Free Papers

Predicting outcomes in patients undergoing keratoplasty: AI machine learning based model

Free Paper Details

First Author: V.Joshi INDIA

Co Author(s):    P. Vaddavalli                    

Abstract Details

Purpose:

Outcomes of corneal transplants depend on a multitude of factors including pre-operative, donor related, intra-operative factors and post-operative compliance.This study was planned to predict the outcomes of corneal transplants by assessing visual acuity, graft survival and compliance for patients over a 10 year period

Setting:

The study was conducted at a Tertitary Eye care centre, L V Prasad Eye Institute at the Hyderabad campus in India

Methods:

The study included pre-operative, surgical and post-operative clinical observations associated with 17176 corneal transplants in the past 20 years at our clinic. The process involved data mining from medical records of the patients and was analyzed for outcomes across several categories such as demographic, clinical, surgical, donor and post-surgical follow-up. Several machine learning algorithms associated with classification and regression trees were trained to develop outcome predictions post-surgery. The accuracy of the outcomes was assessed on the test split (80:20) of dataset for each of the models trained and the model with highest accuracy was picked for final predictions.

Results:

The dataset included 17176 corneal transplants of 14000 patients of which 13509 underwent a single transplant and 3887 underwent more than one transplant. 10629 of the surgeries were performed in male patients. Based on the age at the time of surgery number of surgeries performed in adults (>16 years), children (1 – 16 years of age) and infants (0-1 years of age) were 14843 (86.41%), 1769 (10.29%) and 564 (3.2%) respectively. Considering the economic background, 7348 surgeries (42.7%) were performed in patients from poorer socioeconomic backgrounds. Infections (29.73%) endothelial disorders (22.33%) and corneal opacities (16.15%) were the commonest indications for corneal transplants.

Conclusions:

This machine learning trained model based on a large dataset was 80.5 % accurate in predicting post op BCVA and 81.2 % accurate in predicting compliance for follow up after surgery at different time points, with the accuracy for BCVA being greater at the initial time points while that of compliance being greater for the later time points. The study aided by machine learning methods enabled us to build a model which could help predict the outcomes of grafts at different time points.

Financial Disclosure:

-