Purpose: To develop a new ectasia risk model through artificial intelligence (AI) and machine learning, enabling the creation of an integrated method without a cut-off point per risk factor, and subsequently better at differentiating patients at higher risk of ectasia with normal topography.
Methods: This comparative case-control study included 339 eyes with normal preoperative topography, with 65 eyes that developed ectasia after laser in situ keratomileusis (ectasia group) and 274 eyes that did not develop ectasia (control group). The AI model used known risk factors to engineer 14 additional ones, totaling 20 features. In this methodology, no variable is used in isolation because its cut-off point is never considered. All separation between cases and controls is made through the interaction detected by the machine learning model that gathers the variables considered relevant. The ability to correctly separate ectatic cases identified as high risk, ectatic cases wrongly classified as low risk, and controls were illustrated by the diagram t-distributed stochastic neighbor embedding (t-SNE).
Results: Only two original variables (percent tissue altered and corneal thickness) and two derived from the feature engineering process (derivative percent tissue altered and age weighted value) were selected by the final AI model (ie, best performing AI-based model to separate patients at higher risk). The t-SNE visualization demonstrated the greater ability to differentiate between patients considered at risk by the AI-based model, without a cut-off point, compared to all other methods used alone (P < .0001).
Conclusions: This study describes a new AI-based model that integrates different risk factors without a cut-off point, increasing the number of cases correctly identified as at higher risk. [J Refract Surg. 2022;38(11):716-724.].