1D Convolutional Neural Networks for Estimation of Compensatory Reserve from Blood Pressure Waveforms

Abstract

We propose a Deep Convolutional Neural Network (CNN) architecture for computing a Compensatory Reserve Metric (CRM) for trauma victims suffering from hypovolemia (decreased circulating blood volume). The CRM is a single health indicator value that ranges from 100% for healthy individuals, down to 0% at hemodynamic decompensationwhen the body can no longer compensate for blood loss. The CNN is trained on 20 second blood pressure waveform segments obtained from a finger-cuff monitor of 194 subjects. The model accurately predicts CRM when tested on data from 22 additional human subjects obtained from Lower Body Negative Pressure (LBNP) emulation of hemorrhage, attaining a mean squared error (MSE) of 0.0238 over the full range of values, including those from subjects with both low and high tolerance to central hypovolemia.

Efficacy of the compensatory reserve measurement in an emergency department trauma population

“…The study personnel recorded and stored continuous analog arterial waveforms on the Clearsight blood pressure monitor and later downloaded these data to a PowerLab (Version 16/35, AD Instruments, Dunedin, New Zealand) data acquisition and integration system along with LabChart Pro (version 8.1.16, AD Instruments, Dunedin, New Zealand). We subsequently used these electronic arterial waveform recordings for retrospective calculation of the CRM using a machine learning algorithm software based on 1‐D convolutional neural networks and designed to interrogate subtle changes in waveform features 26 . We evaluated the CRM in 20‐s intervals for each individual patient on a scale of 100%–0%, where 100% reflects a maximal capacity to compensate for reduced circulating blood volume and 0% represents the onset of decompensated shock with thresholds of >60%, 30%–60%, and <30% 4,21 .…”

“…We subsequently used these electronic arterial waveform recordings for retrospective calculation of the CRM using a machine learning algorithm software based on 1-D convolutional neural networks and designed to interrogate subtle changes in waveform features. 26 We evaluated the CRM in 20-s intervals for each individual patient on a scale of 100%-0%, where 100% reflects a maximal capacity to compensate for reduced circulating blood volume and 0% represents the onset of decompensated shock with thresholds of >60%, 30%-60%, and <30%. 4,21 Thus, we used the CRM in the present investigation to assess the ability of the algorithm to distinguish trauma patients who received blood transfusion or airway management in the ED from those who received no interventions.…”

Photoplethysmography Analysis with Duffing–Holmes Self-Synchronization Dynamic Errors and 1D CNN-Based Classifier for Upper Extremity Vascular Disease Screening

“…Hence, chaotic scatter maps can be applied to classification tasks. For the automatic screening function, a one-dimensional (1D) or 2D convolutional neural network (CNN) [30][31][32][33][34][35] can be used to design a multilayer classifier consisting of the convolutional layers, pooling layers, flattening layers, and a fully connected network. In previous studies [33][34][35], PPG analysis with 2D CNN-based classifiers has been applied to HR and hypertension detection, PAD-related CVD risk detection, and PAD detection, exhibiting high accuracy for screening mild to severe diseases, arterial hardening, and Nor.…”

“…The conventional CNN is designed to process 2D images or videos exclusively. The 1D CNN is a modified scheme and is preferable to the 2D CNN in dealing with 1D signals (e.g., ECG signals, blood pressure waveforms, vibration signals, and phonoangiographic signals [30][31][32]) because the computational requirement and complexity of the 1D CNN are lower than those of a 2D CNN, the scheme of the 1D CNN is easier to train and implement that that of a 2D CNN, and training the 1D CNN with a few hidden layers and nodes is faster than training a 2D CNN [30,31]. The configuration of the 1D CNN consists of 1D convolution operations, subsampling (pooling) processes, and a multilayer classifier, as shown in Figure 5.…”

“…graphic signals [30][31][32]) because the computational requirement and complexity of the 1D CNN are lower than those of a 2D CNN, the scheme of the 1D CNN is easier to train and implement that that of a 2D CNN, and training the 1D CNN with a few hidden layers and nodes is faster than training a 2D CNN [30,31]. The configuration of the 1D CNN consists of 1D convolution operations, subsampling (pooling) processes, and a multilayer classifier, as shown in Figure 5.…”