Application Of Machine Learning In Predicting Refractive Outcome After Lasik Surgery

To train and validate a machine learning model using preoperative and surgery-related data to predict refractive deviation from target sphere after LASIK for myopia.

Optegra Eye Health Care and University of Manchester

Retrospective data from myopic patients undergoing LASIK were included. Refractive, morphological, and surgical parameter data stored on the electronic medical record system and/or diagnostic devices, were extracted for analysis. 703 eyes from 447 patients were included in this preliminary study. After data extraction, cleaning and preprocessing, 22 features related to preoperative refractive data, topography data, and surgical parameters (flap thickness) were fed into the automated machine learning (AutoML) to find the best model for classifying the postoperative sphere deviation from its targeted value (Sph_Deviation). The selected model was optimised and trained to predict the Sph_Deviation of unseen data.

Classifying Sph_Deviation into three classes [within ±0.25D, within ±0.50D, outside ±0.50D], and using AutoML  to automatically check the performance of multiple models, ExtraTreesClassifier model showed the highest performance. Optimising it resulted in correct class predictions of 72% of unseen test cases, i.e. overall accuracy of 0.72. Performance metrics (Precision, Recall, F1-score) highlighted that the model excels at predicting Sph_Deviation that falls within ±0.25D with a 93% of true positive rate.

Machine learning models, trained on historical data, hold promise for optimising surgical planning and patient counselling to reduce complications and improve patient satisfaction. Our model accurately predicts outcomes within ±0.25D of the target, facilitating pre-operative enhanced decision making and targeted interventions. Further studies are warranted to refine predictions for wider ranges, incorporating more data and clinical factors. Future expansions will include additional post-surgery metrics, assessing the model's clinical utility and expanding its applicability to other surgery types.