A study of the predictive effects of machine learning for the relationship between axial length elongation and the progression of myopia in school-aged children

Authors: Tang Tao,  Fan Yuzhuo,  Xu Qiong,  Peng Zisu,  Wang Kai,  Zhao Mingwei

DOI: 10.3760/cma.j.issn.2095-0160.2020.02.010
Published 2020-02-10
Cite as Chin J Exp Ophthalmol, 2020,38(02): 134-139.

Abstract                              [View PDF] [Read Full Text]

Objective

To investigate the relationship between axial length (AL)elongation and the progression of spherical equivalent refraction (SER) and its influential factors in school-aged children with myopia based on machine learning (ML).

Methods

A cross-sectional study evaluated 1 011 eyes of school-aged myopic children admitted to the optometry center of Peking University People’s Hospital from January 2017 to December 2018, and data from the right eyes were used for analysis.All the collected data were used to train ML algorithms.When building predictive models, the input features included age, gender, central corneal thickness (CCT), mean K readings (K-mean), horizontal visible iris diameter (HIVD), lens power, and axial length (AL), and the output parameter was SER.A five-fold cross validation scheme randomly divided all the data into five groups, of which four were used as training data, and one group was used as validation data.This process was repeated five times so that all the data were validated by this model, which allowed a better prediction of the overall sample.The prediction accuracy of different models was evaluated by the R-value and R2.The best-performing algorithm was applied to investigate the relationship between AL elongation and the progression of SER and its influencing factors.Written informed consent was obtained from each guardian of each patient prior to entering the study cohort.This study followed the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Peking University People’s Hospital (No.2019PHB280-01).

Results

In the comparison of the R-value and R2 of six ML algorithms based on five-fold cross validation, among all models, the best was the quadratic SVM regression model, with an R-value and R2 of 0.99 and 0.98, respectively.The results of Pearson correlation analysis showed that lens power was negatively correlated with age (r=-0.301, P<0.01). According to the results calculated by the Bennett-Rabbetts formula, the average lens power of the 6-year-old myopic group was higher than that of the 18-year-old myopic group.According to this model, the SER change caused by AL elongation was not a constant value, which was estimated from plano to nearly -3.00 D, depending mainly on the time needed for 1-mm AL elongation.According to the results calculated by this model, the longer it took for the AL to grow by 1 mm, the smaller the corresponding SER change.In myopic children over an age span of one year, for example, from 6-7 years or 12-13 years, 1-mm elongation of the AL corresponded to -2.50 D and -2.33 D of SER change, respectively.Over a three years span, for example, from 6-9 years, a 1-mm elongation of the AL corresponded to -1.77 D of SER change.

Conclusions

For myopic children, the longer the age span required for 1-mm elongation of the AL, the smaller the SER change.An ML algorithm can provide clinical practitioners with a relatively precise estimation for the relationship between AL elongation and myopia progression.

Key words:

Myopia; Machine learning; Axial length; Spherical equivalent refraction

Contributor Information

Tang Tao
Department of Ophthalmology & Clinical Centre of Optometry, Peking University People’s Hospital, Eye Diseases and Optometry Institute, Beijing Key Laboratory of Diagnosis and Therapy of Retinal and Choroid Diseases, College of Optometry, Peking University Health Science Center, Beijing 100044, China
Fan Yuzhuo
Xu Qiong
Peng Zisu
Wang Kai
Zhao Mingwei
(Read 68 times, 1 visits today)
Updated: December 23, 2022 — 7:08 am