A deep learning system for detecting silent brain infarction and predicting stroke risk

Current brain imaging to detect silent brain infarctions (SBIs) is not feasible for the general population. Here, to overcome this challenge, we developed a retinal image-based deep learning system, DeepRETStroke, to detect SBI and refine stroke risk. We use 895,640 retinal photographs to pretrain the DeepRETStroke system, which encodes a domain-specific foundation model for representing eye-brain connections. Then, we validated the downstream clinical tasks of DeepRETStroke using 213,762 retinal photographs from diverse datasets across China, Singapore, Malaysia, the USA, the UK and Denmark to detect SBI and predict stroke events. DeepRETStroke performed well in internal validation datasets, with areas under the curve of 0.901 for predicting incident stroke and 0.769 for predicting recurrent stroke. External validations demonstrated consistent performances across diverse datasets. Finally, in a prospective study comprising 218 participants with stroke, we assessed the performance of DeepRETStroke compared with clinical traits in guiding strategies for stroke recurrence prevention. Altogether, the retinal image-based deep learning system, DeepRETStroke, is superior to clinical traits in predicting stroke events, especially by incorporating the detection of SBI, without the need for brain imaging.

Authors

Institutions

Publication Details

Journal
Nature Biomedical Engineering
Published
2025-06-06
DOI
https://doi.org/10.1038/s41551-025-01413-9
Citations
19
Primary Topic
Acute Ischemic Stroke Management
Type
article
Field-Weighted Citation Impact
25.73

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
article

A deep learning system for detecting silent brain infarction and predicting stroke risk

Haotian Zheng, Zhouyu Guan, Weiping Jia, Jocelyn Hui Lin Goh et al.
19 citations
Nature Biomedical Engineering
Acute Ischemic Stroke Management
25.73
article

A deep learning system for detecting silent brain infarction and predicting stroke risk

Haotian Zheng, Zhouyu Guan, Weiping Jia, Jocelyn Hui Lin Goh, Charumathi Sabanayagam, Yuqian Bao, Ya Xing Wang, An Ran Ran, Huating Li, Ten Cheer Quek, Carol Y. Cheung, Xiaokang Yang, Yingfeng Zheng, Xiaolong Yang, Jakob Grauslund, Shiqun Lin, Frederik Nørregaard Pedersen, Ching‐Yu Cheng, Hongwei Ji, Yuchen Guo, Changchang Yin, Minyi Guo, Tingyao Li, Chan Wu, Sarkaaj Singh, Ruth Hogg, Ting‐Li Chen, Ping Zhang, Lei Ruan, Yanwen Zhou, Yuwei Lu, Gareth J. McKay, Rongping Dai, Weiwen Zou, Dawei Yang, David M. Wright, Tien Yin Wong, Xiaoer Wei, Chenxin Deng, Gavin Siew Wei Tan, Jie Shen, Qionghai Dai, Nan Jiang, Zixiao Li, Yih Chung Tham, Benjamin Sommer Thinggaard, Lee‐Ling Lim, Anis Syazwani Abd Raof, Dan Liu, Jing Wang, Yongjun Wang, Seungmin Lee, Yixiao Jin, Yuehua Li, Shan Huang, Yiming Qin, Zheyuan Wang, Jia Shu, Cuntai Zhang, Tingting Hu, Bin Sheng, Yu Zhao, Miao Li Chee, Haslina Hamzah, Shujie Yu, Yi-Lan Wu, Qingsheng Peng, Yuesong Pan, Shiyu Wang, Jian Wen Samuel Lee-Boey
article en
19 citations

Abstract

Current brain imaging to detect silent brain infarctions (SBIs) is not feasible for the general population. Here, to overcome this challenge, we developed a retinal image-based deep learning system, DeepRETStroke, to detect SBI and refine stroke risk. We use 895,640 retinal photographs to pretrain the DeepRETStroke system, which encodes a domain-specific foundation model for representing eye-brain connections. Then, we validated the downstream clinical tasks of DeepRETStroke using 213,762 retinal photographs from diverse datasets across China, Singapore, Malaysia, the USA, the UK and Denmark to detect SBI and predict stroke events. DeepRETStroke performed well in internal validation datasets, with areas under the curve of 0.901 for predicting incident stroke and 0.769 for predicting recurrent stroke. External validations demonstrated consistent performances across diverse datasets. Finally, in a prospective study comprising 218 participants with stroke, we assessed the performance of DeepRETStroke compared with clinical traits in guiding strategies for stroke recurrence prevention. Altogether, the retinal image-based deep learning system, DeepRETStroke, is superior to clinical traits in predicting stroke events, especially by incorporating the detection of SBI, without the need for brain imaging.

Nature Biomedical EngineeringVol. 9(11)
Queen's University Belfast (GB), Beijing Tongren Hospital (CN), Sun Yat-sen University (CN), National University of Singapore (SG), Chinese University of Hong Kong (HK), Shanghai Jiao Tong University (CN), Capital Medical University (CN), Chinese Academy of Medical Sciences & Peking Union Medical College (CN), Singapore National Eye Center (SG), Peking Union Medical College Hospital (CN), Odense University Hospital (DK), University of Malaya (MY), National Clinical Research Center for Digestive Diseases (CN), Asia Diabetes Foundation (CN), Singapore Eye Research Institute (SG), Duke-NUS Medical School (SG), Shanghai Sixth People's Hospital (CN), National University Health System (SG), Beijing Tsinghua Chang Gung Hospital (CN), Tongji Hospital (CN), Huazhong University of Science and Technology (CN), The Ohio State University (US), Ministry of Industry and Information Technology (CN), Tsinghua University (CN)
National Natural Science Foundation of China, China Postdoctoral Science Foundation, Tsinghua University, Shanghai Jiao Tong University, National Science and Technology Major Project, Natural Science Foundation for Young Scientists of Shanxi Province, Young Scientists Fund
Openalex Percentile: Top 0%
Acute Ischemic Stroke Management
25.73
DOIOpen Access PDFOpenAlex
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.