Artificial Intelligence Applications in Laser-Assisted Refractive Surgery—A Comprehensive Review

Background/Objectives: Artificial Intelligence (AI) technology—in the form of Machine Learning (ML) models and other models specialized in image recognition and processing, such as Convolutional Neural Networks Visual Transformers (CNN-ViTs)—is increasingly being used in ophthalmology. This review summarizes the main AI applications for laser-assisted refractive surgery. Methods: A wide-ranging literature search on the PubMed and ISI Thompson Web of Science databases was conducted. Results: AI is an all-encompassing technological concept that refers to either specific applications of ML/CNN models, such as in keratoconus screening or diagnosis, or agentic/multimodal AI applications that process different data types using a variety of AI models. The most clinically established applications found were corneal ectasia and keratoconus screening, in which AI models can reliably exceed 95% diagnostic accuracy, specificity, and sensitivity; nomogram-derived laser ablation programming; optimization of ray-tracing calculations for ray-traced ablation profiles; patient screening and selection for refractive surgery; visual and refractive result prediction; assessment of postoperative corneal ectasia risk; detection of intraoperative complications; research on the laser coupling effect; and other applications such as monitoring eye rubbing behavior in corneal ectasia suspects. Thus, singular, specific applications of AI models have made the largest clinical contributions. These results must be interpreted with a comparison of internal versus external validation, calibration, and generalizability. Overall the technology is highly promising. Agentic/multimodal AI frameworks are currently being discussed, with pilot studies published and applicability proposed in the context of large, high-patient-volume clinical practices. The technological advent of SHAP force plot graphs and saliency maps has improved understanding of the factors that contribute mechanistically to AI model reasoning, leading to more informed clinical decisions or research applicability, such as in finding parameters or factors linked to postoperative ectasia. Conclusions: AI is already established in laser-assisted refractive surgery in the form of models specialized in data-intensive analysis. At present, the best-performing novel AI models originate from the domain of large or multi-hospital settings due to requirements for multiple latest-generation diagnostic devices, high-volume patient data, computing expertise, and resources. Future directions could include clinical standardization and trickle-down dissemination of the technology. Agentic/multimodal AI has the potential to benefit large clinical/patient settings or offer access to expert knowledge in smaller practice settings. Agentic/multimodal AI implementation is offset by challenges in complexity, the technological limitations of LLMs (“hallucinations”), and ethical and legislative factors.