Facial dysmorphism as a biomarker for autism spectrum disorder

BACKGROUND: Previous studies have indicated that children with autism spectrum disorder (ASD) have certain facial dysmorphisms. However, the means of exploiting these unique facial characteristics in clinical practice remain to be developed. Facial recognition technology (FRT) has tremendously improved in recent years, with it currently being utilized for diverse applications. Here tested the accuracy of FRT in detecting facial photos of children with ASD.

METHODS: High-resolution 2D facial photographs of 81 children with ASD from the Negev Autism Center (NAC) and 162 typically developing controls (matched by age, sex and ethnic origin) were evaluated by the deep-learning facial image algorithm, DeepGestalt. Classification accuracy into cases vs controls was assessed by a cross-validation approach as well as by external validation using another 40 matched case-control pairs.

RESULTS: Characteristics of the children with ASD in this study (mean age=4.77±2.042 years, 78% males, and 72% Jewish) were not significantly different from the other children with ASD in the NAC database. DeepGestalt was highly efficient in classifying cases from controls (AUC=0.89, P<0.001), which was remarkably better classification into males vs females (AUC=0.73 among cases, and AUC=0.81 among controls), or into Jewish vs Bedouin cases (AUC=0.72). Interestingly, the nose and eyes were the most distinctive ASD facial regions (AUC=0.92 and AUC=0.91 respectively). No significant facial differences were found between children with different severity levels of ASD.

CONCLUSIONS: Our findings suggest that the unique facial features of children with ASD are accurately detectable by FRT and could hence serve as a biomarker for ASD.