Validation of neuron activation patterns for artificial intelligence models in oculomics.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Cardiovascular Health

Background:

• Oculomics leverages artificial intelligence (AI) to link retinal features with systemic health.
• Subtle retinal phenotypes used in oculomics models pose challenges for conventional AI interpretability tools like saliency maps.
• Existing neuron activation pattern (NAP) methods primarily focus on failure diagnosis in AI models.

Purpose of the Study:

• To design and validate a novel NAP framework for interpreting oculomics AI models.
• To assess the utility of NAPs in understanding AI predictions of systolic blood pressure from retinal images.
• To explore the correlation of NAPs with cardiovascular risk and biological distinctions within predicted blood pressure groups.

Main Methods:

• Developed a novel NAP framework for interpreting AI models in oculomics.
• Applied the NAP framework to an AI model predicting systolic blood pressure from fundus images.
• Utilized the United Kingdom Biobank dataset for model training and validation.

Main Results:

• The generated NAP correlated with the clinical endpoint of cardiovascular risk.
• The NAP framework successfully distinguished two biologically distinct groups among participants with the same predicted systolic blood pressure.
• Demonstrated the feasibility of the NAP framework for interpreting oculomics AI models.

Conclusions:

• The proposed NAP framework offers a viable method for gaining deeper insights into oculomics AI model functioning.
• NAP analysis provides a link between retinal imaging AI, systolic blood pressure prediction, and cardiovascular risk assessment.
• Further validation on external datasets is necessary to confirm the generalizability of the findings.