Artificial Intelligence-Based Clinical Decision-Making System for Cataract Surgery

Abstract

Cataracts are a common eye condition characterized by the clouding of the natural crystalline lens that impairs vision. In cataract surgery, the natural lens of the eye is replaced by an intraocular lens (IOL) implant to restore vision to the patient. A cataract surgery procedure involves the removal of the natural cloudy lens by means of ultrasound phacoemulsification, followed by the replacement of the lens with an IOL implant chosen specifically for the patient. The power of the IOL should be selected appropriately in order to attain optimal postoperative vision. IOL power calculation has evolved over multiple generations, starting with regression and theoretical optics-based formulas to machine learning (ML) formulas, as well as hybrid formulas. Nevertheless, there is still room for improvement in terms of prediction accuracy. At University of Michigan’s Kellogg Eye Center, we gathered a large collection of medical records of cataract patients and retrieved demographic and surgical information from the Sight Outcomes Research Collaborative (SOURCE) database. In this dissertation, I present research projects in which we used artificial intelligence-based approaches to improve the functionality and accuracy of tools assisting the clinical decision-making procedure. Through a series of studies, we developed and investigated methodologies for achieving better refractive results with cataract surgery with the aid of artificial intelligence algorithms. Predicting where the intraocular lens resides within the eye after cataract surgeries is a critical step in the determination of the IOL power and several other applications (e.g., ray tracing). In Chapter 2, we gathered a dataset of 847 patients and developed methods for predicting the postoperative IOL position (postoperative anterior chamber depth) which outperformed existing methods including Haigis, Hoffer Q, Holladay I, Olsen, and SRK/T. Further, we explored whether the ML-predicted lens position can be used to improve the performance of existing IOL formulas. In Chapter 3, we combined the ML-based postoperative ACD prediction method described in Chapter 2 with existing optics-based methods including Haigis, Hoffer Q, Holladay, and SRK/T. These methods use theoretical Gaussian optics-based equations to estimate the effective lens position (ELP). When combine with our ML-predicted lens position, all four formulas achieved significantly better prediction performance. In Chapter 4, we used our ML-predicted postoperative ACD with a ray tracing-based IOL formula (OKULIX), and significantly improved its performance. In Chapter 5, we described a new machine learning-based formula, Nallasamy formula, that predicts the most appropriate IOL power based on the preoperative biometry alone. We showed that this new formula outperformed existing formulas including the Barrett Universal II formula, and the Emmetropia Verifying Optical (EVO) formula. In Chapter 6, we demonstrate the risks of using standard evaluation metrics for ML-based IOL formulas, and present two new metrics for more robust evaluation of IOL power prediction formulas. In Chapter 7, we summarize the main findings and discuss future directions. Together, the projects presented in this dissertation examined and proved the possibility of utilizing artificial intelligence in cataract surgery decision-making. These studies are among the pioneers of the use of ML in cataract surgery, and they will lay the foundation for the next era of cataract surgery planning.