Automated Evaluation of D-Score for Facial Dysmorphism Analysis in Central African Children With Developmental Disorders

Abstract

Introduction

Dysmorphism is an important characteristic, but its evaluation is largely subjective. A good clinical assessment (dysmorphism) can facilitate a more accurate and efficient diagnosis. We therefore evaluated an automated artificial intelligence tool for facial dysmorphism, D-score, available in Face2Gene.

Methodology

We evaluated 2D frontal facial photographs of pediatric individuals with a developmental delay/intellectual disability from the Democratic Republic of Congo (144) and Belgium (137) as being dysmorphic or not, first clinically, and second by D-score analysis. We determined the performance of D-score by calculating sensitivity, specificity, positive predictive value, negative predictive value, F1-score and Cohen's Kappa (κ). We also evaluated the effects of sex, age, and ethnicity on D-Score.

Results

Of the 144 Congolese children, 69 (47.9%) were dysmorphic, compared to 40.9% in the Belgian cohort. D-score in the Congolese cohort showed a sensitivity of 85.5%, a specificity of 68%, a PPV of 71.1%, and an NPV of 83.6%. The F1-score was 0.78. The k was 0.531 (0.395–0.666) with a standard error of 0.069, p = 0.000. In the Belgian cohort, sensitivity was 71.4%, specificity 71.6%, PPV 63.5%, and NPV 78.4%. The F1-score was 0.672. The k was 0.422 (0.270-0.574) with a standard error of 0.078, p = 0.000. There was no statistically significant difference depending on age and sex.

Conclusion

D-score is not replacing the gestalt facial evaluation, but it is promising to be used in clinical practice as a supplementary tool for precision in the dysmorphism evaluation, especially when dealing with rare or less common genetic conditions.