E-POSTER DETAIL | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Validation of a machine learning algorithm to predict compliance to follow up of keratoplasty patients across an eye care network

Authors
Tanmay Gokhale, Poornima Tandra, Kavya Chandran, Pravin Krishna

Presenting
Tanmay Gokhale

PURPOSE:
Purpose: The Asia-Pacific region has one of the highest rates of high risk keratoplasties owing to the high incidence of corneal infections. Most of the population in need of a keratoplasty are economically backward and have to travel great distances to receive specialised care. Identifying patients who are at high risk of complications, and stratification according to their likelihood of missing a follow up visit is therefore a major challenge for healthcare teams involved in management of such patients. Our aim was to validate a multifactorial machine learning algorithm capable of predicting compliance to follow up visits in the post operative period following keratoplasty

PURPOSE:

Purpose: The Asia-Pacific region has one of the highest rates of high risk keratoplasties owing to the high incidence of corneal infections. Most of the population in need of a keratoplasty are economically backward and have to travel great distances to receive specialised care. Identifying patients who are at high risk of complications, and stratification according to their likelihood of missing a follow up visit is therefore a major challenge for healthcare teams involved in management of such patients. Our aim was to validate a multifactorial machine learning algorithm capable of predicting compliance to follow up visits in the post operative period following keratoplasty

METHODS:
32,900 keratoplasties performed across a network of eye hospitals between January 2016-December 2024 were included and divided in a 80:20 ratio into training and retrospective validation set. Following data optimization, a machine learning algorithm was created taking into account various clinical parameters to predict the outcome measures probability of visiting an eye care centre at various time points following keratoplasty. The same algorithm was then validated prospectively using 5800 keratoplasties that occurred between January and December 2025

METHODS:

32,900 keratoplasties performed across a network of eye hospitals between January 2016-December 2024 were included and divided in a 80:20 ratio into training and retrospective validation set. Following data optimization, a machine learning algorithm was created taking into account various clinical parameters to predict the outcome measures probability of visiting an eye care centre at various time points following keratoplasty. The same algorithm was then validated prospectively using 5800 keratoplasties that occurred between January and December 2025

RESULTS:
The keratoplasty compliance prediction model was able to predict likelihood of a patient visiting an eye care centre at 1,3,6 and 12 months with 98%, 93%, 89% and 81% accuracy respectively.

CONCLUSIONS:
The compliance prediction tool is capable of reasonably predicting whether a patient is likely to be lost to follow up following keratoplasty for a period of upto 10 years.