt i forskningsmiljøer og som et supplerende screeningsværktøj snarere end en diagnostisk standard.

— Beriget 13. maj 2026 · Kilde: Nature Medicine

Background

Artificial-intelligence systems can extract suggestive facial cues—texture changes, asymmetry, pigmentation shifts and subtle swelling—that correlate with metabolic, cardiac and endocrine disorders, but these biomarkers overlap with normal variation and other conditions. Reported accuracies for diseases such as diabetes, chronic kidney disease and coronary artery disease typically range from 60 % to 80 % AUC, relying on large labeled datasets and deep-learning models trained on tens of thousands of images.

Facial phenotyping has been explored as a non-invasive, low-cost screening approach for genetic and neurodegenerative disorders. Convolutional neural networks have improved detection of conditions such as Down syndrome, DiGeorge syndrome, Parkinson’s disease and Alzheimer’s disease in research settings. However, facial traits are heavily influenced by age, sex, lighting and ethnicity, and published results remain investigational; the technique is not approved for clinical diagnosis and is currently used mainly in research and as an adjunctive screening tool rather than a diagnostic standard.

Sources: Nature Medicine; National Institutes of Health (enriched May 13, 2026).