Evaluation of an AI facial recognition system for Turner Syndrome screening and facial complexity: a prospective cohort.

Abstract

Artificial intelligence-based facial recognition (AI-FR) is promising in diagnosis of diseases with distinct facial features. Our team has retrospectively constructed an AI-FR system for Turner Syndrome (TS) based on 1295 facial photographs in previous research. This study aims to evaluate this AI-FR system for TS screening in a prospective cohort in real-world clinic setting. We also aim to elucidate the impact of complexity of facial features on diagnostic accuracy of AI-FR in this cohort. Patients were recruited in a single-gate prospective cohort at a clinic. Facial images were collected for AI-FR diagnosis. Karyotype analyses were performed as the gold-standard diagnosis. Diagnostic performance of the AI-FR system was evaluated. Individual facial recognition intensity (iFRI) was proposed to characterize the complexity of individual facial features in patients. iFRI was calculated based on 26 facial landmarks of seven parts. Group comparison and regression analysis were performed on iFRI according to diagnosis classification. The prospective cohort included 218 patients (29 TS and 189 control). The AI-FR system showed performance: 89.0 % accuracy (95 % CI 84.1-92.5), 72.4 % sensitivity (95 % CI 54.3-85.3), and 91.5 % specificity (95 % CI 86.7-94.7). TS vs. control diagnosed by AI-FR had higher iFRI (p = 0.009), while TS vs. control diagnosed by gold standard showed no iFRI difference (p > 0.05). AI diagnosis classification positively correlated with iFRI (p = 0.033), while gold-standard classification did not (p > 0.05). This was one of the pioneering prospective cohorts in disease diagnosis with AI-FR. This system for TS screening achieved ideal performance and improved the diagnosis of TS. iFRI was proposed and proved influential to AI-FR diagnostic performance.