Document resource
Objectives Computational phenotyping tools, like Face2Gene, are increasingly used to support genetic diagnosis by analysing facial features. These tools perform well in Global North populations but are less accurate in other groups. This study aimed to assess the performance of Face2Gene’s DeepGestalt, FeatureMatcher, and D-Score in South African children with neurodevelopmental disorders (NDDs) and unaffected controls.Methods Facial photographs of 301 children from the NeuroDev South Africa study were analysed including 36 children with NDDs with a confirmed molecular diagnosis, 176 with NDDs without a confirmed molecular diagnosis, and 89 controls. Diagnostic accuracy of DeepGestalt and FeatureMatcher was assessed by the presence of the correct diagnosis among the top-1 and top-10 ranked suggestions generated. D-Scores across sub-groups were compared to determine the tool’s sensitivity, specificity, and predictive value for detecting dysmorphia.Results Among children with genetic diagnoses, 19% had the correct condition ranked within the top-1 and 34% within the top-10 suggestions. Accuracy improved to 33% (top-1) and 61% (top-10) when limited to conditions included in the DeepGestalt training set. Compared with clinician assessment of dysmorphism, the D-Score showed 78% sensitivity and 42% specificity. It had a low positive predictive value (21%) but a high negative predictive value (91%), making it reliable for ruling out dysmorphia, but not ruling in.Conclusions Face2Gene tools show promise, but current performance in South African children limits clinical utility. These findings underscore the need for datasets that represent children from diverse populations to enhance diagnostic accuracy of computational phenotyping tools.