Abstract
PURPOSE: To identify and analyze clinical risk factors and imaging parameters influencing the outcomes of deep anterior lamellar keratoplasty (DALK) for advanced keratoconus (KC) using an artificial intelligence (AI) model. METHODS: This study included 250 DALK eyes with a 5-year follow-up for advanced KC. The DALK eyes were classified as having "favorable" or "unfavorable" outcomes based on graft clarity, scarring at the graft-host interface involving the visual axis which was not pre-existing, early suture loosening less than 3 months after the surgery, corneal vascularization reaching up to or into the graft-host junction at any follow up period, persistent corneal edema greater than 3 months after surgery, and change in visual acuity. Clinical risk factors were determined through a detailed clinical evaluation and questionnaire assessment and included the presence of systemic allergy, ocular allergy, or eye rubbing. Immunoglobulin E (IgE) and vitamin D and B12 levels were obtained from blood investigations. A total of 37 tomographic parameters were exported from an OCULUS Pentacam HR. An AI model was then built to assess these risk factors and imaging parameters. The area under the curve (AUC) and other metrics were evaluated. RESULTS: The AI model classified 92.2% and 89.4% cases as favorable or unfavorable, respectively, based on clinical risk factors and imaging parameters. Systemic allergy, IgE, eye rubbing, and vitamin D had the highest information gains followed by posterior corneal data from the Pentacam HR. The AI model achieved an AUC of 0.957 with sensitivity of 98% and specificity of 85.6%. CONCLUSIONS: Our findings demonstrate the importance of preoperative risk stratification, which can affect surgical outcomes of DALK using AI. TRANSLATIONAL RELEVANCE: Better identification and control of these factors would enable better management and outcomes of DALK for advanced KC.