Neural Network -Supported Prognostic Assessment of Apoptotic and Inflammatory Markers in Neovascular Glaucoma

Abstract

Purpose: To assess the prognostic value of apoptotic (CD95), endothelial (CD54), and systemic inflammatory biomarkers (SII, SIRI, AISI) for predicting the effectiveness of modified selective diode transscleral cyclophotocoagulation (TSCPC) in patients with neovascular glaucoma using regression and neural network-based models. Materials and methods: This prospective cohort study included 258 eyes: 224 eyes with NVG (155 treatment successes and 69 failures) and 34 age-matched healthy controls. Baseline assessments comprised intraocular pressure (IOP), best-corrected visual acuity (BCVA, logMAR), and serum levels of CD95 (Fas), CD54 (ICAM-1), HbA1c, and systemic inflammatory indices (SII, SIRI, AISI). Treatment success was defined as IOP between 10 and 21 mmHg, stabilization or improvement of BCVA, and absence of additional glaucoma surgery. Between-group comparisons were performed using the Mann-Whitney U test and Dunn-Bonferroni post hoc analysis. Predictors of treatment failure were assessed using univariable and multivariable logistic regression. Discriminative performance was evaluated by receiver operating characteristic (ROC) analysis. A gradient-boosting machine-learning model with SHAP-based interpretability was applied to an independent validation set. Results: Patients with TSCPC failure exhibited significantly higher biomarker levels than those who had successful outcomes: CD95 (+41%), CD54 (+71%), and AISI (+164%) (all p < 0.001). In univariable analysis, CD95 (OR = 1.80; 95% CI: 1.55-2.15), AISI (OR = 1.60), and CD54 (OR = 1.50) showed the strongest associations with treatment failure. In the multivariable model, CD95 (OR = 1.65), CD54 (OR = 1.30), and AISI (OR = 1.40) remained independently significant, with robust explanatory power (Nagelkerke R² = 0.60). The combined apoptosis-endothelial model (CD95 + CD54 + AISI) demonstrated superior discriminatory performance (AUC = 0.87; sensitivity 82%; specificity 76%). SHAP analysis confirmed that these three biomarkers accounted for approximately 75% of total predictive importance. Discussion: The findings of this study indicate that the effectiveness of modified selective TSCPC in neovascular glaucoma is determined not merely by baseline IOP, but predominantly by the patient's integrated biological profile, particularly apoptotic (CD95), endothelial (CD54), and systemic inflammatory markers (SII, SIRI, AISI). The observed convergence of endothelial activation, systemic inflammation, and Fas-mediated apoptosis supports the concept of NVG as a multifactorial ischemic–inflammatory disorder and provides a mechanistic basis for treatment resistance. The combined biomarker model demonstrated high predictive performance (AUC = 0.87), outperforming individual parameters and underscoring the value of a multimarker approach. These findings support biomarker-guided risk stratification and reinforce the need for personalized therapeutic strategies, including early neuroprotective and anti-inflammatory interventions in patients at high risk of unfavorable outcomes. Conclusions: Apoptosis (CD95), endothelial activation (CD54), and systemic inflammation (AISI, SII, SIRI) constitute key biological determinants of resistance to TSCPC in NVG. An integrated CD95-CD54-AISI biomarker panel provides high prognostic accuracy and may support personalized risk stratification and treatment optimization in patients with neovascular glaucoma. Keywords: neovascular glaucoma, transscleral cyclophotocoagulation, CD95, CD54, AISI, SII, systemic inflammation, apoptosis, endothelial activation, prognosis, machine learning