Fast and accurate visual acuity prediction based on optical aberrations and machine learning

In this work, we propose three machine learning-based methods for predicting visual acuity (VA). Two methods utilize regression trees (LSBoost and XGBoost), and the third employs a neural network that classifies simulated aberrated optotypes as "recognized" or "unrecognized". The overall VA is estimated by replicating the clinical procedure in which the subject reads optotypes and the VA is determined based on their responses. Here, the neural network acts as a substitute for the subject. Data were collected from a clinical trial involving 135 subjects providing for each sample 36 Zernike coefficients, amplitudes of accommodation, age, and VA values. Evaluation of the regression tree models demonstrates that LSBoost outperformes XGBoost in prediction accuracy, especially when incorporating amplitudes of accommodation. However, XGBoost is faster in computation time, making it more suitable for large datasets and design of visual compensations. The neural network, while achieving high optotype recognition accuracy, is less accurate in VA prediction due to its reliance on synthetic data and complex simulation processes, which requires large processing times. Overall, LSBoost offers the best performance in terms of accuracy, while XGBoost provides faster computation. These findings highlight the suitability of regression tree-based models for VA prediction using tabulated data.