Non-Contact Optical Blood Pressure Biometry Using AI-Based Analysis of Non-Mydriatic Fundus Imaging

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Assessing blood pressure non-invasively is crucial for managing hypertension.
• Traditional methods rely on arm cuff measurements, which can be inconvenient.
• Exploring alternative, less invasive methods for blood pressure assessment is an active area of research.

Purpose of the Study:

• To develop and evaluate a machine learning model for blood pressure assessment.
• To determine if the model's accuracy is comparable to standard arm cuff measurements.
• To investigate the potential of using fundus images for blood pressure estimation.

Main Methods:

• A deep learning model was developed using the UK Biobank dataset.
• The model was trained to predict both systolic and diastolic blood pressure.
• Performance was evaluated against arm cuff measurements using Mean Absolute Error, Mean Squared Error, and R^2.

Main Results:

• Systolic pressure prediction: Mean Absolute Error of 9.81, Mean Squared Error of 165.13, R^2 of 0.36.
• Diastolic pressure prediction: Mean Absolute Error of 6.00, Mean Squared Error of 58.21, R^2 of 0.30.
• Model errors were comparable to the inherent variability of arm cuff measurements.

Conclusions:

• The developed machine learning model shows accuracy comparable to existing methods.
• Fundus image-based blood pressure assessment may offer insights into long-term hypertension.
• Further clinical trials and prospective studies are recommended for validation.