Prediction of Shock-Refractory Ventricular Fibrillation during Resuscitation of Out-of-Hospital Cardiac Arrest

Abstract

Introduction:

Out-of-hospital cardiac arrest (OHCA) due to shock-refractory ventricular fibrillation (VF) is associated with relatively poor survival. The ability to predict refractory VF (requiring ≥3 shocks) in advance of repeated shock failure could enable preemptive targeted interventions aimed at improving outcome, such as earlier administration of antiarrhythmics, reconsideration of epinephrine use or dosage, changes in shock delivery strategy, or expedited invasive treatments.

Methods:

We conducted a cohort study of VF-OHCA to develop an ECG-based algorithm to predict patients with refractory VF. Patients with available defibrillator recordings were randomized 80%/20% into training/test groups. A random forest classifier applied to 3-second ECG segments immediately before and one minute after the initial shock during CPR was used to predict the need for ≥3 shocks based on singular value decompositions of ECG wavelet transforms. Performance was quantified by area under the receiver operating characteristic curve (AUC).

Results:

Of 1376 VF-OHCA patients, 311 (23%) were female, 864 (63%) experienced refractory VF, and 591 (43%) achieved functional neurologic survival. Total shock count was associated with decreasing likelihood of functional neurologic survival, with a relative risk (95% CI) of 0.95 (0.93–0.97) for each successive shock (p<0.001). In the 275 test patients, the AUC (95% CI) for predicting refractory VF was 0.85 (0.79–0.89), with specificity of 91%, sensitivity of 63%, and a positive likelihood ratio of 6.7.

Conclusions:

A machine learning algorithm using ECGs surrounding the initial shock predicts patients likely to experience refractory VF, and could enable rescuers to preemptively target interventions to potentially improve resuscitation outcome.

Introduction

Out-of-hospital cardiac arrest (OHCA) is a leading cause of mortality, claiming hundreds of thousands of lives each year in the United States and millions worldwide.¹ Ventricular fibrillation (VF) is a common initial OHCA dysrhythmia characterized by chaotic electrical activity on the electrocardiogram (ECG). Resuscitation from VF-OHCA can be achieved through coordinated administration of cardiopulmonary resuscitation (CPR) to provide circulatory support, electric shock to terminate VF, and medications to increase perfusion and facilitate defibrillation.

Current resuscitation guidelines follow a graduated protocol whereby emergency medical services (EMS) provide CPR interrupted every two minutes to allow rhythm assessment and shock if VF is present.² If VF is refractory (requiring ≥ 3 shocks), guidelines recommend administration of antiarrhythmics (such as amiodarone) to improve the likelihood of successful defibrillation and help prevent recurrent fibrillation.^3,4 The rationale for antiarrhythmic treatment in response to refractory VF is supported in part by investigations which have observed that a greater number of shocks and increased time in VF are associated with a lower likelihood of survival.^5–7 Hence, the ability to predict refractory VF in advance of repeated shock failure could enable preemptive interventions targeted at improving the relatively poor outcomes of the refractory VF subgroup. Such therapies might include earlier or increased antiarrhythmic dose administration,^8–12 reconsideration of epinephrine use or dosage,^13–15 changes in how shocks are administered,¹⁶ or expedited invasive interventions.^17–21 By contrast, empiric treatment of all VF OHCA patients using such strategies may unnecessarily expose patients who achieve best outcomes under current protocol, introducing potential risk among those unlikely to benefit.^22–24

There is currently no automated strategy to proactively identify patients for whom three or more shocks will be required during resuscitation. However, real-time evaluation of defibrillator electrocardiographic signals could offer such predictive potential. In this study we hypothesized that shock-refractory VF can be predicted by real-time analysis of the ECG using machine learning, a subset of artificial intelligence (AI) that can learn patterns from data. Such a development could enable preemptive clinical strategies intended to reduce refractory VF and potentially improve survival.

Methods

Study Design, Population, and Setting

The investigation was a retrospective cohort study of patients presenting to EMS with VF-OHCA in greater King County, WA between 2008–2020. Cases were not eligible if the patient was < 18 years, had a do-not-resuscitate order, or received a shock by laypersons or police prior to EMS arrival. Given the goal to predict refractory VF using the ECG signal early during EMS resuscitation, we excluded cases if the defibrillator recording did not contain valid, uninterrupted ECG and transthoracic impedance signals from the time of rhythm analysis before the initial shock through at least the first minute following the initial shock.

The study region has a population of 1.5 million persons and is served by a two-tiered EMS system. First responders are emergency medical technician-firefighters who are equipped with automated external defibrillators, provide Basic Cardiac Life Support, and respond to the scene approximately 5 minutes following emergency (9-1-1) dispatch. The second response tier is paramedics who are equipped with manual defibrillators, provide Advanced Cardiac Life Support, and arrive on-scene an average of 8 minutes following dispatch.²⁵ The EMS follows American Heart Association resuscitation guidelines, which include repeated 2-minute CPR cycles separated by rhythm and pulse checks, as well as pharmacologic interventions.²

Data Collection and Pre-Processing

The study system maintains an OHCA registry that follows the Utstein template.²⁶ The registry collects demographic, circumstance, care, and outcome information from the dispatch recording, EMS reports, defibrillator recording, hospital record, and death certificates. Patient outcomes include return of spontaneous circulation (ROSC) at the end of EMS care and survival to hospital discharge with functional neurologic status as defined by a Cerebral Performance Category (CPC) score of 1–2.²⁷

ECG signals were collected from Philips MRx and Forerunner 3 (Philips, Bothell, WA), and Physio-Control Lifepak 12 and 15 (Stryker, Redmond, WA) defibrillator recordings. ECGs were resampled from native sampling rates of between 125–250 Hz to a common frequency of 250 Hz, and filtered from 1–40 Hz using a 4th-order Butterworth filter with forward-backward implementation to remove drift and high-frequency noise.

Investigators used custom MATLAB software (Mathworks, Natick, MA) to identify two discrete ECG intervals surrounding the initial shock. Specifically, from each patient we collected 1) a 3-second ECG segment prior to the first VF shock during the concurrent CPR pause for defibrillator rhythm analysis, and 2) a 3-second post-shock ECG segment 1 minute following initial shock during the scheduled 2-minute period of ongoing CPR (Figure 1). In the pre-shock segments, absence of CPR and presence of VF were confirmed by two-reviewer consensus examination of the ECG and transthoracic impedance signals. By contrast, post-shock segments were collected during the scheduled period of CPR (agnostic of the ECG rhythm type) to ensure study algorithm generalizability consistent with current best-practice CPR.

Figure 1: Example application of algorithm.

Current guidelines recommend antiarrhythmic administration following shock 3. One potential application of the proposed method might include an earlier antiarrhythmic recommendation based on a prediction of ≥ 3 shocks. (The ECG shown is an illustration only, and is not to scale.) (CPR = cardiopulmonary resuscitation, ECG = electrocardiogram, VF = ventricular fibrillation)

Outcome

The study outcome was the number of shocks provided by EMS over the course of VF-OHCA treatment. We a-priori defined patients with ≥ 3 shocks as shock-refractory, as this shock count corresponds to the guideline threshold for drug initiation (specifically antiarrhythmics).^2–4 As a sensitivity analysis, we also redefined refractory VF as patients who received ≥ 4 shocks.

Algorithm Description

To predict refractory VF, a given patient’s 3-second pre-shock and post-shock ECG segments were first converted into time-frequency representations (scalograms) by convolution with complex Morlet wavelets (Table S1).²⁸ We used scalograms rather than Fourier transform-based methods because they have properties useful for ECG analysis, such as good temporal resolution at high frequencies and the ability to detect coherent structures in a signal.^28–30 The pre-shock and post-shock scalograms were computed between 2.5–30 Hz and 10–30 Hz, respectively, to provide a more comprehensive representation of the CPR-free pre-shock ECG while excluding the majority of CPR artifact (which is concentrated below 10 Hz) in the post-shock ECG.³¹ Post-shock scalogram energy was normalized within each frequency to further emphasize high-frequency content above CPR frequencies. Each scalogram was then projected via dot products onto a stored “Eigenscalogram” scalogram library (described below) to generate a corresponding set of projection (similarity) values. These projection values were used as input features to a machine learning classifier which predicted the probability of shock-refractory VF (Figure 2, Figure S1).

Figure 2: Algorithm prediction examples.

Examples are shown for a patient requiring 2 shocks (left) versus 20 shocks (right). (F = frequency, t = time, M = magnitude, VF = ventricular fibrillation.)

Algorithm Development

Patients were randomly divided into 80% training and 20% test groups for algorithm development and evaluation. We generated scalogram singular modes by transforming the training dataset via singular value decomposition, a linear transformation which synthesized the training scalograms into an equal number of singular mode scalograms sorted by the amount of information (variance) in the data they represented.^32,33 Because they were ranked by the amount of information they encompassed, a small subset of the leading singular mode scalograms could be retained to succinctly approximate the unique characteristics of the original training data. We retained the ranked singular mode scalograms as candidates for potential inclusion in the algorithm’s Eigenscalogram libraries.

We performed a parameter search which varied the number of leading singular modes included in the pre-shock and post-shock Eigenscalogram libraries, the machine learning classifier type, and the classifier parameters. Different classifier types were considered including random forest, support vector machine, and logistic regression. The parameter search was conducted by four-fold cross-validation within training data to select the combination of classifier, classifier parameters, and number of leading modes retained in the Eigenscalogram libraries which yielded the greatest mean area under the receiver operating characteristic curve (AUC) for predicting refractory VF among the cross-validation holdout folds (Figure S2).

The Eigenscalogram libraries yielding the highest mean cross-validation holdout AUC comprised the first 12 and first 13 singular mode scalograms generated from the training dataset’s pre-shock and post-shock ECG scalograms, respectively (Figures S3-S4). The best-performing classifier to predict the probability of refractory VF was a random forest, a machine learning model that uses an ensemble of decision trees to make majority vote-based predictions. After selecting the optimal Eigenscalogram modes, classifier type, and classifier parameters based on training data cross-validation folds, the entire training dataset was used to then generate the final Eigenscalogram libraries and train the final classifier (Figure S2). We used odds ratios to interpret the univariate association between each of the Eigenscalograms and refractory VF in training data (Figure S5). Two Eigenscalograms were particularly useful to the random forest classifier as measured by relative feature importance, a characteristic confirmed by their univariate association with refractory VF (Figure S6).

Statistical Methods

We compared Utstein characteristics among study patients according to refractory VF status (< 3 shocks versus ≥ 3 shocks) and according to inclusion status using Wilcoxon rank-sum tests for continuous variables and chi-squared tests for proportions. We used Poisson regression to determine the association (relative risk) between increasing shock count and clinical outcome (ROSC and functional neurologic survival).

Algorithm classification performance was evaluated in training and test data using AUC. To assess operational implications, classification decision thresholds were selected from training data according to minimum target specificities (80–95%), which were applied to determine corresponding sensitivity, specificity, predictive values, and likelihood ratios for the test cohort. The threshold for statistical significance was 0.01. All analyses were performed using MATLAB.

Code and Data Availability

MATLAB code to execute the method on example data is available at https://github.com/jcoult/refractory-vf. Study data sufficient to replicate the reported results will be provided by the corresponding author upon reasonable request, contingent on applicable agreements for data use.

Human Subjects

The Research Review Committee of King County Public Health and the Institutional Review Board at the University of Washington Human Subjects Division approved the study and waiver of informed consent.

Results

Study Group and Outcomes

Of the 2193 potentially eligible patients who presented to EMS with VF-OHCA between 2008–2020, 451 (21%) were excluded due to missing or incompatible defibrillator downloads, and 366 (17%) were excluded due to issues with the defibrillator ECG signal (Figure 3). Of the resulting 1376 patients in the study group, 311 (23%) were female, 1000 (73%) achieved ROSC at the end of EMS care, and 591 (43%) survived to hospital discharge with functional neurologic status (Table 1). Included versus excluded patients were generally comparable, with similar demographics, circumstances, and survival outcomes, though differences were observed in the proportion of arrests before EMS arrival and in the rate of hospital admission (Table S2).

Figure 3: Study cohort and exclusions.

Inclusion and exclusion of study patients are illustrated. (OHCA = out-of-hospital cardiac arrest, VF = ventricular fibrillation.)

Table 1:

Study group characteristics

	Study Group	Non-Refractory (< 3 total shocks)	Refractory (≥ 3 total shocks)
Patients, n(%)	1376 (100)	512 (37.2)	864 (62.8)
Female, n(%)	311 (22.6)	129 (25.2)	182 (21.1)
Age, median (IQR)	62 (53, 72)	62 (52, 73)	62 (53, 72)
Cardiac etiology, n(%)	1269 (92.2)	468 (91.4)	801 (92.7)
Location, n(%)
Home	882 (64.1)	311 (60.7)	571 (66.1)
Public	443 (32.2)	176 (34.4)	267 (30.9)
Nursing Home	51 (3.7)	25 (4.9)	26 (3.0)
Arrest before EMS arrival, n(%)	1315 (95.6)	478 (93.4)	837 (96.9)*
Witnessed arrest, n(%)	1075 (78.1)	411 (80.3)	664 (76.9)
Bystander cardiopulmonary resuscitation, n(%)	1042 (79.2†)	398 (83.3†)	644 (76.9†)*
EMS Response (minutes), median (IQR)	5.2 (4.2, 6.6)	5.2 (4, 6.2)	5.4 (4.3, 6.9)
Total shocks, median (IQR)	3 (1, 6)	1 (1, 2)	5 (4, 8)*
ROSC at end of EMS care, n(%)	1000 (72.7)	440 (85.9)	560 (64.8)*
Admit to hospital, n(%)	998 (72.5)	433 (84.6)	565 (65.4)*
Survive to hospital discharge, n(%)	633 (46.0)	299 (58.4)	334 (38.7)*
Survive with CPC 1 or 2, n(%)	591 (43.0)	280 (54.7)	311 (36.0)*

^*p<0.01 versus non-refractory patients

^†Percentage among patients who arrested before EMS arrival and were thus eligible for bystander CPR

Study group characteristics are illustrated overall and for refractory versus non-refractory patient subsets. (CPC = cerebral performance category, EMS = emergency medical services, IQR = interquartile range, ROSC = return of spontaneous circulation.)

The median total number of shocks administered was 3 (interquartile range (IQR): 1–6), with 864 (63%) patients having refractory VF as defined by receipt of ≥ 3 shocks. Refractory VF patients were more likely to have arrested prior to EMS arrival, less likely to have received bystander CPR, and had worse outcomes (Table 1). When modeling continuous shock count versus outcome, the unadjusted relative risk of good outcome decreased with each successive shock, with a relative risk (95% confidence interval (CI)) of ROSC = 0.96 (0.94–0.97, p<0.001), and relative risk of functional neurologic survival = 0.95 (0.93–0.97, p<0.001) (Figure 4).

Figure 4: Unadjusted association of total shocks with patient outcome.

The proportions of ROSC and functional survival outcomes versus total shock count are shown with trendlines (A), and the number of patients within each shock count is illustrated (B). (IQR = interquartile range, ROSC = return of spontaneous circulation.)

Algorithm Performance

The AUC (95% CI) values for prediction of shock-refractory patients were 0.89 (0.87–0.91) in the training group (N=1101) and 0.85 (0.79–0.89) in the test group (N=275) (Figure 5). Using classification cutoffs derived from the training data, we determined the specificity, sensitivity, predictive values, and likelihood ratios for refractory VF prediction in the test group, observing for example a sensitivity of 63%, specificity of 91%, a positive predictive value of 93%, and a likelihood ratio of 6.7 using the cutoff selected to achieve 90% specificity (Table 2).

Figure 5: Training and test results.

Distributions of classifier predicted probabilities of refractory VF for training (A) and test (C) data are shown for each class. Receiver operating characteristic curves for training (B) and test (D) predictions are also illustrated. (AUC = area under the receiver operating characteristic curve, CI = confidence interval, N = number of patients, VF = ventricular fibrillation.)

Table 2:

Algorithm performance

Target Specificity	Dataset	Specificity	Sensitivity	Positive Predictive Value	Negative Predictive Value	Positive Likelihood Ratio	Negative Likelihood Ratio
95%	Training	95% (398/417)	60% (409/684)	96%	59%	13.1	0.42
	Test	93% (88/95)	58% (105/180)	94%	54%	7.9	0.45
90%	Training	90% (376/417)	66% (452/684)	92%	62%	6.7	0.38
	Test	91% (86/95)	63% (114/180)	93%	57%	6.7	0.41
85%	Training	85% (355/417)	72% (492/684)	89%	65%	4.8	0.33
	Test	84% (80/95)	74% (133/180)	90%	63%	4.7	0.31
80%	Training	80% (335/417)	77% (524/684)	86%	68%	3.9	0.29
	Test	81% (77/95)	79% (142/180)	89%	67%	4.2	0.26

Sensitivity and specificity values for predicting refractory ventricular fibrillation patients (receiving at least 3 shocks) using training (N=1101) and test (N=275) data are listed. Four classification decision thresholds were selected from training data to achieve each of the four minimum target specificity values, and used to compute the resulting specificity, sensitivity, and likelihood ratio values.

Sensitivity Analysis

When refractory VF was defined as requiring ≥ 4 shocks (instead of ≥ 3), the number of patients meeting the definition of refractory VF decreased from 864 (63%) to 670 (49%). The AUC (95% CI) for predicting ≥ 4 shocks was 0.86 (0.84–0.88) and 0.78 (0.72–0.83) in training and test data, respectively.

Discussion

In this retrospective cohort investigation of VF-OHCA, an ECG-based algorithm predicted refractory VF. Most VF patients requiring ≥ 3 shocks were correctly identified while maintaining high specificity. These results suggest potential for automated identification of patients at risk of shock-refractory VF without compromising current guidelines for minimally-interrupted CPR. Such insight could be used to better direct early treatment with the goal of reducing total shock count and potentially improving survival.

Refractory VF and Clinical Outcome

We observed that nearly two-thirds of the study cohort required ≥ 3 shocks and about half required ≥ 4 shocks, highlighting the common circumstance of refractory VF. Moreover, a higher total shock count was associated with worse resuscitation outcomes, findings which are supported by prior studies.^5–7 Hence, refractory VF occurs commonly in VF-OHCA and is a group that may benefit from preemptive alternative treatment strategies, given the relatively worse prognosis for patients experiencing this condition.

Algorithm Performance versus Prior Study

A prior investigation proposed a clinical decision rule which relies upon patient-specific variables to predict refractory VF.³⁴ Although this approach is a useful conceptual model, the dependence on individual clinical and arrest circumstance characteristics may be challenging for time-sensitive operational implementation during resuscitation. In contrast, the current study’s ECG-based method can operate automatically without rescuer input, and its performance (AUC = 0.85) surpasses that of the prior study (AUC = 0.67).

Clinical Implications of Algorithm Performance

The goal of resuscitation treatment is to optimize benefit and minimize risk across what is often a heterogeneous physiologic cohort. The study algorithm predicts patients who will require ≥ 3 EMS shocks to treat VF-OHCA in the context of best practices that continue CPR throughout the scheduled 2-minute CPR intervals. Given the algorithm’s favorable performance when applied within current clinical guidelines and processing requirements suitable for defibrillator hardware, the proposed method could be integrated into resuscitation practice to better guide treatment. For example, one strategy might use the algorithm to provide earlier or higher-dose antiarrhythmic treatment and/or a lower or delayed dose of a vasopressor such as epinephrine (given its potential proarrhythmic effects) for those at high risk of refractory VF.^8–15 The algorithm might also be used to inform which patients could be considered for modified shock delivery strategies (vector change or double sequential defibrillation) or early transport for advanced hospital care such as an emergent coronary artery intervention or extracorporeal CPR.^16–21

Implications of Operational Cutoffs

A decision to modify standard treatment may depend on operational cutoffs informed by the algorithm’s performance characteristics (Table 2, Figure S7). For example, a more substantial treatment change might require a very high level of specificity (and consequent lower sensitivity) that would maintain conventional care for patients who only require < 3 shocks. However, a change considered to have relatively less risk may permit a lower specificity in order to benefit a greater number of patients with refractory VF. The ideal approach however is speculative and would require prospective evaluation to determine clinical utility.

Artificial Intelligence and Prediction of Refractory VF

The use of AI to automate a more patient-specific approach to resuscitation care using real-time signals such as the ECG is a promising area of investigation.^30,35–38 However, clinician adoption of AI to inform medical decisions can be challenged by the “black box” nature of many classification algorithms (e.g. deep neural networks).^39,40 This is especially salient for AI applications that might modify guideline care in real time during OHCA resuscitation. In contrast, the proposed machine learning algorithm allows insight into how refractory VF is predicted by using features that are relatively interpretable. Examination of pre-shock ECG characteristics used by the method suggests that lower initial VF frequency and amplitude may potentially be indicative of shock-refractory patients (Figure S5-S6). Likewise, examination of post-shock features suggests that following an initial defibrillation attempt, shock-refractory patients may have lower energy at higher ECG frequencies. These observations are supported by previous investigations which have observed that pre-shock^29,41 and post-shock^42,43 ECG characteristics predict patient outcomes. In particular, assessment of the ECG specifically 1 minute following shock is useful in evaluating the interaction between defibrillation and patient pathophysiology.⁴⁴ The current study combines these concepts using machine learning to leverage ECG information collected immediately before and 1 minute after initial shock (the latter without requiring CPR interruption) to distinguish refractory VF patients, supporting the broader advancement towards a greater role of AI in guiding resuscitation care.⁴⁰

Limitations and Directions for Future Study

The current study occurred in a system with good clinical outcomes, and used specific defibrillator models from two manufacturers, which may limit generalizability in other systems with distinct patients, different outcomes, or other defibrillators. A number of eligible VF patients (37%) were excluded, primarily due to missing or incompatible defibrillator files or issues with the ECG recording. Ultimately external validation in other EMS systems and patient populations would therefore be important to confirm our results. We did not include Utstein variables in the model of shock count versus outcome or as predictors in the study algorithm. However, in supplementary analyses, adjusting for Utstein variables did not significantly change the estimated association of shock count with worse outcome, and combining Utstein variables with ECG information did not improve prediction of refractory VF (Figure S8, Table S3).

The proposed algorithm in its present form does not distinguish between different presentations of shock-refractory VF; that is, incessant VF despite shock, versus recurrence of VF after an initial termination of VF. The study algorithm operates within current resuscitation guidelines, and thus presumes availability of a 3-second CPR-free VF ECG segment prior to initial shock as well as an ECG segment during CPR 1 minute following the first shock. Future work may seek to evaluate different analysis intervals, input signals, or AI methods in order to improve refractory VF prediction, distinguish between different presentations of refractory VF, or refine the ability to align specific treatments with patients most likely to benefit. Finally, the study retrospectively tested an automated ECG-based algorithm to predict refractory VF but did not prospectively implement changes in care to evaluate the effect on clinical outcome.

Conclusion

A novel ECG-based algorithm enables prediction of refractory VF, information that may help advance targeted treatment strategies designed to improve resuscitation. Future study should seek to develop additional methods for real-time patient evaluation during resuscitation, and to understand how such methods might be implemented to optimize clinical care and improve outcome.

Supplementary Material

Supplemental materials include Tables S1-3 and Figures S1-8.

Clinical Perspective.

What is new?

• Among patients with ventricular fibrillation (VF) out-of-hospital cardiac arrest, those requiring ≥ 3 shocks (refractory VF) have worse clinical outcomes.

• A novel machine learning algorithm can automatically identify patients likely to experience refractory VF during resuscitation using the defibrillator electrocardiogram.

What are the clinical implications?

• The ability to predict which patients may experience refractory VF in advance of repeated shock failure could enable preemptive interventions targeted at improving outcomes for these patients, such as earlier antiarrhythmic administration, reconsideration of epinephrine use or dosage, or changes in shock delivery strategy.

Non-standard Abbreviations and Acronyms

AI

artificial intelligence

AUC

area under the receiver operating characteristic curve

CI

confidence interval

CPC

cerebral performance category

CPR

cardiopulmonary resuscitation

ECG

electrocardiogram

EMS

emergency medical services

IQR

interquartile range

ROSC

return of spontaneous circulation

OHCA

out-of-hospital cardiac arrest

VF

ventricular fibrillation