Posters

Predicting postoperative visual acuity for cataract surgeries using machine learning

Poster Details

First Author: O.Eloka GERMANY

Co Author(s):    F. Rombold                    

Abstract Details

Purpose:

Goal of this study is to predict the uncorrected visual acuity (UVCA) after cataract surgery (4 weeks) using machine learning techniques. Being able to estimate the UVCA helps surgeons to manage their patient’s expectations and to understand their patient’s priorities. Thus, unrealistic expectations by the patient and subsequent postoperative dissatisfaction may be drastically reduced.

Setting:

Since 2015, we have been working with 20 different surgical centers across Germany, where all of their cataract surgeries are documented for this observational study.

Methods:

Cooperating ophthalmologists documented medical data across the whole cataract treatment cycle for more than three years. A total of 4706 complete datasets were collected. Data was analyzed using machine learning algorithms, with pre-operative examination details (e.g. gender, age, UVCA, best corrected visual acuity (BCVA) with glasses, sphere, cylinder, axis) used as features. Four different regression models were considered: Boosted Decision Tree Regression, Bayesian Linear Regression, Neural Network Regression and Linear Regression. Three metrics were used: Mean Absolute Error (MAE), Mean Squared Error (MSE) and Root Mean Squared Error (RMSE). A k-fold cross-validation was performed with k=10.

Results:

Our results showed that the Boosted Decision Tree Regression model performed best of all algorithms with values of MAE= 0.093, MSE= 0.023 and RMSE= 0.152. The Bayesian Linear Regression model achieved values of MAE= 0.127, MSE= 0.032 and RMSE= 0.181. The Linear Regression model achieved values of MAE= 0.125, MSE= 0.034 and RMSE= 0.185 and the Neural Network Regression model performed worst with values of MAE= 0.148, MSE= 0.037 and RMSE= 0.194.

Conclusions:

We found that machine learning models can reliably predict post-operative UVCA using pre-operative examination details. The variation between the different models is consistent with various other studies. Our next step will be to identify the most important factors influencing the prediction. Furthermore, we will use classification methods to find out if there are particularly suitable/unsuitable IOLs for specific patient subgroups. Finally, we will use classification methods to identify high risk patient subgroups.

Financial Disclosure:

is employed by a for-profit company with an interest in the subject of the presentation