Preoperative predictive factors for opaque bubble layer formation and area during small-incision lenticule extraction: predictive models based on machine learning.

Abstract

Background: We aimed to develop exploratory machine learning (ML)-based predictive models for the occurrence and area of an opaque bubble layer (OBL) during small-incision lenticule extraction (SMILE) and identify associated preoperative and surgical-planning factors.

Methods: This retrospective study included 216 eyes (72 with an OBL, 144 controls) that underwent SMILE at the Zhongshan Ophthalmic Center between August 2024 and August 2025, which were matched 1:2 by age and sex. Comprehensive preoperative ocular examinations were performed using a Pentacam HR camera system and standard equipment. The OBL area was quantified from intraoperative videos using ImageJ software. Fifty-four preoperative and surgical features were used to develop the ML models: 15 classification algorithms for OBL occurrence, and 19 regression algorithms for the relative OBL area. Shapley additive explanations analysis was applied for model interpretability.

Results: The Extra Trees model achieved optimal performance for predicting OBL occurrence (area under the receiver operating characteristic curve = 0.885, accuracy = 0.820). The top five key factors included femtosecond laser energy, intraocular pressure, residual stromal thickness, 10-mm corneal volume, and total astigmatism, all of which showed positive correlations. For OBL area prediction, the random forest regression model performed best in the test sets (mean absolute error = 2.89%, root mean square error = 3.37%). Corneal optical density (within the central 2-mm zone of the anterior 120-μm corneal layer) and age were negatively associated with OBL area, whereas keratoconus index showed the strongest positive association.

Conclusion: The ML models showed exploratory predictive ability for OBL occurrence and area during SMILE using preoperative and surgical-planning parameters. External validation and recalibration in consecutive cohorts with a natural OBL prevalence are required before clinical use.