Deep multimodal learning based management of myopia in children
Description: Objectives: We aimed to 1) predict cumulative myopic shift in SER in advance of 3 years, and classify children into high or low risk of myopia progression; 2) predict SER after a 2-year low concentration atropine treatment, and to classify them into good or poor response to low-concentration atropine treatment.
Hypothesis: We hypothesize that deep learning models are able to robustly predict cumulative myopic shift in SER and classify children into high or low risk of myopia progression. Furthermore, the models are able to predict SER after a 2-year low concentration atropine treatment and classify children into good or poor response to low concentration atropine treatment based on clinical and imaging data.
Design and subjects: Retrospective. Our training and validation datasets for aim 1 will come from the HKCES with more than 30,000 recruited subjects. The training and validation for aim 2 will come from the LAMP study with 912 randomized children based on dosage of atropine eye drop.
Main outcome measures: For aim 1, the outcomes are predicting SER and diagnostic performance of classifying low or high risk of myopia progression. For aim 2, the outcomes are predicting SER after atropine treatment and diagnostic performance of classifying good or poor response to atropine treatment.
Data analysis: We will calculate RMSE and MAE between the predicted SERs and the ground truth values for the first outcome, and AUROCs will be used to evaluate the diagnostic performance.
Expected results: The system will achieve RMSE less than 0.1 and AUROCs over 0.7.
Funder: Health and Medical Research Fund
Amount: HKD 1,390,000
Funding type: grant
Funder scheme: HMRF - Advanced Medical Research
Start date: 2022-01-01
Investigator: Jason Cheuk-sing YAM
Institution: Department of Ophthalmology and Visual Sciences, The Chinese University of Hong Kong
Funded outputs count: 19
Funder award ID: 11220206
OpenAlex ID: G127482126