Out-of-hospital cardiac arrest (OHCA) is a major public health issue and the most common cause of death worldwide (Kiguchi et al., 2020). Defined as a disorganized electrical activity without the presence of distinguishable QRS complexes, ventricular fibrillation (VF) is the most common etiology in patients suffering from OHCA. Although patients with VF as an initial rhythm were more likely to be successfully resuscitated than with other rhythms, the less than 30% survival rate remains unsatisfactory (Rajan et al., 2017). Survival from OHCA depends on a complex system working together to secure the best outcome, and early defibrillation with concurrent high-quality cardiopulmonary resuscitation (CPR) is the most important life-saving intervention for VF (Koike et al., 2011). To provide a general treatment strategy, the latest guidelines for CPR and emergency cardiovascular care recommend the initiation of high-quality CPR, delivery of an electrical shock as soon as a defibrillator is available and every 2 min thereafter if VF persists (Panchal et al., 2020).

With the deepening understanding of the physiology of cardiac arrest and resuscitation, it is increasingly clear that not all patients in VF benefit from being treated with the same intervention. The development of new technologies also enables unprecedented ability to personalize resuscitation according to the interval after onset of the VF, the effectiveness of CPR and the consequent myocardial metabolic state (Chalkias et al., 2019). Electrocardiogram (ECG) waveforms, which are routinely collected by the automated external defibrillators (AEDs), have been extensively investigated to identifying VFs and predicting defibrillation outcomes (Bessen et al., 2021). Based on the observation that the characteristics of VF signals reflect the pathophysiological and metabolic status of the fibrillating heart, a variety of measures have been developed to characterize the underlying organization of the myocardial electrical activity and with the ultimate goal of guiding CPR (Coult et al., 2019). Amplitude spectrum area (AMSA), as a well-established measure to predict defibrillation outcome, has been shown to be correlated with coronary perfusion pressure, reflect the energy state of the myocardium, and reveal whether myocardial perfusion is improved (Reynolds et al., 2012). Clinical studies have confirmed that both preshock AMSA and relative changes in the AMSA during CPR are associated with shock success (Indik et al., 2014; Schoene et al., 2014; Ristagno et al., 2015). Therefore, real-time monitoring of AMSA may serve as a strategy for quality control of CPR and for individualization of resuscitation (Babini et al., 2021).

Unfortunately, real-time monitoring of AMSA remains unachievable with present algorithms. Because the mechanical activity of chest compression (CC) induces artifacts in the ECG, reliable AMSA calculation can only be achieved during CC pauses. The life-saving benefit of CPR will be markedly compromised if CC is repetitively interrupted to calculate AMSA during resuscitation efforts (Shimizu et al., 2021). Therefore, it is a major challenge for accurate AMSA assessment during uninterrupted CPR to complete the goal of patient specific, time-sensitive and physiology-directed strategy of personalized resuscitation.

To perform reliable and accurate ECG waveform analysis without interrupting CPR, a number of signal processing solutions have been proposed to remove CC artifacts in the past two decades (Gong et al., 2013). One solution is to suppress artifacts using only the ECG waveform, such as the Kalman filter (Ruiz de Gauna et al., 2008), independent component analysis (Granegger et al., 2011), coherent line removal algorithm (Amann et al., 2010), empirical mode decomposition (Lo et al., 2013) and condition-based filtering algorithm (Hajeb-Mohammadalipour et al., 2021). Although artifacts can be strongly suppressed, and the signal-to-noise ratio (SNR) is markedly improved, the specificity for VF detection is insignificantly improved using these methods. The other solution is to remove the artifacts using additional CPR-related reference waveforms, including Gabor multiplier (Werther et al., 2009), Wiener filter (Aase et al., 2000), recursive adaptive matching pursuit algorithm (Husøy et al., 2002), adaptive filtering based on the least mean square (LMS) algorithm (Irusta et al., 2009; Gong et al., 2017) and variable-frequency notch filter (Coult et al., 2021). CPR artifacts are first modeled by a reference waveform, such as compression depth, transthoracic impedance, and compression force, and are subsequently subtracted from the corrupt waveform (Ruiz de Gauna et al., 2014). Although the time-frequency variability of the artifacts can be reconstructed to a certain extent, additional equipment is required to obtain these references, and they are not available in all existing AEDs.

Recently, powerful tools developed for machine learning have been successfully applied in the field of biomedical signal processing by end-to-end architectures of deep neural networks (Li et al., 2021). A currently popular tool is the convolutional neural network (CNN), which is a hierarchical neural network model with alternating convolutional and subsampling layers, followed by a fully connected layer that is identical to a multilayer perceptron. The biggest advantage of CNN is that it convolves the learned features with the input data without manually extracting features. Inspired by the feature learning capacity of CNN for image classification, CNNs have also been shown to be able to classify non-image time series and waveform data. Specifically, several attempts have been made in ECG waveform analysis, such as signal denoising (Fotiadou et al., 2020), QRS detection (Zahid et al., 2022), heartbeat classification (Kiranyaz et al., 2016), arrhythmia detection (Hannun et al., 2019) and defibrillation success prediction (Ivanović et al., 2020). Although these studies demonstrated that CNNs have a wide application prospect in ECG waveform analysis, they focused mainly on the problem of signal classification and/or tested on clinical ECG records with low-level motion artifacts. Therefore, whether CNN can be used for reliable VF signal analysis during CPR when the ECG waveform is severely corrupted with a high level of artifacts remains unsolved.

In this paper, we propose a novel AMSA estimation algorithm that can provide continuous guidance for personalizing CPR in realtime during resuscitation efforts using a 1D CNN. To accomplish this, an architecture consisting of 1D CNN blocks and fully connected layers is used for feature extraction and AMSA estimation. The proposed approach can estimate AMSA from CC-corrupted VF signals without reference, thus, eliminating the need for artifact filtering, feature extraction and postprocessing.

The flowchart of this study is shown in Figure 2. First, the VF signals were preprocessed, labeled and distributed to different datasets. Second, the model was trained by the augmented training data and optimized by the validation data. Finally, the performance of the developed algorithm was evaluated using the testing data and compared with traditional adaptive filtering and FFT-based techniques. There was no crossover between subjects in the training, validation and testing sets.

A total of 6642 S _UVF/S _CVF and S _CVF/S _UVF segment pairs were extracted in the derivation set. The segment pairs randomized to training and validation sets in the five-fold validation iteration are 5,372/1,270, 5,435/1,207, 5,175/1,467, 5,260/1,382 and 5,318/1,324, respectively. Additionally, 1,108 S _SVF/S _UVF segment pairs, 2,423 S _UVF/S _CVF or S _CVF/S _UVF segment pairs and 284 S _PVF segments (98 successful and 186 unsuccessful shocks) were extracted in the testing set. The optimal hyperparameters of the model were W = 11, N = 5, and C = 32.

Given the importance of continuous AMSA evaluation during uninterrupted CPR for quality control and personalized defibrillation, this study introduces a solution for reliable AMSA estimation from VF signals using the 1D CNN model. Compared with the traditional AMSA calculation method combining adaptive filtering and FFT techniques, the proposed algorithm can reliably estimate AMSA from VF signals. Also, the proposed method is validated to preserve the predictive performance of defibrillation success.

Analyzing ECG waveforms during uninterrupted CPR remains a major challenge because signals are always contain artifacts. An earlier study indicated that CC-produced artifacts are an additive noise predominantly generated from electrode motion and thoracic muscle contraction (Fitzgibbon et al., 2002). When the underlying cardiac rhythm of a patient is asystole, the ECG waveform recorded during CPR can be regarded as pure artifacts produced by CC (Gong et al., 2013). In the time domain, the artifact features a relatively high amplitude and an nearly periodic waveform. In the frequency domain, the energy content is concentrated around the harmonics of the fundamental frequency being that of the CC with a bandwidth of approximately 0-20 Hz. Because the energy of VF signals lies in the frequency band of 0-18 Hz and completely overlaps with the artifacts, calculating AMSA directly from the corrupted VF signal will lead to an overestimation of its value (Coult et al., 2019). Using the reference waveform recorded from an additional channel independent of the ECG waveform but related to CPR artifacts, the artifacts can be modeled and suppressed by adaptive filters (Irusta et al., 2009; Ruiz de Gauna et al., 2014; Gong et al., 2017). However, the accuracy of shockable rhythm classification is still lower than that used for uncorrupted ECG waveforms. The primary reason for this result is that the artifacts are variable, and their characteristics depend on how the CC is administered and on the characteristics of the patient and the recording system (Ruiz de Gauna et al., 2014).

Recently, strategies have shifted from suppressing artifacts to mining hidden features with existing deep neural networks. Research was first performed in the ECG waveform without CPR. Picon et al. (2019) introduced an approach that combines a 2-layer CNN and a long short-term memory (LSTM) network for the detection of VF. Krasteva et al. (2020) demonstrated that the optimal hyperparameters of CNN with five convolutional blocks, a filter number between 5 and 20, and a kernel size of 10 resulted in maximum validation performance for the detection of shockable rhythm. Panda et al. (2020) used the first five decomposed subband signals, while Sabut et al. (2021) used a feature set of 24 time-frequency-based parameters of ECG waveforms as the input of CNN for shockable rhythm detection. Additionally, Ivanović et al. (2020) proposed a 3-stage CNN feature extractor model to predict defibrillation success. Using the raw 4 s VF signals immediately prior to the first shock in 260 adult patients, the strategy was demonstrated to be capable of learning useful representations for defibrillation outcome. These studies revealed that CNN approaches have equivalent or superior performance to classical feature extraction based machine learning algorithms and have the potential to be used in AEDs. Then, studies focused on rhythm analysis for the CC-corrupted ECG waveforms. Isasi et al. (2020) designed an algorithm to classify shockable rhythms using filtered ECG as the input of a CNN that contained 3 convolution blocks and 2 fully connected layers. Hajeb-M et al. (2021) designed a CNN model comprised of convolutional layers, residual networks, and bidirectional LSTM using 2-dimensional images that combined the ECG waveform with the amplitude and phase information derived from the short-time Fourier transform as input. Jekova and Krasteva (2021) showed that the optimal architecture of a deep CNN with the best performance for shockable rhythm analysis during CPR was with three convolutional layers when the raw ECG waveform was used. Although the datasets and validation methods of these studies were considerably different, their results unanimously confirmed that CNN can be used for reliable rhythm analysis during CPR.

In this study, we proposed an algorithm to estimate AMSA and guide CPR continuously during resuscitation effort using a 1D CNN. Instead of suppressing the CC-related artifacts and computing AMSA from the frequency domain, we implemented an end-to-end architecture of CNN, applying differentiated VF at the input and obtaining a continuous variable AMSA at the output, without determining the presence of CC with additional sensors. The architecture design and a high-quality database are the two primary determinants of the performance of CNN based methods. In this study, an architecture consisting of a 6-layer 1D CNN and three fully connected layers was developed based on previous studies concerning ECG waveform denoising, VF detection and defibrillation success prediction using CNN. Each block was composed of convolution, ReLU activation, dropout and max-pooling. Considering that the output result is a continuous variable rather than a classified variable, the classifier was composed of three fully connected perceptrons. The optimized hyperparameters of the model were similar to the results reported in previous studies, which further confirmed the robustness of the CNN network for VF signal feature extraction (Picon et al., 2019; Isasi et al., 2020; Krasteva et al., 2020; Panda et al., 2020; Hajeb-M et al., 2021; Jekova and Krasteva, 2021; Sabut et al., 2021). Conversely, the model must be trained by a representative and accurately labeled database with a sufficient sample size. Because it is impossible to have the underlying clean VF signals during CPR, prior efforts used simulated data that were constructed by randomly adding different scaled pure CC artifacts to the uncorrupted VF signals (Lo et al., 2013). Considering that VF is a quasi stationary signal, we used the adjacent uncorrupted signals as surrogate data of the corrupted VF to calculate AMSA. The simulated data indicated that the AMSA values between the uncorrupted segment pairs were highly correlated and that the AMSA value calculated from the uncorrupted VF could be used as the true value of the adjacent corrupted VF. The real-life dataset of corrupted/uncorrupted segment pairs was used to train the model due to the heterogeneity of CPR artifacts in real-life data and homogeneity of CPR artifacts in simulated data. As shown in Table 4, the performance of the proposed method is demonstrated to be superior to that of the traditional AMSA calculation method using simulated data and independent real-life CPR data (Zuo et al., 2021; Coult et al., 2022). Experimental results demonstrated the excellent feature extraction capability to exploit all information reflecting the energy state of the myocardium hidden in the VF signal. To our knowledge, this is the first study that uses a deep neural network model to calculate AMSA during uninterrupted CPR. This study indicated that the paradigm shift represented by end-to-end deep learning may enable a new approach to monitor the effectiveness of CC and predict defibrillation success during CPR using VF signals alone.

There are several limitations/drawbacks that must be addressed regarding the current study. First, the dataset used to train the model was not uniformly distributed. As the samples with low AMSA values were much more abundant than those with high AMSA values of the real-life patient data, it is still unknown whether the overall performance of the proposed method will be improved and bias in the Bland-Altman plots will be decreased if a sufficiently large uniformly distributed dataset is available. Second, although the proposed method can provide continuous information about the dynamic changes in the AMSA during CPR, the VF features in their hidden layers were extracted blindly without a comprehensive explanation of the mechanism for future clinical application. Third, the model was established based on CC-corrupted VF signals using the least mean square method, so for the uncorrupted VF signals, such as segments during CC pauses and prior to shock delivery, the AMSA values would be underestimated. Four, the algorithm is designed to VF waveform analysis; thus, it cannot be used without the help of a reliable shockable rhythm detection algorithm. Although several algorithms have been demonstrated to correctly classify shockable rhythms during CPR using CNN, the performance of combining the current algorithm with those shockable rhythm detection algorithms also must be validated in future studies.

This study introduces a 1D CNN based continuous AMSA estimation approach, which is promising for reliable AMSA estimation during uninterrupted CPR using solely the ECG waveform. Experimental results indicate that this method performs better than traditional adaptive filtering and FFT-based techniques. Combined with the shockable rhythm detection algorithms during CC, the algorithm has the potential to be incorporated into current versions of AEDs, and personalized CPR can be implemented by real-time monitoring of the effectiveness of CC and predicting the probability of defibrillation success.