Artificial Intelligence, Machine Learning and Calculation of Intraocular Lens Power

Abstract

Background and Purpose In the last decade, artificial intelligence and machine learning algorithms have been more and more established for the screening and detection of diseases and pathologies, as well as for describing interactions between measures where classical methods are too complex or fail. The purpose of this paper is to model the measured postoperative position of an intraocular lens implant after cataract surgery, based on preoperatively assessed biometric effect sizes using techniques of machine learning. Patients and Methods In this study, we enrolled 249 eyes of patients who underwent elective cataract surgery at Augenklinik Castrop-Rauxel. Eyes were measured preoperatively with the IOLMaster 700 (Carl Zeiss Meditec), as well as preoperatively and postoperatively with the Casia 2 OCT (Tomey). Based on preoperative effect sizes axial length, corneal thickness, internal anterior chamber depth, thickness of the crystalline lens, mean corneal radius and corneal diameter a selection of 17 machine learning algorithms were tested for prediction performance for calculation of internal anterior chamber depth (AQD_post) and axial position of equatorial plane of the lens in the pseudophakic eye (LEQ_post). Results The 17 machine learning algorithms (out of 4 families) varied in root mean squared/mean absolute prediction error between 0.187/0.139 mm and 0.255/0.204 mm (AQD_post) and 0.183/0.135 mm and 0.253/0.206 mm (LEQ_post), using 5-fold cross validation techniques. The Gaussian Process Regression Model using an exponential kernel showed the best performance in terms of root mean squared error for prediction of AQDpost and LEQpost. If the entire dataset is used (without splitting for training and validation data), comparison of a simple multivariate linear regression model vs. the algorithm with the best performance showed a root mean squared prediction error for AQD_post/LEQ_post with 0.188/0.187 mm vs. the best performance Gaussian Process Regression Model with 0.166/0.159 mm. Conclusion In this paper we wanted to show the principles of supervised machine learning applied to prediction of the measured physical postoperative axial position of the intraocular lenses. Based on our limited data pool and the algorithms used in our setting, the benefit of machine learning algorithms seems to be limited compared to a standard multivariate regression model.

Considerations on the Castrop formula for calculation of intraocular lens power

“…In the present paper we describe the strategy of intraocular lens power calculation using the Castrop formula as described in Materials and Methods. The large dataset of 1452 eyes was split into training and test subsets [ 25 ]. The training data were used for formula constant optimization, and the test data for cross validation.…”

“…For cross validation, the N = 1452 measurements were split randomly into a training set (70%, 1017 cases) and a test set (30%, 435 cases) [ 25 ]. For the training set, the constants for the Castrop formula (C, H, R), the SRKT (A), the Hoffer-q (pACD), the Holladay1 (SF), the simplified Haigis (a0) and Haigis formula with triplet optimization (a0/a1/a2) were optimized in terms of minimizing the rmsPE using the nonlinear Levenberg-Marquardt algorithm [ 21 , 22 ].…”

IOL Formula Constants: Strategies for Optimization and Defining Standards for Presenting Data

“…With repeated random subsampling, a random partition is separated out before calculation of the formula constant and this separated dataset then used for validation of the formula constant. However, this carries the risk that some data points may never be included in the training or validation data sets [10].…”

Artificial intelligence applications and cataract management: A systematic review

“…12 Langenbucher and coworkers attempted to model the measured postoperative position of an intraocular lens implant after cataract surgery using machine learning techniques. 42 Based on preoperative effect sizes of AL, corneal thickness, internal ACD, LT, mean corneal radius, and corneal diameter, 17 machine learning algorithms were tested for prediction performance for calculation of internal anterior chamber depth (AQD post) and axial position of the equatorial plane of the lens in the pseudophakic eye (LEQ_post). In terms of root mean squared error for AQDpost and LEQpost prediction, the Gaussian Process Regression Model with an exponential kernel outperformed the machine learning algorithms.…”

“…Recent studies have shown that AI algorithms have the potential to exhibit comparable or higher accuracy compared to that of experienced clinicians. 42 Using rigorous validation approaches, these AI applications might be To date, only a few clinical trials have compared the efficacy of AI systems in real-world settings.…”