Machine Learning Detects Subtle Facial Patterns Associated With Sickle Cell Disease

An interpretable machine-learning model analyzing front-facing smartphone photographs has identified distinctive geometric and texture facial features associated with sickle cell disease in children and adolescents, according to research conducted in the Democratic Republic of the Congo. Sickle cell disease is a common inherited genetic disorder that contributes substantially to global childhood morbidity and mortality, particularly across Sub-Saharan Africa. While craniofacial and skeletal alterations have previously been recognized in individuals living with the condition, subtle variations across soft-tissue facial morphology had not been quantitatively characterized using computational methods.

To investigate these morphometric patterns, researchers evaluated 308 standardized front-facing smartphone photographs collected from 154 participants with electrophoresis-confirmed sickle cell disease and 154 age- and sex-matched controls. Study participants ranged in age from 5 months to 19 years and were recruited in Kinshasa. Using manually annotated anatomical landmarks, the research team extracted predefined geometric descriptors alongside texture-based facial features to capture detailed soft-tissue variations across the cohort.

The team analyzed the extracted descriptors using an interpretable linear support vector machine combined with recursive feature elimination, evaluating the framework through leave-one-out cross-validation. This pipeline ranked facial characteristics based on their individual relevance for distinguishing participants with sickle cell disease from control subjects. In total, the facial analysis algorithm isolated eight geometric features and six texture features that differed significantly between the two evaluated groups.

A specific subset of six geometric measurements contributed most consistently to internal discrimination between the groups. These key features comprised the distance between the medial and lateral canthi, the angle at the nasal ala, the distance from the nasion to the philtrum, the distance from the medial canthi to the columella, the distance from the columella to the lower lip, and the distance between the nasal alae. Incorporating these parameters, the model achieved an area under the receiver operating characteristic curve of 0.843, with a 95 percent confidence interval between 0.798 and 0.887, and attained an overall classification accuracy of 0.795, with a 95 percent confidence interval spanning 0.750 to 0.840.

Cross-sectional analyses stratified by age and sex revealed notable variations in the expression of selected morphological features over time. Morphological differences between participants with sickle cell disease and control subjects were limited among the youngest age groups, whereas disparities became distinctly more pronounced across several older cohorts. This divergence suggests that soft-tissue facial variations associated with the disorder become increasingly evident across developmental stages.

The study authors emphasize that these results provide preliminary evidence that sickle cell disease is linked to quantifiable morphometric facial patterns. However, the researchers noted key limitations and outlined necessary future directions, stating that external validation, longitudinal tracking, and clinically characterized studies are essential to determine whether these computational features can reliably serve as reproducible digital biomarkers within a broader supportive disease-assessment framework.

Subject of Research: Quantitative facial phenotyping in sickle cell disease using machine learning

Article Title: Quantitative facial phenotyping and age-associated differences in Sickle cell disease using interpretable machine learning

Article References: Spencer, D., Liu, X., Mosema-Be-Amoti, K., Kandosi, G., Bramble, M. S., Al Munajjed, F., Likuba, E., E-Andjafono, D. O. L., Tshibambe, L., Colwell, B., Howell, K., O’Brien, N., Moxon, C., Anwar, S. M., Porras, A. R., Ngoyi, D. M., Vilain, E., Tshala-Katumbay, D., & Linguraru, M. G. (2026). Quantitative facial phenotyping and age-associated differences in Sickle cell disease using interpretable machine learning. PLOS Digital Health, 5(11), e0001813. https://doi.org/10.1371/journal.pdig.0001813

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001813

Keywords: Sickle cell disease, Facial phenotyping, Machine learning, Linear support vector machine, Smartphone photography, Kinshasa, Democratic Republic of the Congo, Sub-Saharan Africa, Morphometrics, Pediatric health, Digital health, Geometric descriptors, Texture analysis, Recursive feature elimination, Digital biomarkers, Canthi, Columella, Facial landmarks