Tiroid Göz Hastalığında Hastalık Şiddetinin Oftalmolojik ve Ortoptik Klinik Bulgular ile Korelasyonu: Klinik Belirteç Olarak Mikroperimetri ve Derin Öğrenme Tabanlı Program Kullanımının Değerlendirilmesi

Abstract

Karaca B. Correlation of Disease Severity with Ophthalmological and Orthoptic Clinical Findings in Thyroid Eye Disease: Evaluation of Microperimetry and Deep Learning-Based Software as Clinical Biomarkers. Hacettepe University Faculty of Medicine, Department of Ophthalmology, Specialty Thesis, Ankara, 2026. Objective: Thyroid eye disease (TED) is the most common extrathyroidal manifestation of Graves' disease. The present study aimed to (i) evaluate whether macular sensitivity and fixation stability abnormalities measured by MAIA™ microperimetry are already present in Graves' disease patients without overt orbitopathy, (ii) test whether the wide bivariate contour ellipse area (BCEA) values observed in these patients represent retinal dysfunction or an adaptive strategy operating within the Panum fusion area, and (iii) assess the diagnostic performance of TED Detector, an artificial intelligence-based photograph analysis pipeline. Methods: This cross-sectional observational study enrolled 180 participants: 90 healthy controls, 48 Graves' disease patients without clinical orbitopathy (Graves w/o GO), and 42 Graves' orbitopathy patients (Graves w/ GO). Disease severity was graded by EUGOGO classification. All participants underwent MAIA™ microperimetry, yielding average threshold (AT, dB), fixation stability (FS) at P1/P2 probability levels and 63%/95% BCEA. Primary position facial photographs were analyzed with the TED Detector for macOS v2.3 pipeline. Three-group comparisons were performed with the Kruskal–Wallis test with Bonferroni-corrected Dunn post-hoc testing. Age effects were controlled with Quade's nonparametric ANCOVA. Logistic regression models were fitted using Firth's penalized maximum likelihood to handle quasi-complete separation caused by near-zero lid retraction values in controls. Results: Macular sensitivity was significantly reduced in both disease groups compared with controls (AT (Average Threshold) OD (oculus dexter)) median: controls 27.1 dB; Graves w/o GO 25.6 dB; Graves w/ GO 25.2 dB; H = 40.98, ε² = 0.229, p < 0.001). After Bonferroni correction, no MAIA™ microperimetry parameter differed significantly between the Graves w/o GO and Graves w/ GO subgroups (all p > 0.05). Fixation stability P1 OD decreased across both disease groups (96.0% vs 90.0% vs 77.5%; p < 0.001). The majority of wide BCEA values remained within the Panum foveal fusion area bounds (≈ 0.1–1 deg²). All group differences persisted after age adjustment. The TED Detector pipeline achieved 87.8% sensitivity, 85.6% specificity and 86.7% accuracy (κ = 0.733). Firth Model 1 (controls vs all Graves) had AUC = 0.902; Model 2 (Graves w/o GO vs w/ GO) had AUC = 0.901. In Model 2, no MAIA™ microperimetry parameter was a significant discriminator; only pipeline Hertel (OR = 2.05; 95% CI 1.20–3.49, p = 0.008) and pipeline CAS (Clinical Activity Score) (OR = 5.69; 95% CI 1.10–29.57, p = 0.039) reached significance. Conclusion: In Graves' disease the principal macular functional deterioration appears to emerge before overt orbitopathy develops, suggesting that objective retinal involvement may be present at the subclinical stage. The observation that wide BCEA values largely remain within Panum foveal fusion limits supports the hypothesis that this "enlarged" fixation distribution may reflect an adaptive compensatory strategy preserving binocular vision rather than a pure motor disturbance. The TED Detector pipeline demonstrated high diagnostic accuracy for both general TED detection and the orbitopathy subgroup distinction, and may serve as an adjunct to photograph-based screening. Keywords: Thyroid eye disease, Graves' orbitopathy, microperimetry, BCEA, Panum fusion area, subclinical orbitopathy, artificial intelligence, deep learning.