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

Background: This study was developed to determine whether a machine learning model could be developed to assess blood pressure with accuracy comparable to arm cuff measurements.

Methods: A deep learning model was developed based on the UK Biobank dataset and was trained to detect both systolic and diastolic pressure. The hypothesis was formulated after data collection and before the development of the model. Comparison was conducted between arm cuff measurements, as ground truth, and results from the model, using Mean Absolute Error, Mean Squared Error, and Coefficient of Determination (R^2).

Results: Systolic pressure was measured with 9.81 Mean Absolute Error, 165.13 Mean Squared Error and 0.36 R^2. Diastolic pressure was measured with 6.00 Mean Absolute Error, 58.21 and 0.30 R^2.

Conclusions: This model improves on existing research and shows errors comparable to the variability of hand cuff measurements. The use of fundus images to assess blood pressure may be more indicative of long-term hypertension. Additional trials in clinical settings may be necessary, as well as additional prospective studies to validate results.