Ophthalmology. Retina · 2026
Multimodal Artificial Intelligence for Predicting 3- and 5-Year Risks of Myopic Choroidal Neovascularization in High Myopia.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Wang Yining, Ishida Takashi, Wang Ziye, Wu Yijin, Kamoi Koju, Wei Ting Daniel Shu, Ohno-Matsui Kyoko
Аннотация
PURPOSE: To develop artificial intelligence-based models for estimating the risk of myopic macular neovascularization (MNV) in highly myopic eyes.
DESIGN: Retrospective, single-center cohort study.
PARTICIPANTS: A total of 4235 eyes from 2501 patients with high myopia who visited the Institute of Science Tokyo between October 2011 and May 2021 were included for analysis.
METHODS: Baseline fundus photographs and 7 clinical variables were used for model development and validation. Three image-based deep learning (DL) models with a DenseNet-121 backbone were trained to predict the risk of MNV within 1, 3, and 5 years after baseline. A multimodal survival model based on the DeepSurv framework was further developed by integrating image-derived features with clinical variables to estimate time-to-event risk. Model performance was evaluated using discrimination, calibration, and clinical utility metrics. Model interpretability was explored using gradient-weighted class activation mapping and SHapley Additive exPlanation analyses.
MAIN OUTCOME MEASURES: Area under the receiver operating characteristic curve for the DL models and concordance index for the multimodal DeepSurv model.
RESULTS: The DL models achieved area under the receiver operating characteristic curves of 0.785 (95% confidence interval [CI], 0.474-0.988), 0.798 (95% CI, 0.687-0.900), and 0.729 (95% CI, 0.614-0.832) for the 1-, 3-, and 5-year predictions, respectively. The multimodal DeepSurv model achieved a concordance index of 0.683 (95% CI, 0.582-0.775) and enabled stratification of eyes into subgroups with distinct long-term probabilities of MNV development. Heatmaps showed activation in clinically relevant regions, and SHapley Additive exPlanation analysis indicated that axial length and pathologic myopia category showed prominent conditional contributions within the multimodal model.
CONCLUSIONS: Artificial intelligence-based models enabled fixed-horizon risk estimation and longitudinal time-to-event risk assessment for MNV in highly myopic eyes, providing clinically meaningful risk stratification. With further external validation, this approach may support risk-adapted long-term monitoring strategies in high myopia.
FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
DESIGN: Retrospective, single-center cohort study.
PARTICIPANTS: A total of 4235 eyes from 2501 patients with high myopia who visited the Institute of Science Tokyo between October 2011 and May 2021 were included for analysis.
METHODS: Baseline fundus photographs and 7 clinical variables were used for model development and validation. Three image-based deep learning (DL) models with a DenseNet-121 backbone were trained to predict the risk of MNV within 1, 3, and 5 years after baseline. A multimodal survival model based on the DeepSurv framework was further developed by integrating image-derived features with clinical variables to estimate time-to-event risk. Model performance was evaluated using discrimination, calibration, and clinical utility metrics. Model interpretability was explored using gradient-weighted class activation mapping and SHapley Additive exPlanation analyses.
MAIN OUTCOME MEASURES: Area under the receiver operating characteristic curve for the DL models and concordance index for the multimodal DeepSurv model.
RESULTS: The DL models achieved area under the receiver operating characteristic curves of 0.785 (95% confidence interval [CI], 0.474-0.988), 0.798 (95% CI, 0.687-0.900), and 0.729 (95% CI, 0.614-0.832) for the 1-, 3-, and 5-year predictions, respectively. The multimodal DeepSurv model achieved a concordance index of 0.683 (95% CI, 0.582-0.775) and enabled stratification of eyes into subgroups with distinct long-term probabilities of MNV development. Heatmaps showed activation in clinically relevant regions, and SHapley Additive exPlanation analysis indicated that axial length and pathologic myopia category showed prominent conditional contributions within the multimodal model.
CONCLUSIONS: Artificial intelligence-based models enabled fixed-horizon risk estimation and longitudinal time-to-event risk assessment for MNV in highly myopic eyes, providing clinically meaningful risk stratification. With further external validation, this approach may support risk-adapted long-term monitoring strategies in high myopia.
FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.