Construction of a hypotension prediction model using deep learning with visible and infrared facial images of hemodialysis patients

During hemodialysis, intradialytic hypotension, a risk factor for cardiovascular complications, can occur due to fluid removal, dialysate-induced vasodilation, and osmotic pressure reductions. Because blood pressure is measured only intermittently during dialysis, episodes of hypotension may go undetected until they manifest clinically. In this study, we developed a deep learning model to predict intradialytic hypotension 15 minutes before its onset, using only facial visible and infrared images captured from hemodialysis patients. We recruited eight patients undergoing routine dialysis, acquired synchronized visible-light (30 fps) and infrared (1 fps) facial images at approximately 15-minute intervals, and labeled each image according to subsequent blood pressure measurements. Convolutional neural networks of varying depth were trained on the red, green, blue, a*, and b* channels of visible images and on infrared images. The model using the red channel of visible images with a shallow network achieved the highest area under the receiver operating characteristic curve, approximately 0.90, outperforming all infrared-based models. Shallower architectures captured local skin blood flow features more effectively, and Grad-CAM visualization confirmed that nasal and upper cheek regions contributed most to the prediction. We also observed that some patients exhibited increases in facial redness prior to hypotension, whereas others showed decreases, reflecting individual microvascular responses. These findings demonstrate that non-contact facial imaging and optimized network architecture can enable early detection of intradialytic hypotension, potentially improving real-time patient monitoring and intervention strategies in dialysis care.