[A machine learning model on orthokeratology lens fitting based on the data of optometry examination]

Objective: To get an orthokeratology lens fitting model according to the research of the optometry examination data, which can help to improve the work efficiency and increase the hitting rate of prescription. Methods: The relationship between the basic optometry examination data and the effective optometry prescription was evaluated. We got 1 467 sets of data, including 80%(1 173) for training and 20%(294) for testing. The trail results were used for the 20% testing sets, and 100% completed sets for verification. With the contrast of accuracy, we got the influencing variables and the fitting model. Logistic regression and linear regression analysis were applied to the training set for building a fitting model, and the receiver operating characteristic curve (ROC curve) was drawn to determine the optimal threshold of the logistic regression model (1.0).The trail results are used for the 20% testing sets, 100% completed sets for verification. With the contrast of accuracy (ACC), we get the influencing variables and the fitting model. Results: Kf-FK and Ks-SK were selected as dependent variables. Through analysis of Stepwise regression, we found the influencing variables of Kf-FK were Kf(P=0.01), E(mean) (P=0.027), and CYL(P<0.001), with FK=6.23+0.866kf-1.69 E(mean)+0.16CYL, and the precision rate was 81%(r(2)=0.95). The influencing variables of Ks-SK were Ks (P<0.001), CYL (P=0.00), and E(mean) (P<0.001), with SK=6.84+0.862ks-0.28E(mean)-1.3CYL, and the precision rate was 80%(r(2)=0.93). By selecting the logistic method and decision tree analysis for glasses design, the influencing variables were KS (P<0.001)and CYL (P<0.001). The logistic function was Design=-45.7+0.82Ks+6.027CYL, and the precision rate was 87%. Conclusions: The combination of AI and medical data can verify the expert experience and guide the optometrist to locate the prescription quickly and accurately. Try to break the traditional mode of fitting with wearing the orthokeratology lens shortly, and it is possible for us to position the prescription quickly through the calculation of a computer. (Chin J Ophthalmol, 2019, 55: 105-110).