Applying Artificial Intelligence Techniques to Identify Factors Associated with Corneal Graft Rejection
Mohammadreza Arzaghi1 , Sepehr Feizi1 *, Mohammad Ali Javadi1 , Kia Bayat2 , Siavash Shirzadeh Barough1
- Ocular Tissue Engineering Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, No. 23, Paidarfard St., Boostan 9 St., Pasdaran Ave, Tehran, Iran.
- Ocular Tissue Engineering Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, No. 23, Paidarfard St., Boostan 9 St., Pasdaran Ave, Tehran, Iran.Ocular Tissue Engineering Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, No. 23, Paidarfard St., Boostan 9 St., Pasdaran Ave, Tehran, Iran.
Abstract: The aim was to develop models for predicting immunologic reaction following corneal transplants in keratoconus patients, utilizing various data mining methods and evaluating their performance.
Methods: This retrospective cohort study included 985 keratoconus patients who received either primary penetrating keratoplasty (462 eyes, 46.9%) or deep anterior lamellar keratoplasty (523 eyes, 53.1%). The study examined donor and recipient characteristics, surgical details, and postoperative events to assess their association to the risk of immunologic reaction. Supervised machine learning methods such as gradient boosting, support vector machine, extra tree classifier, and random survival forests were utilized to predict initial graft rejection episodes and their risk factors. The data were split into a training set (70%, 690 eyes) and a validation set (30%, 295 eyes), with the models appraised on calibration (Brier scores) and discrimination (C-Statistics) for comparing their performance and discriminatory capacity.
Results: Overall, 286 eyes (29.0%) developed at least one episode of graft rejection. Random survival forests (Brier score=0.156, C-statistics=0.723) outperformed extra tree classifier (Brier score=0.228, C-statistics=0.667), gradient boosting (Brier score=0.363, C-statistics=0.633), and support vector machine (Brier score=0.448, C-statistics=0.559) in predicting graft rejection. Key predictors of graft rejection identified by random survival forests were the duration of postoperative corticosteroid use, keratoplasty technique, and time until complete suture removal.
Conclusion: Random survival forests proved to be a more effective method for identifying predictive factors of corneal graft rejection in keratoconus. The findings suggest maintaining a low dose of topical corticosteroids post corneal transplant in keratoconus patients until all sutures are removed, particularly for those undergoing penetrating keratoplasty.
