Clinical Scorecard: Bridging the Gap in Silent Stroke Detection: The Role of Deep Learning in Analyzing Retinal Images

At a Glance

Key Highlights

• DeepRETStroke AI system detects SBIs with 85.2% sensitivity, closing 78% of the gap between traditional retinal screening and MRI.
• Trained on ~900,000 unlabeled retinal images using a three-stage deep learning pipeline combining self-supervised, semi-supervised, and knowledge transfer methods.
• Validated across 12 multiethnic cohorts with robust performance in SBI detection and stroke risk prediction, including incident and recurrent strokes.

Guideline-Based Recommendations

Diagnosis

• Utilize retinal imaging analyzed by DeepRETStroke AI for non-invasive detection of silent brain infarctions.
• Consider retinal fundus photography as a scalable screening tool to identify covert cerebrovascular injury before symptomatic stroke.

Management

• Integrate DeepRETStroke-guided risk stratification into clinical practice to guide personalized stroke prevention strategies.
• Apply integrated management for high-risk patients identified by the AI to reduce recurrent stroke incidence.

Monitoring & Follow-up

• Use longitudinal retinal imaging and AI analysis to monitor cerebrovascular health and stroke risk over time.
• Perform ongoing validation and adaptation of AI models to maintain accuracy across diverse populations.

Risks

• Recognize potential generalizability limitations due to model training predominantly on Chinese cohorts.
• Be aware of possible biases in stroke prediction for underrepresented ethnic groups; consider federated learning approaches to mitigate.

Patient & Prescribing Data

Adults at risk for stroke, including those with silent brain infarctions, diabetes, or hypertension

DeepRETStroke-guided integrated management demonstrated an 82.44% relative reduction in recurrent strokes compared to conventional screening in a preliminary prospective study.

Clinical Best Practices

• Employ retinal fundus photography combined with DeepRETStroke AI for early detection of silent brain infarctions.
• Incorporate AI-driven stroke risk prediction into routine screening for high-risk populations to enable timely preventive interventions.
• Address population diversity by validating and adapting AI models using federated learning with multiethnic datasets.
• Use AI outputs to guide personalized management plans and monitor treatment efficacy longitudinally.

References

• Stroke global burden and prevention importance
• Silent brain infarctions prevalence and stroke risk
• DeepRETStroke AI system development and validation
• Retinal imaging as a window to cerebrovascular health
• Federated learning for AI generalizability