ESCRS - FP31.03 - Predicting Outcomes In Lasik Treatment: An Explainable Machine Learning Approach With Uncertainty Quantification

Predicting Outcomes In Lasik Treatment: An Explainable Machine Learning Approach With Uncertainty Quantification

Published 2026 - 44th Congress of the ESCRS

Reference: FP31.03 | Type: Free paper | DOI: 10.82333/9r28-z988

Authors: Dr Anthony Man 1 , Dr Nabila Jones* 1 , Prof. Hema Radhakrishnan 2 , Dr Ajay Harish 3 , Mr Amir Hamid 1 , Prof. Clare ODonnell 4

1Optegra Eye Sciences,Optegra Eye Health Care,Manchester,United Kingdom;Faculty of Biology, Medicine and Health,University of Manchester,Manchester,United Kingdom, 2Faculty of Biology, Medicine and Health,University of Manchester,Manchester,United Kingdom, 3Faculty of Science and Engineering,University of Manchester,Manchester,United Kingdom, 4Optegra Eye Sciences,Optegra Eye Health Care,Manchester,United Kingdom;Faculty of Biology, Medicine and Health,University of Manchester,Manchester,United Kingdom;College Health and Life Sciences ,Aston University,Birmingham,United Kingdom

Purpose

To develop a machine learning model that predicts post-operative uncorrected distance visual acuity (UDVA). This predictive tool will assist surgeons in identifying suitable LASIK candidates and in making informed clinical decisions through individualised outcome predictions.

Setting

Optegra Eye Health Care, UK

Methods

Retrospective clinical data from 6416 eyes of 3945 patients were used as input features to train a machine learning model for predicting postoperative UDVA. The eyes underwent LASIK vision correction surgery aiming for emmetropia. The preoperative input measures included patient demographics, preoperative refraction, planned correction, and planned surgical parameters. XGBoost was chosen as the model type due to its popular and proven effective usage in tabular regression, such as the present work. Shapley Additive Explanations (SHAP) were used to assess the impact of each input feature on model predictions. A calibration process was undertaken to embed uncertainty quantification into the predictions to provide confidence intervals.

Results

The model was able to predict postoperative UDVA with an average error of 0.09 logMAR superseding state-of-the-art performance reported in the literature of 0.10 logMAR. SHAP analysis revealed that preoperative corrected distance visual acuity (CDVA), planned cylinder correction, and planned central ablation depth are the most consequential features in determining postoperative UDVA. Based on the retrospective data, the uncertainty quantification functionality was able to calculate the probability of a patient receiving a suboptimal outcome. The probability of achieving 0.2 logMAR or higher for a model prediction of -0.2 logMAR was 0.001% while for a model prediction of 0.1 logMAR, the probability was approximately 20%.

Conclusions

The machine learning model was able to predict postoperative UDVA to state-of-the-art accuracy, with supporting explainability and uncertainty quantification functionalities. Predictions from this model provide an additional offering to inform patients of their risk and suitability for laser vision correction procedures. Future research will focus on generalising this model architecture to other refractive surgery modalities to enhance preoperative decision support across multiple treatment options.