To provide objective and efficient technical support for orthokeratology lens fitting and address the issues of strong subjectivity,poor repeatability,and low efficiency in clinical evaluation of lens fitting status,decentration,and myopia control effect,we developed a comprehensive artificial intelligence(AI)-based evaluation scheme covering the entire orthokeratology lens fitting process.This retrospective study included 360 lens fitting evaluation videos,corneal topography data from 200 eyes,and follow-up data from 115 left eyes of eligible pediatric myopia patients who received orthokeratology treatment.Strict inclusion and exclusion criteria were applied to ensure the homogeneity of research subjects and minimize potential confounding factors,thereby laying a solid foundation for the reliability and credibility of subsequent experimental results.
Three specialized AI analytical methods were employed to address key clinical challenges at different stages of orthokeratology fitting:first,a deep learning algorithm integrated with spatiotemporal attention mechanisms classified fitting status into appropriate,tight or loose fitting via a two-stage classification strategy,effectively improving evaluation accuracy by focusing on key image features related to lens fitting.Second,K-means clustering combined with the weighted least squares method was used to analyze defocus rings on corneal topography,fit optimal ellipses for the defocus ring area,and quantitatively determine the degree of lens decentration,enabling objective and quantitative assessment of lens position.Third,a hybrid prediction model integrating multiple linear regression and a temporal convolutional network(TCN)was established to predict 12-month axial length elongation,which could simultaneously capture both linear and non-linear relationships between influencing factors and axial growth.
The fitting evaluation algorithm achieved over 80%accuracy,sensitivity,and specificity in both classification stages,with 80%to 83%correct classification for each status and excellent repeatability.For decentration evaluation,AI results were highly consistent with senior clinicians' consensus(83%agreement)and significantly more repeatable than manual evaluation.Regarding the prediction of myopia control effect,Pearson correlation analysis showed that age(r=-0.30,P<0.001),spherical equivalent refraction(r=0.20,P=0.032),and gender(r=0.19,P=0.037)were significantly correlated with axial length elongation.In contrast,other baseline ocular biological parameters(e.g.,pupil diameter,corneal curvature)showed no significant correlation with axial growth(all P>0.1).Additionally,changes in corneal topographic parameters,including defocused height,defocused volume,and treatment zone area,were also identified as important factors associated with axial length elongation(all P<0.01).The hybrid prediction model,built based on key baseline factors and dynamic changes in corneal topography,showed good predictive validity,with its root mean square error(RMSE),mean absolute error(MAE),and coefficient of determination(R2)all superior to those of the traditional linear regression model in both the training and test sets.
In conclusion,this AI-based evaluation scheme demonstrates high accuracy,good repeatability,and strong clinical applicability across all key stages of the entire orthokeratology lens fitting process.It can effectively realize intelligent auxiliary evaluation in clinical fitting,follow-up,and myopia control effect prediction,reduce human errors caused by subjective judgment,improve the efficiency and standardization of orthokeratology fitting,and provide a reliable technical tool for personalized myopia management in children and adolescents,thus holding important clinical application value and broad promotion prospects in pediatric myopia prevention and control.