ASCRS 2025: Joaquin De Rojas, MD, leverages machine learning model to predict arcuate outcomes

Joaquin De Rojas, MD, shared highlights of his presentation on implementing machine learning to refine arcuate keratotomy (AK) nomograms, which earned him the Best Paper of Session award at the 2025 American Society of Cataract and Refractive Surgery annual meeting, held April 25 to 28, in Los Angeles, California.

“In an effort to make the predictions of arcuate keratotomies better, we leveraged machine learning to help understand this complex interplay of multiple… ‘features’,” De Rojas explained, underscoring the model’s ability to incorporate a broader range of input variables than traditional methods.

Central to the model’s success was the concept of treated astigmatism—calculated by subtracting postoperative refractive error from preoperative astigmatism, aligned on the same axis using basic trigonometry. “This is the amount of astigmatism that was actually treated with the arcuate. And that’s how you want to train the model,” he said. Treated astigmatism emerged as the most predictive feature for incision length, followed closely by preoperative astigmatism and patient age.

Notably, the model also identified secondary but influential variables such as white-to-white distance, laterality, handedness, sex, axial length, and even device-specific differences in biometry. These nuanced factors are often overlooked in current nomograms. “A wider cornea… that same arcuate incision at 9 mm diameter is going to have a different effect,” De Rojas explained.

The machine learning framework was built on XGBoost, a gradient boosting algorithm well-suited to tabular data, offering robustness with limited datasets while avoiding overfitting. While neural networks were also explored, they proved too data-hungry and erratic without stringent controls. “You have to find the best balance… more sophistication isn’t always better,” he said.

Clinicians can test the live model at DeRojas.info, where two versions of the tool offer actionable guidance, including advising against AKs for high against-the-rule astigmatism. “You just have to get some good data… and within seconds, you have a customized, tailored approach," he added.