Detection of shockable rhythms using convolutional neural networks during chest compressions provided by a load distributing band

Abstract

Load Distributing Band (LDB) mechanical chest compression devices are used to treat out-of-hospital cardiac arrest (OHCA) patients. The artefacts that LDB chest compressions induce in the ECG impede a reliable shock/noshock diagnosis, resulting in compression interruptions to analyze the ECG. The aim of this study was to design a deep learning algorithm to accurately detect shockable rhythms with concurrent LDB compressions. The dataset was comprised of 780 shockable and 2644 nonshockable rhythms from 242 OHCA patients treated with the LDB device. Underlying rhythms were annotated by expert reviewers in artefact-free intervals. The method consisted of two stages: a Recursive Least Squares (RLS) filter to remove LDB compression artefacts and a shock/no-shock decision algorithm based on CNNs. Shock/no-shock diagnoses were compared with the rhythm annotations to obtain the sensitivity (Se) and specificity (Sp) of the method. The median Se, Sp were 92.2%, 96.6%, respectively. The proposed algorithm met the American Heart Association's (AHA) requirements for rhythm analysis.

Optimization of End-to-End Convolutional Neural Networks for Analysis of Out-of-Hospital Cardiac Arrest Rhythms during Cardiopulmonary Resuscitation

“…Those reference methods have been selected because they use test set from human ECGs and either have analyzed only ECG signals without the need for additional channels for CPR assessment, or have performed the rhythm classification via neural networks. Although the notable disparities between the studies in Table 6, concerning different test conditions and databases, their comparison reveals that neural networks (this study and [47][48][49]) present equal or better performance than traditional machine learning classifiers [29,38,40] for analysis of cardiac arrest rhythms during CPR. Formally, the relatively simple fully convolutional architecture of CNN3-CC-ECG network in this study shows similar performance to a hybrid DNN architecture [49] (including convolutional layers, residual blocks and bidirectional LSTM layers), i.e., Se of the hybrid DNN model is 5.2% points higher than CNN3-CC-ECG (94.2% vs. 89%), while Sp of CNN3-CC-ECG is 5.2% points better than the hybrid DNN network (91.3% vs. 86.1%).…”

“…Hybrid architectures, including recurrent long short-term memory (LSTM) layers have also been applied, although LSTM is computationally expensive and alone has not demonstrated better performance than fully CNN [38]. In conditions of CPR artifacts, only three recent publications were found to investigate DNNs for detection of shockable (Sh) and non-shockable (NSh) rhythms [47][48][49]. Similarly to their previous studies [15,16,50], Isasi et al [47,48] rely on the strategy for pre-filtering of CC artifacts by a recursive least squares filter using information for instantaneous CC periods derived either through compression depth signal from external accelerometer sensor or strictly controlled by a mechanical chest compression device.…”

“…In conditions of CPR artifacts, only three recent publications were found to investigate DNNs for detection of shockable (Sh) and non-shockable (NSh) rhythms [47][48][49]. Similarly to their previous studies [15,16,50], Isasi et al [47,48] rely on the strategy for pre-filtering of CC artifacts by a recursive least squares filter using information for instantaneous CC periods derived either through compression depth signal from external accelerometer sensor or strictly controlled by a mechanical chest compression device. The filtered ECG samples are fed to a CNN model with three convolutional blocks and two fully connected layers for binary classification of Sh/NSh rhythms.…”

“…Asystoles have not been tested in [49], although they are the most common rhythm in cardiac arrests that define the border amplitude between Sh and NSh rhythms, and therefore, are important for adequate training of end-to-end DNN classifiers. The other two studies with neural networks [47,48] present similar architecture as CNN3-CC-ECG with three convolutional layers, although the max-pooling size, number of filters and kernel sizes are different. It is worthy to note that the input to the networks is different, i.e., CNN3-CC-ECG network is optimized to directly process ECG with CC artifacts, while the networks in [47,48] are optimized to process ECG after adaptive filtration with the help of an external sensor.…”

“…The other two studies with neural networks [47,48] present similar architecture as CNN3-CC-ECG with three convolutional layers, although the max-pooling size, number of filters and kernel sizes are different. It is worthy to note that the input to the networks is different, i.e., CNN3-CC-ECG network is optimized to directly process ECG with CC artifacts, while the networks in [47,48] are optimized to process ECG after adaptive filtration with the help of an external sensor. The latter network is linked to a specific signal processing chain, which is hardly reproducible without that external sensor.…”