Construction and validation of an XGBoost-based predictive model for orthokeratology lens ordering parameters
10.3980/j.issn.1672-5123.2026.11.21
- VernacularTitle:基于XGBoost的角膜塑形镜订单参数预测模型构建及验证
- Author:
Lei DING
1
;
Erying HAN
1
;
Shuna ZHAO
1
;
Dan WU
1
Author Information
1. Department of Optometry and Ophthalmology, Puyang Second People's Hospital;Puyang Eye Hospital, Puyang 457000, Henan Province, China
- Publication Type:Journal Article
- Keywords:
orthokeratology lens;
machine learning;
extreme gradient boosting (XGBoost);
corneal topography;
parameter prediction
- From:
International Eye Science
2026;26(11):1991-1997
- CountryChina
- Language:Chinese
-
Abstract:
AIM:To explore the feasibility of predicting orthokeratology lens ordering parameters(flat K, spherical power, lens diameter, and lens toricity)using the extreme gradient boosting(XGBoost)algorithm based on corneal topographic parameters and basic clinical examination indicators, and to provide references for intelligent clinical fitting.METHODS:Clinical data were retrospectively collected from myopic patients who underwent initial orthokeratology lens fitting and achieved successful fitting at Puyang Second People's Hospital between January 2019 and March 2024. Data from right and left eyes were treated as independent samples, and the dataset was partitioned at the patient level into training and test sets in a 7∶3 ratio. Input features included uncorrected visual acuity(UCVA), axial length(AL), horizontal visible iris diameter(HVID), flat K, steep K, horizontal eccentricity, and vertical eccentricity. The lens toricity prediction model additionally incorporated keratometric astigmatism(ΔK)and surface asymmetry index(SAI). The output labels were the final ordering parameters(flat K, spherical power, lens diameter, and lens toricity). Separate XGBoost regression models were constructed for each output parameter, with hyperparameters optimized independently through grid search combined with 5-fold cross-validation. Model performance was evaluated on the test set using the coefficient of determination(R2), mean absolute error(MAE), and root mean square error(RMSE). Feature importance analysis was also performed.RESULTS:Totally 400 patients(800 eyes)were included in the final analysis. The 400 patients comprised 208 male(52.0%)and 192 female(48.0%)patients, with a mean age of 10.1±2.3 y. After partitioning at the patient level, the training set comprised 280 patients(560 eyes; 146 males and 134 females; mean age 10.1±2.3 y), and the test set comprised 120 patients(240 eyes; 62 males and 58 females; mean age 10.0±2.2 y). No statistically significant differences in sex composition or age were found between the two sets(P>0.05). Toric lenses were used by 140 of the 280 patients(50.0%)in the training set and by 50 of the 120 patients(41.7%)in the test set(χ2=2.017, P=0.156). The flat K prediction model achieved an MAE of 0.34 D and an R2 of 0.84. The spherical power model attained an MAE of 0.43 D and an R2 of 0.70. The lens diameter model yielded an MAE of 0.12 mm and an R2 of 0.42. The lens toricity model(Tweedie regression, power=1.5)achieved an MAE of 0.44 D and an R2 of 0.38. The MAE for each parameter fell within the clinically acceptable tolerance limits(flat K≤0.50 D, spherical power ≤0.75 D, lens diameter ≤0.25 mm, lens toricity ≤0.75 D). Scatter plots and Bland-Altman analysis demonstrated strong agreement between predicted values and actual observations. Feature importance analysis revealed that flat K(0.70)and steep K(0.11)were the most influential predictors in the flat K prediction model.CONCLUSION: The XGBoost model based on routine clinical parameters provides predictions of flat K, spherical power, and lens diameter with MAEs within clinically acceptable ranges, and offers useful reference for cylindrical power prediction. This approach may serve as a useful adjunct to clinical lens fitting. Nevertheless, future prospective studies are warranted to validate its clinical utility in improving actual fitting success rates.
