Marfan Syndrome Prediction Via Graph Neural Networks on 3D Facial Cues

Marfan syndrome (MFS) is a rare genetic connective tissue disorder whose early detection is critical to prevent life-threatening cardiovascular complications. To support the diagnostic process, which currently relies on complex and costly clinical assessments and genetic testing, we propose a 3D vision-based method for rapid, non-invasive screening for MFS using facial scans. The method analyzes 3D facial morphology using Graph Neural Networks on either manually or automatically extracted landmarks, leveraging geometric relationships to capture subtle craniofacial patterns and a statistically derived sparse graph topology to improve robustness and accuracy; moreover, an adjustable decision threshold enables context-dependent prioritization of a low false-negative rate for MFS diagnosis. Experimental evaluation on a high-resolution dataset of 440 subjects (126 MFS and 314 controls) demonstrates strong discriminative performance (AUC up to 0.98, and classification accuracy above 93%).