Machine Learning Prediction of Mean Arterial Pressure from the Photoplethysmography Waveform During Hemorrhagic Shock and Fluid Resuscitation.

We aimed to evaluate the non-invasive photoplethysmography waveform as a means to predict mean arterial pressure using artificial intelligence models. This was performed using datasets captured in large animal hemorrhage and resuscitation studies. An initial deep learning model trained using a subset of large animal data and was then evaluated for real-time blood pressure prediction. With the successful proof-of-concept experiment, we further tested different feature extraction approaches as well as different machine learning and deep learning methodologies to examine how various combinations of these methods can improve the accuracy of mean arterial pressure predictions from a non-invasive photoplethysmography sensor. Different combinations of feature extraction and artificial intelligence models successfully predicted mean arterial pressure throughout the study. Overall, manual feature extraction fed into a long short-term memory network tracked the mean arterial pressure through hemorrhage and resuscitation with the highest accuracy.