Multi-modal AI-enabled steatotic liver disease diagnostics using facial images and metabolomics.

Abstract

Steatotic liver disease (SLD) affects one-third of the global population, yet current non-invasive diagnostic methods are too costly or operator-dependent for population-scale screening. Here, we present 3D-FAICE, a deep learning system that uses three-dimensional facial imaging for non-invasive SLD detection. Trained and tested on 11,456 participants, the facial model achieves robust performance across internal, external, and self-controlled longitudinal cohorts and remains effective in a smartphone-based point-of-care setting. Metabolomic analysis reveals that facial risk scores correlate with glycolipid and amino acid pathways, supporting biological plausibility. Multimodal fusion of facial and metabolomic data further improves accuracy, and a cross-modal distillation strategy significantly elevates the performance of the facial-only model. These findings establish facial image-based AI as a non-invasive, scalable, and privacy-aware tool for SLD screening, with potential applications in self-monitoring and population health management.