Pilot Validation of ChatGPT-Based Strabismus Assessment Using 9Gaze Photographs: A Single-Center Feasibility Study

Authors : Ata Baytaroğlu, Şerife Nur Çiftci

Pages : 155-160

Doi:10.33713/egetbd.1829121

View : 13 | Download : 34

Publication Date : 2025-12-31

Article Type : Research Paper

Abstract :OBJECTIVE: To investigate the degree of agreement between artificial intelligence (AI)-based strabismus measurements obtained from images of nine diagnostic gaze positions and the actual diagnosis and amount of deviation recorded during clinical examination. MATERIALS and METHODS: The study included twenty cases diagnosed with horizontal strabismus. For each patient, nine gaze position photographs taken using the 9gaze application (See Vision LLC, Virginia, USA) under fixation on a near target were used, and horizontal and vertical deviation values were recorded during clinical examination. Data on the amounts of horizontal and vertical deviations, incomitance status, pattern presence, and type of strabismus were reviewed from clinical records. The same photographs were uploaded to ChatGPT-5.0-Plus, and the diagnosis, incomitance, pattern, and deviation amounts generated by the AI algorithm were documented. RESULTS: The average age of the 20 cases included in the study was 21.0±20.9 (1–65) years; 10 (50%) were female and 10 (50%) were male. According to the actual diagnosis, 11 (55%) had esotropia and 9 (45%) had exotropia. The number of cases correctly classified in the clinical diagnosis classification of the YZ was 19/20 (95%), showing excellent agreement with Cohen\\\'s kappa = 0.90. Sensitivity for esotropia was 90.9%, specificity was 100%, and overall accuracy was 95%. Clinical and AI analyses showed 75% agreement for incomitance (Kappa=0.38). The AI algorithm was found to be inadequate in detecting pattern shift (%80 agreement, Kappa=-0.05). Strong correlations were observed in horizontal and vertical shift analyses (r=0.87, p<0.001 and r=0.77, p<0.001). No significant relationship was found between age and gender and the absolute error magnitude (p>0.05 for all). CONCLUSION: AI-based analysis of nine diagnostic gaze position photographs shows a high level of agreement with clinical measurements in estimating strabismus type and deviation magnitude. However, agreement is much lower for more subtle diagnostic features such as incommitance and A/V-pattern. The findings suggest that properly trained AI systems can serve as a useful diagnostic support tool in strabismus practice but cannot replace clinical examination, especially in cases of incomitant and patterned strabismus.
Keywords : Şaşılık, 9 bakış pozisyonu, yapay zeka, derin öğrenme

ORIGINAL ARTICLE URL