Unveiling Common LASIK Complications: A Machine-Learning Based Analysis of FDA Reported Cases

Purpose: This study explores complications associated with Laser-Assisted In Situ Keratomileusis (LASIK) surgery by applying natural language processing (NLP) and machine learning to a dataset of FDA-reported adverse events. The goal is to identify common complication patterns and predictors to support patient safety, preoperative counseling, and clinical decision-making. Methods: A retrospective analysis was performed using the publicly available “LASIK Complications” dataset from Kaggle, which includes 1,368 reports submitted to the U.S. Food and Drug Administration (FDA) between June 1997 and April 2019. Text data were cleaned, tokenized, and vectorized using term frequency-inverse document frequency (TF-IDF). A Random Forest classifier was developed and tuned through three-fold cross-validation. Model performance was assessed using accuracy, kappa statistic, sensitivity, specificity, and feature importance metrics. Results: The model achieved a cross-validation accuracy of 48.76% and a test accuracy of 28.21% (κ = 0.1988, p = 0.2661). Among 19 complication types, the Visual Symptoms category showed the highest sensitivity (76.39%) and specificity (85.57%). Feature importance analysis revealed that terms like “halos,” “dry,” and “night” were the most influential descriptors, aligning with well-known post-LASIK experiences. In contrast, complications such as infection and corneal flap issues showed no sensitivity. Conclusion: While overall accuracy was limited, the model successfully identified clinically relevant descriptors, highlighting the potential of NLP and machine learning in analyzing patient-reported outcomes. Future work should address class imbalance, enrich contextual features (e.g., n-grams), and apply explainable AI tools like SHAP or LIME to improve clinical adoption and interpretability.