Photograph-Based Skeletal Malocclusion Screening with Geometric Machine Learning: External Validation Across Turkish and Non-European Cohorts

A photograph based screening classifier for skeletal malocclusion maps 100 engineered geometric feature vectors, derived from a side-profile smartphone photograph, to orthodontic class labels, enabling non-invasive triage without radiographic imaging. This study reports external validation across two independent cohorts that contributed no cases to training. Two binary tasks were evaluated; Class III against non-Class III, and Class I against Class II; each with its own development set and model, applied without modification at the default threshold. One external cohort was Turkish, drawn from an institution not represented in the development set; the other was a non-European cohort from a Belgian private orthodontic clinic. Sensitivity, specificity and the area under the receiver operating characteristic curve (AUC) were pre-specified as primary outcome measures, since none depends on outcome prevalence. In the Class III task, sensitivity was similar across all three evaluation sets (0.801, 0.836, 0.848; all adjusted p>0.7), while specificity fell from 0.766 in the Turkish external cohort to 0.697 in the non-European cohort. AUC was 0.902 and 0.835 respectively (difference 0.066; p=0.029). In the Class I and Class II task the corresponding values were 0.875 and 0.723 (difference 0.152; p<0.001). A subgroup analysis showed lower sensitivity in children than adults across both cohorts (0.71 to 0.80 versus 0.89 to 0.90). AUC was lower in the non-European cohort for all three models in both tasks. The observed performance differences are consistent with a population-specific generalizability effect rather than model-specific failure and support mandatory external validation on target populations before clinical deployment.