Deep learning model for risk assessment of myopic progression in children wearing orthokeratology lenses

To construct a risk stratification model for axial length elongation in orthokeratology lenses, facilitating the identification of high-risk patients. This study collected case data on Ortho-K lenses at the Ophthalmology Department and Optometry Center of Peking University People’s Hospital from January 2018 to May 2021. The Ortho-K cohort used for model development included 1,639 myopia patients (3,271 eyes), and the cohort of 216 patients (356 eyes) collected at different times was used for temporal verification. Hierarchical Attention Model (HAM) based on the transformer was trained, and performance was assessed for validation temporally. The HAM was designed based on attention mechanism, which could be embedded into each feature group independently, and combined with local and global attention mechanisms to deal with complex interactions in medical data. Permutation Importance and Shapley additive explanations analysis were used to interpret key contributing factors. The overall precision of HAM on the training set was 0.852, the recall rate was 0.853, and the F1 score was 0.852. The performance was obviously superior to other nine models. The model was exceptionally effective at identifying high-risk patients, achieving the corresponding area under the curve was the highest at 0.98. It maintained the clear advantage in temporal verification, which achieved a precision of 85.33%, recall of 75.12%, and an F1-score of 79.90%. Calibration curves and decision curve analysis showed good consistency (all Brier scores < 0.15) and clinical value. Permutation Importance and SHapley additive explanations analysis revealed that patients characterized by a younger age, less myopia, higher cylinder power, and a higher degree of parental myopia was prone to rapid axial length elongation. Those with a larger pupil diameter, less weight change, and an older age at diagnosis of myopia exhibited a lower risk of rapid axial elongation. The HAM, developed from routine ocular examinations and questionnaires, can rapidly identify patients at high risk of rapid myopia progression. Its high performance and temporal validation enhance the reliability. This allows for timely suggestions and adjustment of individualized myopia prevention and control measures in real world clinical settings.

Authors

Institutions

Publication Details

Journal
BMC Medical Informatics and Decision Making
Published
2026-11-19
DOI
https://doi.org/10.1186/s12911-026-03762-0
Primary Topic
Ophthalmology and Visual Impairment Studies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep learning model for risk assessment of myopic progression in children wearing orthokeratology lenses

Kai Wang, Shujian Gao, Zekuan Yu, Xuewei Li et al.
BMC Medical Informatics and Decision Making
Ophthalmology and Visual Impairment Studies
article

Deep learning model for risk assessment of myopic progression in children wearing orthokeratology lenses

Kai Wang, Shujian Gao, Zekuan Yu, Xuewei Li, Tonglei Zheng, Yan Li, Yuanqi Deng, Mingwei Zhao
article en

Abstract

To construct a risk stratification model for axial length elongation in orthokeratology lenses, facilitating the identification of high-risk patients. This study collected case data on Ortho-K lenses at the Ophthalmology Department and Optometry Center of Peking University People’s Hospital from January 2018 to May 2021. The Ortho-K cohort used for model development included 1,639 myopia patients (3,271 eyes), and the cohort of 216 patients (356 eyes) collected at different times was used for temporal verification. Hierarchical Attention Model (HAM) based on the transformer was trained, and performance was assessed for validation temporally. The HAM was designed based on attention mechanism, which could be embedded into each feature group independently, and combined with local and global attention mechanisms to deal with complex interactions in medical data. Permutation Importance and Shapley additive explanations analysis were used to interpret key contributing factors. The overall precision of HAM on the training set was 0.852, the recall rate was 0.853, and the F1 score was 0.852. The performance was obviously superior to other nine models. The model was exceptionally effective at identifying high-risk patients, achieving the corresponding area under the curve was the highest at 0.98. It maintained the clear advantage in temporal verification, which achieved a precision of 85.33%, recall of 75.12%, and an F1-score of 79.90%. Calibration curves and decision curve analysis showed good consistency (all Brier scores < 0.15) and clinical value. Permutation Importance and SHapley additive explanations analysis revealed that patients characterized by a younger age, less myopia, higher cylinder power, and a higher degree of parental myopia was prone to rapid axial length elongation. Those with a larger pupil diameter, less weight change, and an older age at diagnosis of myopia exhibited a lower risk of rapid axial elongation. The HAM, developed from routine ocular examinations and questionnaires, can rapidly identify patients at high risk of rapid myopia progression. Its high performance and temporal validation enhance the reliability. This allows for timely suggestions and adjustment of individualized myopia prevention and control measures in real world clinical settings.

BMC Medical Informatics and Decision MakingVol. 26(1)
Peking University (CN), Fudan University (CN), Peking University People's Hospital (CN)
Ophthalmology and Visual Impairment Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon