Optimising ELP Prediction with AI

With postoperative effective lens position (ELP) prediction remaining a major limiting factor in the accuracy of IOL power calculation, researchers have applied artificial intelligence (AI) techniques to improve ELP accuracy. Results of a study conducted by Woong-Joo Whang MD and colleagues indicate that enhancing deep learning models with an attention mechanism can optimise ELP prediction.

“I think concepts of attention are very useful for improving ELP prediction and the accuracy of IOL power calculations,” Dr Whang said.

“We know that the contribution of different ocular parameters to ELP changes depending on axial length (AL) range. This makes it challenging to define a single equation that applies across the entire spectrum of AL, and so we need alternatives.”

Structural equation modelling is one option, but even with its introduction, results of a previous study showed the optimal combination of variables and the best-fitting equations change depending on AL.¹

“Attention mechanisms identify the relationships between input variables, dynamically weigh which inputs are most important for prediction, and capture complex interactions between many variables,” Dr Whang explained.

“For example, in IOL calculations, AL might be more significant when considered together with corneal power (K) in certain scenarios, while in others, the relationship of ELP with anterior chamber depth (ACD) could be more relevant. The attention mechanism automatically learns these contextual relationships, so it is ideal for understanding the relationships between various ocular parameters and finding the optimal combination.”

Testing the concept

The potential for improving ELP prediction using attention-enhanced deep learning models was investigated in a retrospective study involving 1,120 patients who underwent conventional cataract surgery at Yeouido St Mary Hospital in Seoul. Data included were AL, K, ACD, lens thickness, and white-to-white (corneal diameter)—all measured by swept-source optical coherence tomography—and ELP was calculated by the vergence formula.

ELP prediction performance was evaluated using 35 models that consisted of five standard deep learning models (multilayer perceptron, 1.0 D-convolutional neural network, recurrent neural network, gated recurrent unit, long short-term memory) used alone and each model enhanced individually by one of six types of attention mechanisms (self attention, multi-head attention, temporal attention, gate attention, additive attention, scaled-dot attention), yielding a total of 30 attention-enhanced deep learning models.

A comparison of the results achieved with the various models versus traditional multiple regression analysis showed that 20 of the 35 deep learning models provided superior ELP prediction performance. A comparison of results obtained with the deep learning models alone versus in combination with an attention mechanism showed that 22 of the 30 attention-enhanced deep learning models offered better performance than the standard deep learning model.

An analysis of the outcomes achieved with each of the five deep learning models showed that different architectures benefited from different types of attention. Two of the deep learning models performed best when enhanced with multi-head attention, the best results for two of the others were obtained using the additive and temporal mechanisms, while gated attention optimised performance of the fifth deep learning model.

“These findings indicate that a tailored attention mechanism selection for each neural network is crucial,” Dr Whang said.

Overfitting analysis was performed to verify the models learned genuine biometric–ELP relationships rather than memorising training data. The test-to-training mean squared error ratio remained below 1.2 across all 35 combinations, indicating true pattern recognition.

Dr Whang spoke on this topic at the 2026 ASCRS annual meeting in Washington, DC.

Woong-Joo Whang MD is an Assistant Professor of Ophthalmology, Catholic University of Korea, South Korea. olokl@nate.com

1. Yoo YS, Whang WJ. J Clin Med, 2022; 11(6): 1469.