This article is more than 5 years old. Information may no longer be current.
Applying artificial intelligence to current IOL calculations
Refractive cataract surgery has the ability to neutralize the patient’s preexisting refractive error at the same time we address the lens opacity. Ideally, this is achieved all together at the same sitting by correctly determining the spherical IOL power and then also addressing the astigmatism. We can even implant IOLs that have multiple focal points, extended depth of field or perhaps even future accommodating designs, but the success of all of these hinges on the ability to have accurate IOL calculations.
By using just the keratometry and axial length, we are able to achieve about 80% of patients within 0.5 D of our intended refractive target using third-generation static formulae such as SRK/T, Holladay 1 and Hoffer Q. Our methods of IOL calculation have progressed over time, and we have learned that additional input variables such as anterior chamber depth and white-to-white measurements can further improve accuracy. Holladay 2, Haigis and Barrett Universal II all make use of this additional biometry and can produce increasingly better results. These formulae are static, however, and they cannot evolve on their own. Additional lines of code must be programmed for any desired improvement.
Artificial intelligence methods have the ability to evolve over time and to self-learn, thereby increasing accuracy as more data are obtained. We have all seen the power of this with complex strategy games such as chess or Go, in which Google’s AlphaZero now reigns as the grand champion, beating all humans. The AlphaZero algorithm was given the rules of these games and then it self-played for a few hours, making billions of hypothetical moves and learning the consequences. We can apply the same methodologies to IOL calculations, and this was shown by John Ladas, MD, PhD, who amalgamated multiple formulae and then applied deep machine learning for his Ladas 2.0 AI formula.
The Ladas 2.0 AI formula quickly proved to be more accurate. Using data sets to learn from multiple surgeons and then applying it to another surgeon, Ladas 2.0 AI gave an accuracy of 87%, beating the Barrett Universal II and the Holladay 2. Even better was using the same surgeon’s data to improve his own calculations, which resulted in 94% accuracy. So why wouldn’t everyone switch over? It turns out that surgeons know and trust one method, and then they are reluctant to change. This is where the PLUS method comes in: Precision Ladas Universal Super-algorithm, which allows us to use the same artificial intelligence methods to improve any existing formula.
The Barrett Universal II has gained popularity as one of the most accurate methods for IOL calculations. I routinely print out Barrett calculations for my cataract patients, and then I apply the PLUS methodology to improve accuracy. In a recent study of 1,400 eyes from a university setting, the results were amazing: All of the different AI methods that were applied by PLUS resulted in a significantly higher refractive accuracy with a lower mean absolute error (Figure). Note that the PLUS methodology is a proprietary form of AI that uses multiple methods and then cherry-picks for the best accuracy.
The methods of artificial intelligence that were used in this test of 1,400 eyes included one linear regression method and three non-linear regression methods. Linear regression (LR) is a linear estimation of the relationship between the dependent output variable and the independent input variables. The ridge regression method can be used to reduce the model complexity and prevent overfitting by limiting the magnitude of the coefficients.
Source: Uday Devgan, MD
Support vector regression (SVR) uses a support vector machine to do the regression in which input variables are projected into higher dimensional space to increase the complexity, allowing the inner non-linear relationships between the input and output to pop up after the projection. Extreme gradient boosting (EGB) ensembles many different prediction models and decision trees, both simple and complex, to reveal the relationship between input and output. Neural networks (NN) are built up by nodes and layers, each being a non-linear filter. Each layer is then built up by parallel nodes, and the connection between the layers is the mapping between nodes. The input will go through the network of these nodes, and the weights of the connections will be learned by training the neural network model. It is then the combination of these nodes that finally reveals the non-linear relationship between input and output.
While this seems complex, the PLUS method runs all of these calculations in the background without requiring user input and then chooses the best combination for highest accuracy. For the surgeon, the entire process is the same as or easier than what is currently being done. Ladas has also developed a self-calibrating biometer that automatically does these calculations at the time of biometry, and then in the postop period, this device remeasures the eye and sends the refractive outcome data back to the central computers to further increase accuracy. This is 1,000 times more detailed than simply honing an A-constant, a technique that is outdated by many years. While this device is under development and will be available in the near future, even today surgeons can access PLUS to improve their refractive accuracy.
Changing the A-constant will either raise or lower all calculations done by the same dioptric amount, and we know that this is erroneous because you cannot treat a small hyperopic eye the same way as a large myopic eye. The PLUS method, which is what drives the self-calibrating biometer, will allow adjusting for 10 ranges of axial length, 10 ranges of keratometry and 10 ranges of anterior chamber depth. This is a cubic grid of 1,000 blocks, each of which will be adjusted individually by the computer algorithm.
If any future biometric parameters are determined to be useful, such as equatorial lens position, posterior cornea or even angle geometry, the PLUS artificial intelligence method will allow easy integration of these variables and continued evolution over time. The future of IOL calculations will certainly involve machine learning methods that will evolve over time. And if you have any doubts, just go ahead and try to beat your computer at chess.
- For more information:
- Uday Devgan, MD, is in private practice at Devgan Eye Surgery, Chief of Ophthalmology at Olive View UCLA Medical Center and Clinical Professor of Ophthalmology at the Jules Stein Eye Institute, UCLA School of Medicine. He can be reached at 11600 Wilshire Blvd. #200, Los Angeles, CA 90025; email: devgan@gmail.com; website: www.CataractCoach.com.
Disclosure: Devgan reports he is a principal in Advanced Euclidean Solutions, which owns IOLcalc.com, the PLUS methodology and related intellectual property.