Rhythm Analysis during Cardiopulmonary Resuscitation Using Convolutional Neural Networks

Machine learning innovations in CPR: a comprehensive survey on enhanced resuscitation techniques

This survey paper explores the transformative role of Machine Learning (ML) and Artificial Intelligence (AI) in Cardiopulmonary Resuscitation (CPR), marking a paradigm shift from conventional, manually driven resuscitation practices to intelligent, data-driven interventions. It examines the evolution of CPR through the lens of predictive modeling, AI-enhanced devices, and real-time decision-making tools that collectively aim to improve resuscitation outcomes and survival rates. Unlike prior surveys that either focus solely on traditional CPR methods or offer general insights into ML applications in healthcare, this work provides a novel interdisciplinary synthesis tailored specifically to the domain of CPR. It presents a comprehensive taxonomy that classifies ML techniques into four key CPR-related tasks: rhythm analysis, outcome prediction, non-invasive blood pressure and chest compression modeling, and real-time detection of pulse and Return of Spontaneous Circulation (ROSC). The paper critically evaluates emerging ML approaches-including Reinforcement Learning (RL) and transformer-based models-while also addressing real-world implementation barriers such as model interpretability, data limitations, and deployment in high-stakes clinical settings. Furthermore, it highlights the role of eXplainable AI (XAI) in fostering clinical trust and adoption. By bridging the gap between resuscitation science and advanced ML techniques, this survey establishes a structured foundation for future research and practical innovation in ML-enhanced CPR. It offers clear insights, identifies unexplored opportunities, and sets a forward-looking research agenda identifying emerging trends and practical implementation challenges aiming to improve both the reliability and effectiveness of CPR in real-world emergencies.

Comparison of Neural Network Structures for Identifying Shockable Rhythm During Cardiopulmonary Resuscitation

Background/Objectives: Minimizing interruptions in chest compressions is very important when resuscitating patients with cardiac arrest. Recently, research has analyzed electrocardiograms (ECGs) during chest compressions using convolutional neural networks (CNNs). This study aimed to compare the accuracy of deeper neural networks and more advanced structures. Methods: ECGs with chest compression artifacts were obtained from patients who visited the emergency department of Korea University Ansan Hospital from September 2019 to February 2024. We compared the accuracy of a deeper CNN, long short-term memory (LSTM), and a CNN with an attention mechanism and residual block against a reference model. The input of the model was 5 s ECG segments with compression artifacts, and the output was the probability that the ECG with the artifacts was a shockable rhythm. Results: A total of 1889 ECGs with compression artifacts from 172 patients were included in this study. There were 969 ECGs annotated as shockable and 920 as non-shockable. The area under the receiver operating characteristic curve (AUROC) of the reference model was 0.8672. The AUROCs of the deeper CNN for five and seven layers were 0.7374 and 0.6950, respectively. The AUROCs of LSTM and bidirectional LSTM were 0.7719 and 0.8287, respectively. The AUROC of the CNN with the attention mechanism and residual block was 0.7759. Conclusions: CNNs with deeper layers or those incorporating attention mechanisms, residual blocks, and LSTM architectures did not exhibit better accuracy. To improve the model accuracy, a larger dataset or advanced augmentation techniques may be required, rather than complicating the structure of the model.

Implementing an intelligent diagnosis and treatment system for in-hospital cardiac arrest in the Utstein style: a multi-center case study

BACKGROUND

Cardiac arrest presents a variety of causes and complexities, making it challenging to develop targeted treatment plans. Often, the original data are either inadequate or lack essential patient information. In this study, we introduce an intelligent system for diagnosing and treating in-hospital cardiac arrest (IHCA), aimed at improving the success rate of cardiopulmonary resuscitation and restoring spontaneous circulation.

METHODS

To compensate for insufficient or incomplete data, a hybrid mega trend diffusion method was used to generate virtual samples, enhancing system performance. The core of the system is a modified episodic deep reinforcement learning module, which facilitates the diagnosis and treatment process while improving sample efficiency. Uncertainty analysis was performed using Monte Carlo simulations, and dependencies between different parameters were assessed using regular vine copula. The system's effectiveness was evaluated using ten years of data from Utstein-style IHCA registries across seven hospitals in China's Hebei Province.

RESULTS

The system demonstrated improved performance compared to other models, particularly in scenarios with inadequate data or missing patient information. The average reward scores in two key stages increased by 2.3-9 and 9.9-23, respectively.

CONCLUSIONS

The intelligent diagnosis and treatment effectively addresses IHCA, providing reliable diagnosis and treatment plans in IHCA scenarios. Moreover, it can effectively induce cardiopulmonary resuscitation and restoration of spontaneous circulation processes even when original data are insufficient or basic patient information is missing.

Artificial intelligence for predicting shockable rhythm during cardiopulmonary resuscitation: In-hospital setting

AIM OF THE STUDY

This study aimed to develop an artificial intelligence (AI) model capable of predicting shockable rhythms from electrocardiograms (ECGs) with compression artifacts using real-world data from emergency department (ED) settings. Additionally, we aimed to explore the black box nature of AI models, providing explainability.

METHODS

This study is retrospective, observational study using a prospectively collected database. Adult patients who presented to the ED with cardiac arrest or experienced cardiac arrest in the ED between September 2021 and February 2024 were included. ECGs with a compression artifact of 5 s before every rhythm check were used for analysis. The AI model was designed based on convolutional neural networks. The ECG data were assigned into training, validation, and testing sets on a per-patient basis to ensure that ECGs from the same patient did not appear in multiple sets. Gradient-weighted class activation mapping was employed to demonstrate AI explainability.

RESULTS

A total of 1,889 ECGs with compression artifacts from 172 patients were used. The area under the receiver operating characteristic curve (AUROC) for shockable rhythm prediction was 0.8672 (95% confidence interval [CI]: 0.8161-0.9122). The AUROCs for manual and mechanical compression were 0.8771 (95% CI: 0.8054-0.9408) and 0.8466 (95% CI: 0.7630-0.9138), respectively.

CONCLUSION

This study was the first to accurately predict shockable rhythms during compression using an AI model trained with actual patient ECGs recorded during resuscitation. Furthermore, we demonstrated the explainability of the AI. This model can minimize interruption of cardiopulmonary resuscitation and potentially lead to improved outcomes.

Clinical performance of AED shock advisory system with integrated Analyze Whilst Compressing algorithm for analysis of the ECG rhythm during out-of-hospital cardiopulmonary resuscitation: A secondary analysis of the DEFI 2022 study

Objective

This study involving automated external defibrillators (AEDs) in early treatment of refibrillation aims to evaluate the performance of a new shock advisory system (SAS) during chest compressions (CC) in out-of-hospital cardiac arrest (OHCA) patients.

Methods

This work focuses on AED SAS performance as a secondary outcome of DEFI 2022 clinical prospective study, which included first-analysis shockable OHCA patients. SAS employs the Analyze Whilst Compressing (AWC) algorithm to interact with both cardiopulmonary resuscitation (CPR) and shock advice by conditional operation of two-stage ECG analysis in presence or absence of chest compressions. AWC is triggered by the first-shock recommendation. Then, after 1 min of CPR, ECG analysis during CC decides between two treatment scenarios. For patients with refibrillation, CPR is paused for immediate confirmation analysis and shock advice. For patients with non-shockable rhythms, CPR is continued for 2 min until standard analysis.

Results

Clinical data from 285 OHCA patients with shock recommendation at the first-analysis by AEDs (DEFIGARD TOUCH7, Schiller Médical) consisted of 576 standard analyses, 2011 analyses during CC, 577 confirmation analyses in absence of CC. Global AED SAS performance meets the standard recommendations for arrhythmia analysis sensitivity (94.9%) and specificity (>99.3%). AWC provided innovative treatment of shockable rhythms by stopping CPR earlier than 2 min in most ventricular fibrillations (92.9%), while most non-shockable patients (86.5-95.2%) benefitted from continuous CPR for at least 2 min.

Conclusion

This study provides positive evidence for routine use of AEDs with AWC-integrated algorithm for ECG analysis during CPR by first-responders in early OHCA treatment.Clinical Trial Registration: Registration number: NCT04691089, trial register: ClinicalTrials.gov.

Patient's airway monitoring during cardiopulmonary resuscitation using deep networks

Cardiopulmonary resuscitation (CPR) is a crucial life-saving technique commonly administered to individuals experiencing cardiac arrest. Among the important aspects of CPR is ensuring the correct airway position of the patient, which is typically monitored by human tutors or supervisors. This study aims to utilize deep transfer learning for the detection of the patient's correct and incorrect airway position during cardiopulmonary resuscitation. To address the challenge of identifying the airway position, we curated a dataset consisting of 198 recorded video sequences, each lasting 6-8 s, showcasing both correct and incorrect airway positions during mouth-to-mouth breathing and breathing with an Ambu Bag. We employed six cutting-edge deep networks, namely DarkNet19, EfficientNetB0, GoogleNet, MobileNet-v2, ResNet50, and NasnetMobile. These networks were initially pre-trained on computer vision data and subsequently fine-tuned using the CPR dataset. The validation of the fine-tuned networks in detecting the patient's correct airway position during mouth-to-mouth breathing achieved impressive results, with the best sensitivity (98.8 %), specificity (100 %), and F-measure (97.2 %). Similarly, the detection of the patient's correct airway position during breathing with an Ambu Bag exhibited excellent performance, with the best sensitivity (100 %), specificity (99.8 %), and F-measure (99.7 %).

Emergence of Artificial Intelligence and Machine Learning Models in Sudden Cardiac Arrest: A Comprehensive Review of Predictive Performance and Clinical Decision Support

Sudden cardiac death/sudden cardiac arrest (SCD/SCA) is an increasingly prevalent cause of mortality globally, particularly in individuals with preexisting cardiac conditions. The ambiguous premortem warnings and the restricted interventional window related to SCD account for the complexity of the condition. Current reports suggest SCD to be accountable for 20% of all deaths hence accurately predicting SCD risk is an imminent concern. Traditional approaches for predicting SCA, particularly "track-and-trigger" warning systems have demonstrated considerable inadequacies, including low sensitivity, false alarms, decreased diagnostic liability, reliance on clinician involvement, and human errors. Artificial intelligence (AI) and machine learning (ML) models have demonstrated near-perfect accuracy in predicting SCA risk, allowing clinicians to intervene timely. Given the constraints of current diagnostics, exploring the benefits of AI and ML models in enhancing outcomes for SCA/SCD is imperative. This review article aims to investigate the efficacy of AI and ML models in predicting and managing SCD, particularly targeting accuracy in prediction.

Extracting physiologic and clinical data from defibrillators for research purposes to improve treatment for patients in cardiac arrest

Background

A defibrillator should be connected to all patients receiving cardiopulmonary resuscitation (CPR) to allow early defibrillation. The defibrillator will collect signal data such as the electrocardiogram (ECG), thoracic impedance and end-tidal CO2, which allows for research on how patients demonstrate different responses to CPR. The aim of this review is to give an overview of methodological challenges and opportunities in using defibrillator data for research.

Methods

The successful collection of defibrillator files has several challenges. There is no scientific standard on how to store such data, which have resulted in several proprietary industrial solutions. The data needs to be exported to a software environment where signal filtering and classifications of ECG rhythms can be performed. This may be automated using different algorithms and artificial intelligence (AI). The patient can be classified being in ventricular fibrillation or -tachycardia, asystole, pulseless electrical activity or having obtained return of spontaneous circulation. How this dynamic response is time-dependent and related to covariates can be handled in several ways. These include Aalen's linear model, Weibull regression and joint models.

Conclusions

The vast amount of signal data from defibrillator represents promising opportunities for the use of AI and statistical analysis to assess patient response to CPR. This may provide an epidemiologic basis to improve resuscitation guidelines and give more individualized care. We suggest that an international working party is initiated to facilitate a discussion on how open formats for defibrillator data can be accomplished, that obligates industrial partners to further develop their current technological solutions.

Impact of Combined Modulation of Two Potassium Ion Currents on Spiral Waves and Turbulent States in the Heart

In the realm of cardiac research, the control of spiral waves and turbulent states has been a persistent focus for scholars. Among various avenues of investigation, the modulation of ion currents represents a crucial direction. It has been proved that the methods involving combined control of currents are superior to singular approaches. While previous studies have proposed some combination strategies, further reinforcement and supplementation are required, particularly in the context of controlling arrhythmias through the combined regulation of two potassium ion currents. This study employs the Luo-Rudy phase I cardiac model, modulating the maximum conductance of the time-dependent potassium current and the time-independent potassium current, to investigate the effects of this combined modulation on spiral waves and turbulent states. Numerical simulation results indicate that, compared to modulating a single current, combining reductions in the conductance of two potassium ion currents can rapidly control spiral waves and turbulent states in a short duration. This implies that employing blockers for both potassium ion currents concurrently represents a more efficient control strategy. The control outcomes of this study represent a novel and effective combination for antiarrhythmic interventions, offering potential avenues for new antiarrhythmic drug targets.

Artificial intelligence to support out-of-hospital cardiac arrest care: A scoping review

Background

Artificial intelligence (AI) has demonstrated significant potential in supporting emergency medical services personnel during out-of-hospital cardiac arrest (OHCA) care; however, the extent of research evaluating this topic is unknown. This scoping review examines the breadth of literature on the application of AI in early OHCA care.

Methods

We conducted a search of PubMed®, Embase, and Web of Science in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews guidelines. Articles focused on non-traumatic OHCA and published prior to January 18th, 2023 were included. Studies were excluded if they did not use an AI intervention (including machine learning, deep learning, or natural language processing), or did not utilize data from the prehospital phase of care.

Results

Of 173 unique articles identified, 54 (31%) were included after screening. Of these studies, 15 (28%) were from the year 2022 and with an increasing trend annually starting in 2019. The majority were carried out by multinational collaborations (20/54, 38%) with additional studies from the United States (10/54, 19%), Korea (5/54, 10%), and Spain (3/54, 6%). Studies were classified into three major categories including ECG waveform classification and outcome prediction (24/54, 44%), early dispatch-level detection and outcome prediction (7/54, 13%), return of spontaneous circulation and survival outcome prediction (15/54, 20%), and other (9/54, 16%). All but one study had a retrospective design.

Conclusions

A small but growing body of literature exists describing the use of AI to augment early OHCA care.

Artificial intelligence and machine learning in prehospital emergency care: A scoping review

Our scoping review provides a comprehensive analysis of the landscape of artificial intelligence (AI) applications in prehospital emergency care (PEC). It contributes to the field by highlighting the most studied AI applications and identifying the most common methodological approaches across 106 included studies. The findings indicate a promising future for AI in PEC, with many unique use cases, such as prognostication, demand prediction, resource optimization, and the Internet of Things continuous monitoring systems. Comparisons with other approaches showed AI outperforming clinicians and non-AI algorithms in most cases. However, most studies were internally validated and retrospective, highlighting the need for rigorous prospective validation of AI applications before implementation in clinical settings. We identified knowledge and methodological gaps using an evidence map, offering a roadmap for future investigators. We also discussed the significance of explainable AI for establishing trust in AI systems among clinicians and facilitating real-world validation of AI models.

Deep Learning Strategy for Sliding ECG Analysis during Cardiopulmonary Resuscitation: Influence of the Hands-Off Time on Accuracy

This study aims to present a novel deep learning algorithm for a sliding shock advisory decision during cardiopulmonary resuscitation (CPR) and its performance evaluation as a function of the cumulative hands-off time. We retrospectively used 13,570 CPR episodes from out-of-hospital cardiac arrest (OHCA) interventions reviewed in a period of interest from 30 s before to 10 s after regular analysis of automated external defibrillators (AEDs). Three convolutional neural networks (CNNs) with raw ECG input (duration of 5, 10, and 15 s) were applied for the shock advisory decision during CPR in 26 sequential analyses shifted by 1 s. The start and stop of chest compressions (CC) can occur at arbitrary times in sequential slides; therefore, the sliding hands-off time (sHOT) quantifies the cumulative CC-free portion of the analyzed ECG. An independent test with CPR episodes in 393 ventricular fibrillations (VF), 177 normal sinus rhythms (NSR), 1848 other non-shockable rhythms (ONR), and 3979 asystoles (ASYS) showed a substantial improvement of VF sensitivity when increasing the analysis duration from 5 s to 10 s. Specificity was not dependent on the ECG analysis duration. The 10 s CNN model presented the best performance: 92-94.4% (VF), 92.2-94% (ASYS), 96-97% (ONR), and 98.2-99.5% (NSR) for sliding decision times during CPR; 98-99% (VF), 98.2-99.8% (ASYS), 98.8-99.1 (ONR), and 100% (NSR) for sliding decision times after end of CPR. We identified the importance of sHOT as a reliable predictor of performance, accounting for the minimal sHOT interval of 2-3 s that provides a reliable rhythm detection satisfying the American Heart Association (AHA) standards for AED rhythm analysis. The presented technology for sliding shock advisory decision during CPR achieved substantial performance improvement in short hands-off periods (>2 s), such as insufflations or pre-shock pauses. The performance was competitive despite 1-2.8% point lower ASYS detection during CPR than the standard requirement (95%) for non-noisy ECG signals. The presented deep learning strategy is a basis for improved CPR practices involving both continuous CC and CC with insufflations, associated with minimal CC interruptions for reconfirmation of non-shockable rhythms (minimum hands-off time) and early treatment of VF (minimal pre-shock pauses).

Current and Future Use of Artificial Intelligence in Electrocardiography

Artificial intelligence (AI) is increasingly used in electrocardiography (ECG) to assist in diagnosis, stratification, and management. AI algorithms can help clinicians in the following areas: (1) interpretation and detection of arrhythmias, ST-segment changes, QT prolongation, and other ECG abnormalities; (2) risk prediction integrated with or without clinical variables (to predict arrhythmias, sudden cardiac death, stroke, and other cardiovascular events); (3) monitoring ECG signals from cardiac implantable electronic devices and wearable devices in real time and alerting clinicians or patients when significant changes occur according to timing, duration, and situation; (4) signal processing, improving ECG quality and accuracy by removing noise/artifacts/interference, and extracting features not visible to the human eye (heart rate variability, beat-to-beat intervals, wavelet transforms, sample-level resolution, etc.); (5) therapy guidance, assisting in patient selection, optimizing treatments, improving symptom-to-treatment times, and cost effectiveness (earlier activation of code infarction in patients with ST-segment elevation, predicting the response to antiarrhythmic drugs or cardiac implantable devices therapies, reducing the risk of cardiac toxicity, etc.); (6) facilitating the integration of ECG data with other modalities (imaging, genomics, proteomics, biomarkers, etc.). In the future, AI is expected to play an increasingly important role in ECG diagnosis and management, as more data become available and more sophisticated algorithms are developed.

The ability of machine learning algorithms to predict defibrillation success during cardiac arrest: A systematic review

OBJECTIVE

To evaluate the existing knowledge on the effectiveness of machine learning (ML) algorithms inpredicting defibrillation success during in- and out-of-hospital cardiac arrest.

METHODS

MEDLINE, Embase, CINAHL and Scopus were searched from inception to August 30, 2022. Studies were included that utilized ML algorithms for prediction of successful defibrillation, observed as return of spontaneous circulation (ROSC), survival to hospital or discharge, or neurological status at discharge.Studies were excluded if involving a trauma, an unknown underlying rhythm, an implanted cardiac defibrillator or if focused on the prediction or onset of cardiac arrest. Risk of bias was assessed using the PROBAST tool.

RESULTS

There were 2399 studies identified, of which 107 full text articles were reviewed and 15 observational studies (n = 5680) were included for final analysis. 29 ECG waveform features were fed into 15 different ML combinations. The best performing ML model had an accuracy of 98.6 (98.5 - 98.7)%, with 4 second ECG intervals. An algorithm incorporating end-tidal CO2 reported an accuracy of 83.3% (no CI reported). Meta-analysis was not performed due to heterogeneity in study design, ROSC definitions, and characteristics.

CONCLUSION

Machine learning algorithms, specifically Neural Networks, have been shown to have potential to predict defibrillation success for cardiac arrest with high sensitivity and specificity.Due to heterogeneity, inconsistent reporting, and high risk of bias, it is difficult to conclude which, if any, algorithm is optimal. Further clinical studies with standardized reporting of patient characteristics, outcomes, and appropriate algorithm validation are still required to elucidate this. PROSPERO 2020 CRD42020148912.

Real-time amplitude spectrum area estimation during chest compression from the ECG waveform using a 1D convolutional neural network

Introduction: Amplitude spectrum area (AMSA) is a well-established measure than can predict defibrillation outcome and guiding individualized resuscitation of ventricular fibrillation (VF) patients. However, accurate AMSA can only be calculated during cardiopulmonary resuscitation (CPR) pause due to artifacts produced by chest compression (CC). In this study, we developed a real-time AMSA estimation algorithm using a convolutional neural network (CNN). Methods: Data were collected from 698 patients, and the AMSA calculated from the uncorrupted signals served as the true value for both uncorrupted and the adjacent corrupted signals. An architecture consisting of a 6-layer 1D CNN and 3 fully connected layers was developed for AMSA estimation. A 5-fold cross-validation procedure was used to train, validate and optimize the algorithm. An independent testing set comprised of simulated data, real-life CC corrupted data, and preshock data was used to evaluate the performance. Results: The mean absolute error, root mean square error, percentage root mean square difference and correlation coefficient were 2.182/1.951 mVHz, 2.957/2.574 mVHz, 22.887/28.649% and 0.804/0.888 for simulated and real-life testing data, respectively. The area under the receiver operating characteristic curve regarding predicting defibrillation success was 0.835, which was comparable to that of 0.849 using the true value of the AMSA. Conclusions: AMSA can be accurately estimated during uninterrupted CPR using the proposed method.

[Effects of automated external defibrillators on hands-off intervals in lay rescuers]

Background

Survival chances after out-of-hospital cardiac arrests caused by hyperdynamic electric cardiac rhythms can be significantly improved by early defibrillation with automated external defibrillators (AEDs). As postulated in international guidelines, the resulting hands-off intervals should not exceed 10 s.

Objectives

We investigated delay in onset of chest compressions and the length of hands-off intervals during defibrillation associated with the application of AEDs.

Materials and methods

In a prospective, randomized, single-blinded observational study, the resuscitation efforts by first year medical students were analyzed in different emergency scenarios on manikins. Delay in onset of chest compressions and the length of hands-off intervals between voice prompts from four conventional devices were compared during shockable and nonshockable rhythms. Satisfaction with the device, difficulties with the application, and suggested improvements were assessed by questionnaire.

Results

In a total of 70 applications, the start with thoracic compressions was delayed by a mean of 115 s. On average, the first shock was administered after 125 s in shockable heart rhythms. Perishock pauses of less than 10 s were achieved with none of the tested devices. Hands-off intervals during defibrillation differed significantly between the devices (p < 0.001). Improvements were suggested regarding marking, voice prompts, and electrodes.

Conclusions

Perishock pause of less than 10 s was not achieved with any of the tested devices. Shortened and more precise voice prompts as well as more clearly arranged labeling and layout of pads are needed to simplify application, reduce delayed onset of chest compressions and shorten hands-off intervals.

Role of artificial intelligence in defibrillators: a narrative review

Automated external defibrillators (AEDs) and implantable cardioverter defibrillators (ICDs) are used to treat life-threatening arrhythmias. AEDs and ICDs use shock advice algorithms to classify ECG tracings as shockable or non-shockable rhythms in clinical practice. Machine learning algorithms have recently been assessed for shock decision classification with increasing accuracy. Outside of rhythm classification alone, they have been evaluated in diagnosis of causes of cardiac arrest, prediction of success of defibrillation and rhythm classification without the need to interrupt cardiopulmonary resuscitation. This review explores the many applications of machine learning in AEDs and ICDs. While these technologies are exciting areas of research, there remain limitations to their widespread use including high processing power, cost and the ‘black-box’ phenomenon.

State-of-the-art Deep Learning Methods on Electrocardiogram Data: A Systematic Review (Preprint)

Background

Electrocardiogram (ECG) is one of the most common noninvasive diagnostic tools that can provide useful information regarding a patient’s health status. Deep learning (DL) is an area of intense exploration that leads the way in most attempts to create powerful diagnostic models based on physiological signals.

Objective

This study aimed to provide a systematic review of DL methods applied to ECG data for various clinical applications.

Methods

The PubMed search engine was systematically searched by combining “deep learning” and keywords such as “ecg,” “ekg,” “electrocardiogram,” “electrocardiography,” and “electrocardiology.” Irrelevant articles were excluded from the study after screening titles and abstracts, and the remaining articles were further reviewed. The reasons for article exclusion were manuscripts written in any language other than English, absence of ECG data or DL methods involved in the study, and absence of a quantitative evaluation of the proposed approaches.

Results

We identified 230 relevant articles published between January 2020 and December 2021 and grouped them into 6 distinct medical applications, namely, blood pressure estimation, cardiovascular disease diagnosis, ECG analysis, biometric recognition, sleep analysis, and other clinical analyses. We provide a complete account of the state-of-the-art DL strategies per the field of application, as well as major ECG data sources. We also present open research problems, such as the lack of attempts to address the issue of blood pressure variability in training data sets, and point out potential gaps in the design and implementation of DL models.

Conclusions

We expect that this review will provide insights into state-of-the-art DL methods applied to ECG data and point to future directions for research on DL to create robust models that can assist medical experts in clinical decision-making.

Automated Condition-Based Suppression of the CPR Artifact in ECG Data to Make a Reliable Shock Decision for AEDs during CPR

Cardiopulmonary resuscitation (CPR) corrupts the morphology of the electrocardiogram (ECG) signal, resulting in an inaccurate automated external defibrillator (AED) rhythm analysis. Consequently, most current AEDs prohibit CPR during the rhythm analysis period, thereby decreasing the survival rate. To overcome this limitation, we designed a condition-based filtering algorithm that consists of three stop-band filters which are turned either 'on' or 'off' depending on the ECG's spectral characteristics. Typically, removing the artifact's higher frequency peaks in addition to the highest frequency peak eliminates most of the ECG's morphological disturbance on the non-shockable rhythms. However, the shockable rhythms usually have dynamics in the frequency range of (3-6) Hz, which in certain cases coincide with CPR compression's harmonic frequencies, hence, removing them may lead to destruction of the shockable signal's dynamics. The proposed algorithm achieves CPR artifact removal without compromising the integrity of the shockable rhythm by considering three different spectral factors. The dataset from the PhysioNet archive was used to develop this condition-based approach. To quantify the performance of the approach on a separate dataset, three performance metrics were computed: the correlation coefficient, signal-to-noise ratio (SNR), and accuracy of Defibtech's shock decision algorithm. This dataset, containing 14 s ECG segments of different types of rhythms from 458 subjects, belongs to Defibtech commercial AED's validation set. The CPR artifact data from 52 different resuscitators were added to artifact-free ECG data to create 23,816 CPR-contaminated data segments. From this, 82% of the filtered shockable and 70% of the filtered non-shockable ECG data were highly correlated (>0.7) with the artifact-free ECG; this value was only 13 and 12% for CPR-contaminated shockable and non-shockable, respectively, without our filtering approach. The SNR improvement was 4.5 ± 2.5 dB, averaging over the entire dataset. Defibtech's rhythm analysis algorithm was applied to the filtered data. We found a sensitivity improvement from 67.7 to 91.3% and 62.7 to 78% for VF and rapid VT, respectively, and specificity improved from 96.2 to 96.5% and 91.5 to 92.7% for normal sinus rhythm (NSR) and other non-shockables, respectively.

A Machine Learning Model for the Prognosis of Pulseless Electrical Activity during Out-of-Hospital Cardiac Arrest

Pulseless electrical activity (PEA) is characterized by the disassociation of the mechanical and electrical activity of the heart and appears as the initial rhythm in 20–30% of out-of-hospital cardiac arrest (OHCA) cases. Predicting whether a patient in PEA will convert to return of spontaneous circulation (ROSC) is important because different therapeutic strategies are needed depending on the type of PEA. The aim of this study was to develop a machine learning model to differentiate PEA with unfavorable (unPEA) and favorable (faPEA) evolution to ROSC. An OHCA dataset of 1921 5s PEA signal segments from defibrillator files was used, 703 faPEA segments from 107 patients with ROSC and 1218 unPEA segments from 153 patients with no ROSC. The solution consisted of a signal-processing stage of the ECG and the thoracic impedance (TI) and the extraction of the TI circulation component (ICC), which is associated with ventricular wall movement. Then, a set of 17 features was obtained from the ECG and ICC signals, and a random forest classifier was used to differentiate faPEA from unPEA. All models were trained and tested using patientwise and stratified 10-fold cross-validation partitions. The best model showed a median (interquartile range) area under the curve (AUC) of 85.7(9.8)% and a balance accuracy of 78.8(9.8)%, improving the previously available solutions at more than four points in the AUC and three points in balanced accuracy. It was demonstrated that the evolution of PEA can be predicted using the ECG and TI signals, opening the possibility of targeted PEA treatment in OHCA.

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

High performance of the shock advisory analysis of the electrocardiogram (ECG) during cardiopulmonary resuscitation (CPR) in out-of-hospital cardiac arrest (OHCA) is important for better management of the resuscitation protocol. It should provide fewer interruptions of chest compressions (CC) for non-shockable organized rhythms (OR) and Asystole, or prompt CC stopping for early treatment of shockable ventricular fibrillation (VF). Major disturbing factors are strong CC artifacts corrupting raw ECG, which we aimed to analyze with optimized end-to-end convolutional neural network (CNN) without pre-filtering or additional sensors. The hyperparameter random search of 1500 CNN models with 2-7 convolutional layers, 5-50 filters and 5-100 kernel sizes was done on large databases from independent OHCA interventions for training (3001 samples) and validation (2528 samples). The best model, named CNN3-CC-ECG network with three convolutional layers (filters@kernels: 5@5,25@20,50@20) presented Sensitivity Se(VF) = 89%(268/301), Specificity Sp(OR) = 91.7%(1504/1640), Sp(Asystole) = 91.1%(3325/3650) on an independent test OHCA database. CNN3-CC-ECG's ability to effectively extract features from raw ECG signals during CPR was comprehensively demonstrated, and the dependency on the CPR corruption level in ECG was tested. We denoted a significant drop of Se(VF) = 74.2% and Sp(OR) = 84.6% in very strong CPR artifacts with a signal-to-noise ratio of SNR < -9 dB, p < 0.05. Otherwise, for strong, moderate and weak CC artifacts (SNR > -9 dB, -6 dB, -3 dB), we observed insignificant performance differences: Se(VF) = 92.5-96.3%, Sp(OR) = 93.4-95.5%, Sp(Asystole) = 92.6-94.0%, p > 0.05. Performance stability with respect to CC rate was validated. Generalizable application of the optimized computationally efficient CNN model was justified by an independent OHCA database, which to our knowledge is the largest test dataset with real-life cardiac arrest rhythms during CPR.

Shock decision algorithm for use during load distributing band cardiopulmonary resuscitation

AIM

Chest compressions delivered by a load distributing band (LDB) induce artefacts in the electrocardiogram. These artefacts alter shock decisions in defibrillators. The aim of this study was to demonstrate the first reliable shock decision algorithm during LDB compressions.

METHODS

The study dataset comprised 5813 electrocardiogram segments from 896 cardiac arrest patients during LDB compressions. Electrocardiogram segments were annotated by consensus as shockable (1154, 303 patients) or nonshockable (4659, 841 patients). Segments during asystole were used to characterize the LDB artefact and to compare its characteristics to those of manual artefacts from other datasets. LDB artefacts were removed using adaptive filters. A machine learning algorithm was designed for the shock decision after filtering, and its performance was compared to that of a commercial defibrillator's algorithm.

RESULTS

Median (90% confidence interval) compression frequencies were lower and more stable for the LDB than for the manual artefact, 80 min^-1 (79.9-82.9) vs. 104.4 min^-1 (48.5-114.0). The amplitude and waveform regularity (Pearson's correlation coefficient) were larger for the LDB artefact, with 5.5 mV (0.8-23.4) vs. 0.5 mV (0.1-2.2) (p < 0.001) and 0.99 (0.78-1.0) vs. 0.88 (0.55-0.98) (p < 0.001). The shock decision accuracy was significantly higher for the machine learning algorithm than for the defibrillator algorithm, with sensitivity/specificity pairs of 92.1/96.8% (machine learning) vs. 91.4/87.1% (defibrillator) (p < 0.001).

CONCLUSION

Compared to other cardiopulmonary resuscitation artefacts, removing the LDB artefact was challenging due to larger amplitudes and lower compression frequencies. The machine learning algorithm achieved clinically reliable shock decisions during LDB compressions.

Optimizing defibrillation during cardiac arrest

PURPOSE OF REVIEW

Current cardiac arrest guidelines are based on a fixed, time-based defibrillation strategy. Rhythm analysis and shock delivery (if indicated) are repeated every 2 min requiring cyclical interruptions of chest compressions. This approach has several downsides, such as the need to temporarily stop cardiopulmonary resuscitation (CPR) for a variable amount of time, thus reducing myocardial perfusion and decreasing the chance of successful defibrillation. A tailored defibrillation strategy should identify treatment priority for each patient, that is chest compressions (CCS) or defibrillation, minimize CCs interruptions, speed up the delivery of early effective defibrillation and reduce the number of ineffective shocks.

RECENT FINDINGS

Real-time ECG analysis (using adaptive filters, new algorithms robust to chest compressions artifacts and shock-advisory algorithms) is an effective strategy to correctly identify heart rhythm during CPR and reduce the hands-off time preceding a shock. Similarly, ventricular fibrillation waveform analysis, that is amplitude spectrum area (AMSA) represents a well established approach to reserve defibrillation in patients with high chance of shock success and postpone it when ventricular fibrillation termination is unlikely. Both approaches demonstrated valuable results in improving cardiac arrest outcomes in experimental and observational study.

SUMMARY

Real-time ECG analysis and AMSA have the potential to predict ventricular fibrillation termination, return of spontaneous circulation and even survival, with discretely high confidence. Prospective studies are now necessary to validate these new approaches in the clinical scenario.

Deep Neural Network Approach for Continuous ECG-Based Automated External Defibrillator Shock Advisory System During Cardiopulmonary Resuscitation

Background Because chest compressions induce artifacts in the ECG, current automated external defibrillators instruct the user to stop cardiopulmonary resuscitation (CPR) while an automated rhythm analysis is performed. It has been shown that minimizing interruptions in CPR increases the chance of survival. Methods and Results The objective of this study was to apply a deep-learning algorithm using convolutional layers, residual networks, and bidirectional long short-term memory method to classify shockable versus nonshockable rhythms in the presence and absence of CPR artifact. Forty subjects' data from Physionet with 1131 shockable and 2741 nonshockable samples contaminated with 43 different CPR artifacts that were acquired from a commercial automated external defibrillator during asystole were used. We had separate data as train and test sets. Using our deep neural network model, the sensitivity and specificity of the shock versus no-shock decision for the entire data set over the 4-fold cross-validation sets were 95.21% and 86.03%, respectively. This result was based on the training and testing of the model using ECG data in both the presence and the absence of CPR artifact. For ECG without CPR artifact, the sensitivity was 99.04% and the specificity was 95.2%. A sensitivity of 94.21% and a specificity of 86.14% were obtained for ECG with CPR artifact. In addition to 4-fold cross-validation sets, we also examined leave-one-subject-out validation. The sensitivity and specificity for the case of leave-one-subject-out validation were 92.71% and 97.6%, respectively. Conclusions The proposed trained model can make shock versus nonshock decision in automated external defibrillators, regardless of CPR status. The results meet the American Heart Association's sensitivity requirement (>90%).