A new approach in autism diagnosis: Evaluating natural interaction using point of view (POV) glasses
Makaleler > Tam Makale · 2025
Yazarlar: Hakan Kayış, Murat Çelik, Çınar Gedizlioğlu, Elif Kayış, Cumhur Aydemir, Arda Hatipoğlu, Seda Sarı, Nurhak DOĞAN, Elif Mumcu, Rabia Arı, Nazlı Burcu ÖZBARAN
Asian Journal of Psychiatry, Uluslararası, Wos, 2025, cilt 116, ss 104798-104798
DOI: 10.1016/j.ajp.2025.104798
Özet
This study introduces an AI-assisted method based on examiner-worn Point of View (POV) glasses and computer vision analysis to provide objective behavioral data for the diagnosis of Autism Spectrum Disorder (ASD). The study included 29 children with ASD and 27 children without ASD, aged between 17 and 36 months. During semi-structured naturalistic interactions, the examiner wore POV glasses equipped with a scene camera that captured the child’s face from an eye-level perspective, preserving ecological validity. Behavioral parameters—including facial expressions, approximate social gaze (operationalized as the child’s eyes orientation toward the POV camera), and head mobility—were extracted using OpenFace and MediaPipe and subsequently analyzed with machine learning techniques. Statistical analyses revealed that total social gaze duration, longest social gaze, social smiling, number of responses to name, response latency, response duration, social responsiveness, and head movements along the z-axis had p-values ≤ 0.05, while head movements on the x- and y-axes, total head movement, and rapid head movements had p-values > 0.05. The classification model developed using decision trees and the AdaBoost algorithm demonstrated high performance, achieving an accuracy of 91.07% and a sensitivity of 89.65%. These findings support the clinical applicability of examiner-worn POV recordings for early ASD detection and highlight their potential to complement traditional, subjective assessment methods. • Examiner-worn POV glasses captured children’s natural social behaviors in real time. • AI-based computer vision quantified gaze, facial expressions, and head movements. • Significant group differences observed in gaze, smiling, and name-response measures. • Machine learning model (AdaBoost) achieved 91% diagnostic accuracy (AUC = 0.95). • AI-driven POV framework offers a feasible tool for early ASD detection in clinics.