Early identification of Prader-Willi syndrome in newborns using facial phenotype-based machine learning.

BACKGROUND: Existing facial models for Prader-Willi syndrome (PWS) demonstrate limited accuracy in newborns with subtle facial dysmorphism. METHODS: Photographs of children from a public platform (training cohort) and of newborns from Children's Hospital of Fudan University (validation cohort) were used for feature comparison. Models were developed using six algorithms with three feature sets (Complete-, Top-10, and ClinSel-) in the training cohort. Model performance, calibration, feature importance, and sensitivity analysis were assessed. External validation via threshold re-estimation was compared with Face2Gene. RESULTS: Of 113 children (21 [19%] PWS), RF and AdaBoost gained the highest Area Under the Precision-Recall Curve across all feature sets. RF Complete-set model achieved highest sensitivity (mean [95% CI], 0.97 [0.84-1.00]) with 0.95 specificity (95% CI, 0.83-1.00) and was well calibrated. Age, face width, and upper lip-related features were important features. Four previously unreported features (jaw symmetry, chin, nose, and jaw widths) were identified. Sensitivity analysis showed 26 facial features contribute more than age. In external validation (10 PWS newborns and 13 controls), RF Complete-set model achieved comparable sensitivity (0.80 vs. 0.70) and significantly higher specificity (0.77 vs. 0.15) at a false positive rate ≤0.25 compared with Face2Gene. CONCLUSIONS: RF Complete-set model demonstrates potential for neonatal PWS identification. IMPACT: Facial phenotype-based machine learning models for identifying Prader-Willi syndrome (PWS) in newborns facilitate early detection of PWS before genetic testing. Models achieved comparable sensitivity but higher specificity than Face2Gene, providing a favorable tool for PWS. Four previously unreported facial features-jaw symmetry, chin, nose, and jaw widths-expand the PWS facial phenotype. Models trained on children successfully identified PWS newborns after threshold adjustment, showing the value of reusing pediatric data when newborn samples are limited. Pilot validation in newborns with 0.80 sensitivity and 0.77 specificity supports the feasibility of machine learning-based screening for genetic syndromes.