Extended Data Fig. 4: The Workflow of the Development of DeepRETStroke System.

The DeepRETStroke system, which was built with a three-stage pretraining strategy, includes an encoder, an SBI Detector, an SBI Learner and a Stroke Predictor. For the first stage, a retinal feature encoder from RETFound were employed to perform self-supervised pretraining on SIM and CNDCS dataset. For the second stage, SDPP dataset was employed to roughly train the encoder and Stroke Predictor. For the third stage, SDPP-MRI dataset was employed along with SDPP dataset to train SBI Detector by semi-supervised learning strategy. The developed SBI Detector was then used to generate the ‘soft label’ for each sample in SDPP dataset. Then, the encoder, SBI Learner and Stroke Predictor were jointly trained on samples in SDPP dataset with their ‘soft label’ of SBI and ground truth label of future stroke. Finally, the developed encoder and Stroke Predictor were fine-tuned on population of NDSP dataset with stroke history to develop a specific model for recurrent stroke prediction. Of note, SBI detection and incident stroke prediction can be iteratively repeated for further optimization of the entire system. Here we highlighted the modules trained in each stage to provide a clearer explanation of the development process.