FaceMesh2HPO: AI Framework for Diagnosing Genetic Disorders via Facial Phenotyping

FaceMesh2HPO, developed by researchers from the GestaltMatcher Database Consortium and multiple universities, addresses a critical challenge in clinical genetics: extracting standardized, ontology-aligned facial phenotypes from patient photos to support rare disease diagnosis. The framework transforms 2D facial images into 3D meshes with 478 landmarks, then applies a hierarchical PointNet architecture with a novel cascading feature elimination strategy. This design allows the model to classify facial descriptors at multiple levels of the Human Phenotype Ontology (HPO), from broad categories (e.g., "abnormality of the face") to specific leaf terms (e.g., "thin upper lip"). The training set included annotations from 124 clinicians across 10 genetic disorders and non-syndromic controls, covering 107 distinct HPO terms. Demographic metadata and facial outline features were also integrated to boost performance.

The best-performing models achieved AUROCs ranging from approximately 0.55 to 0.89, with notably higher accuracy at parent ontology nodes compared to specific leaf terms. This hierarchical structure enables interpretability—clinicians can see which broad facial features drive predictions before drilling down to precise descriptors. However, external validation revealed variable generalizability across different disorders, and performance on rare leaf terms (those with few training examples) remained weak. The authors conclude that while 3D facial geometry modeling offers a promising, interpretable path for AI-assisted phenotyping, substantial improvements in data diversity, feature selection, and rare-term coverage are necessary before clinical deployment. The work represents a step toward automated, standardized facial phenotyping in genetic medicine.