Neural Network-Based Feature Prediction Using Remote Photoplethysmogram Waveform Analysis

The origin of blood pulse waveform is mainly affected by various factors, such as vascular age, lifestyle, and ability of vascular disorders. Non-contact optical measurements of blood volume pulse in microvascular tissue were performed by remote photoplethysmography. This work aims to classify subjects into a specific vascular health condition based on a given pulse waveform using network-based machine learning models. For this reason, we trained our models by 9000 waveforms taken from palm’s dorsal side of 18 subjects (14 healthy, age 21-54 yrs. and 4 patients affected by septic shock and taking vasopressors, aged 45-81 yrs.). To get variant waveforms, we employed a bilateral thigh supra-systolic occlusion test in healthy subjects to temporarily compromise leg blood supply, potentially altering vascular resistance. To train and validate our models we used five relevant hemodynamic parameters which are related to reflected waves from the periphery. The classification model validation tests showed the following best accuracy taken from five classes: true-positive 98.8 % and false-negative 1.2%. The neural network-based approach could be valuable for prediction of vascular health state from blood pulse waveform in cases when the signal is weak and noisy.