Prediction of Sudden Cardiac Death Manifesting with Documented Ventricular Fibrillation or Pulseless Ventricular Tachycardia

Abstract

Background:

Sudden cardiac death manifests as ventricular fibrillation (VF)/ ventricular tachycardia (VT) potentially treatable with defibrillation, or nonshockable rhythms (pulseless electrical activity/asystole) with low likelihood of survival. There are no available clinical risk scores for targeted prediction of VF/VT.

Objectives:

This study aimed to develop a novel clinical prediction algorithm for avertable sudden cardiac death.

Methods:

Subjects with out-of-hospital sudden cardiac arrest presenting with documented VF or pulseless VT (33% of total cases) were ascertained prospectively from the Portland, Oregon, metro area with population ≈1 million residents (n = 1,374, 2002–2019). Comparisons of lifetime clinical records were conducted with a control group (n=1600) with ≈70% coronary disease prevalence. Prediction models were constructed from a training dataset using backwards stepwise logistic regression and applied to an internal validation dataset. Receiver operating characteristic curves (C statistic) were used to evaluate model discrimination. External validation was performed in a separate, geographically distinct population (Ventura county, California, population ≈850,000, 2015–2020).

Results:

A clinical algorithm (VFRisk) constructed with 13 clinical, electrocardiogram, and echocardiographic variables had very good discrimination in the training dataset (C statistic 0.808; [95% CI: 0.774 – 0.842]) and was successfully validated in internal (C statistic 0.776 [95% CI: 0.725 – 0.827]) and external (C statistic 0.782 [95% CI: 0.718 – 0.846]) datasets. The algorithm substantially outperformed the left ventricular ejection fraction (LVEF) ≤35% (C statistic 0.638) and performed well across the LVEF spectrum.

Conclusions:

An algorithm for prediction of sudden cardiac arrest manifesting with VF/VT was successfully constructed using widely available clinical and noninvasive markers. These findings have potential to enhance primary prevention, especially in patients with mid-range or preserved LVEF.

Graphical Abstract

Development of a Clinical Algorithm for Prediction of Potentially Avertable Sudden Cardiac Death

The prediction algorithm (VFRisk) was developed from prospectively ascertained consecutive individuals who suffered out-of-hospital sudden cardiac arrest with documented ventricular fibrillation (VF)/pulseless ventricular tachycardia (VT), and control subjects, of whom the majority (67%) had significant coronary artery disease. Individuals with documented nonshockable rhythms (pulseless electrical activity or asystole), were excluded. VFRisk was constructed with 13 clinical, electrocardiographic (ECG), and echocardiographic variables obtained from each subject’s lifetime clinical record. VFRisk had very good discrimination in the training dataset as well as the internal validation dataset; and was replicated in an external, geographically distinct population. LVEF = left ventricular ejection fraction; SCD = sudden cardiac death.

Condensed Abstract

Ventricular fibrillation (VF) and pulseless ventricular tachycardia (VT) are potentially treatable with defibrillation, thereby averting sudden cardiac death (SCD). Targeting an intermediate risk population, we developed a novel and user-friendly clinical prediction algorithm (VFRisk) for avertable SCD. Constructed with 13 clinical, ECG, and echocardiographic variables, VFRisk performed well and was successfully validated in internal, and external datasets. The algorithm substantially outperformed left ventricular ejection fraction (LVEF) ≤ 35% and performed equally well across the LVEF spectrum. These findings have the potential to enhance primary prevention, especially in patients with mid-range or preserved LVEF.

Introduction

Out of hospital sudden cardiac arrest (SCA) claims at least 300,000 US lives on an annual basis.(1) Patients who experience SCA can present with ventricular fibrillation/tachycardia (VF/VT), pulseless electrical activity (PEA) or asystole.(2) The only lethal rhythms amenable to treatment with defibrillation are VF/VT. If individuals at risk for VF/VT could be identified before their SCA and implanted with a prophylactic implantable cardioverter-defibrillator (ICD), SCD could be averted. The ICD is not effective for nonshockable rhythms (PEA or asystole). It stands to reason that targeted risk prediction of SCA likely to manifest with VF/VT would have the highest likelihood of identifying individuals who would benefit from a prophylactic ICD, thus averting sudden cardiac death (SCD).(3) However, there are no current risk prediction algorithms for SCA presenting with VF/VT.

In addition, the current clinical approach to primary prevention of SCD relies mainly on provision of the ICD to patients with severely reduced left ventricular ejection fraction (LVEF).(4,5) However, the majority (~70%) of individuals with LVEF measured before SCD do not have severely reduced LVEF, highlighting the importance of identifying novel clinical approaches to risk stratification for mid-range and preserved LVEF.(6)

For 2 decades, the Oregon Sudden Unexpected Death Study (SUDS) has prospectively ascertained SCD cases and control subjects in the general population, with the primary goal of developing a clinical algorithm for VF/VT. (7) The study has a unique design, enrolls control subjects with a high proportion of coronary artery disease (CAD), and has the ability to evaluate SCD risk across the spectrum of LVEF values. The geographically distinct, but otherwise similarly designed and conducted Ventura Prediction of Sudden Death in Multi-ethnic Communities (PRESTO) study provides the opportunity to perform an external validation of such a risk score. Therefore, we developed and validated a clinical risk prediction score to identify patients who present with SCA manifesting as VF or VT that could potentially be applied to patient populations at intermediate risk of SCA.

Methods

This study received the proper ethical oversight and was approved by institutional review boards at Oregon Health & Science University, Cedars-Sinai Medical Center and Ventura County Medical Center.

Study Design

CAD is the most common underlying risk factor for SCA, but it is not a specific marker of SCA risk. Therefore, we designed our study to control for underlying CAD so as to identify risk factors beyond CAD. Although our SCA cases were drawn from the general population, as expected, ~70% had documented CAD before arrest or CAD identified at arrest. (8) Our control group was selected to represent an intermediate-risk subset of the general population (with ~70% prevalence of CAD) to facilitate development of a risk score that could be applied in clinical practice to intermediate-risk patients populations rather than the general population.

Study population

Study participants were identified (2002–2019) from the Oregon SUDS, an ongoing investigation of SCA in the Portland, Oregon, metropolitan area. Methods have been described in detail previously. (6,7) Briefly, consecutive individuals with out-of-hospital SCA were identified prospectively through collaboration with the region’s emergency medical services (EMS) system and state medical examiner’s office (Central Illustration). A 3-physician panel adjudicated each case to determine if they met inclusion criteria for SCA of likely cardiac etiology, using all available information from patients’ existing medical records, the EMS pre-hospital report with circumstances of arrest, and autopsies when available. Patients with noncardiac causes of SCA or terminal illness were excluded (eg, trauma, overdose, pulmonary embolism, stroke, cancer not in remission, severe lung disease on home oxygen). Oregon SUDS control subjects were identified and enrolled prospectively in the same period from the same geographic region and were selected to have a high proportion with documented CAD but no history of ventricular arrhythmias or SCA, from multiple sources. These included subjects transported by EMS for symptoms of acute coronary ischemia; patients undergoing angiography or visiting an outpatient cardiology clinic at one participating hospital; and members of a regional health maintenance organization.

Inclusion criteria

All cases and controls eligible for the present analysis were age ≥18 with lifetime clinical history available from archived medical records. Because the focus of this analysis was to identify factors related to treatable SCA, only case subjects with an EMS-documented shockable rhythm (VF/VT) at the time of cardiac arrest (one-third of total cases, Figure 1) were included. All individuals with an EMS-documented non-shockable rhythm (pulseless electrical activity or asystole) or with missing rhythm were excluded (Figure 1). Among cases with a shockable rhythm, most (93%) had VF and the remainder (7%) had VT.

Figure 1. Cases and Controls in the Training, Internal Validation, and External Validation Datasets for Prediction of SCA Presenting With VF/VT.

The analysis was conducted among sudden cardiac arrest (SCA) cases with shockable rhythms (VF/VT). Cases with nonshockable presenting rhythm (PEA/asystole) or with missing rhythm information were excluded. The analysis dataset from the Oregon SUDS discovery population was divided into the training dataset (67%) for development of the prediction models, and the validation dataset (33%) to test the prediction models. Prediction models were externally validated in the geographically distinct Ventura PRESTO study. ECG = electrocardiogram; PEA = pulseless electrical activity; PRESTO = Prediction of Sudden Death in Multi-ethnic Communities; SUDS = Sudden Unexpected Death Study; VF = ventricular fibrillation; VT = ventricular tachycardia.

Data sources

For each SCA case and control, we reviewed archived medical records to obtain a complete clinical history, including cardiovascular risk factors, cardiac tests/procedures, history of cardiac events and non-cardiac comorbidities. Archived 12-lead electrocardiograms (ECGs) were obtained (before and unrelated to the SCA event for cases). When more than 1 ECG was available, the one closest and unrelated to the SCA event (or the ascertainment date for controls) was selected. Archived echocardiograms were obtained in a similar manner. Available clinical history was required for this analysis (resulting in inclusion of 88% of total cases and 100% of total controls), while ECGs and echocardiograms were available for a subset of participants.

External validation

SCA cases for the external validation population were identified (2015–2020) from the Ventura PRESTO study, an ongoing population-based investigation of SCA in Ventura County, California. Methods for case ascertainment, adjudication, inclusion criteria, and data retrieval and definitions were identical to those used in Oregon SUDS. Validation control subjects were selected from out-patients in a health maintenance organization in the region (Cedars-Sinai Medical Center Medical Network, Los Angeles, CA; n=4251) that were age- and sex-matched to all PRESTO SCA cases. From this larger sample, 2 controls per VF/VT case were selected using further frequency matching by race/ethnicity. We prioritized inclusion of subjects with documented coronary disease, resulting in a control group with 77% documented coronary disease.

Statistical Methods:

Study variables:

To construct the risk score model, we considered patient demographics (age, sex, and race/ethnicity), and the following variables based on prior reported associations with SCD: Clinical predictors (n = 12) included smoking (current smoker vs not), body mass index (BMI, calculated as weight [kg]/height [m]²), diabetes mellitus, hypertension, heart failure, stroke, atrial fibrillation, history of myocardial infarction, chronic obstructive pulmonary disease (COPD), chronic renal insufficiency, seizure disorder, and history of syncope. (7,9–18)

ECG predictors (n=11) (19,20) :

ECG predictors (n = 11) included heart rate, corrected QT interval (Bazett’s), Tpeak-Tend interval, delayed intrinsicoid deflection, left ventricular hypertrophy (LVH) by ECG, QRS duration, QRST angle, delayed QRS transition, left bundle branch block, prolonged PR interval, and left atrial enlargement.(19,20)

Echocardiographic predictors (n=4)(21,22):

Echocardiographic predictors (n = 4) included LVEF, LVH defined as an LV mass index greater than 134 g/m² for men and 110 g/m² for women, moderate to severe mitral regurgitation, and moderate to severe aortic stenosis.(21–23)

Training and internal validation sets:

We divided the Oregon SUDS case-control data into a training dataset (67%) and a validation dataset (33%), each with 4 separate analysis data subsets. Dataset 1 (all subjects, regardless of ECG and echocardiogram data available) and Dataset 4 (restricted to subset with both ECG and echocardiogram data available) were the primary analysis datasets (Figure 1). Dataset 2 (restricted to subset with ECG data, regardless of echocardiogram availability) and Dataset 3 (restricted to subset with echocardiogram data, regardless of ECG availability) were used for secondary analysis (Supplementary Figure). These data subsets were constructed using stratified random sampling based on data availability, pooling cases, and controls to ensure that both the training and validation datasets had approximately equal proportions of ECG and echo data available for both cases and controls. For internal training and validation datasets, no matching was performed on age, sex, or other patient characteristics.

Preliminary case-control analysis:

We used Student t tests and chi square tests to compare means and frequencies of each predictor in cases and controls in the training and validation datasets 1 through 4. SAS version 9.4 (SAS Institute, Cary, North Carolina) was used for all analyses. To assess potential bias from complete case analysis in subgroups, we compared effect estimates from logistic regression models which included the same clinical predictors in datasets 1 through 4. The effect estimates for each clinical predictor were relatively consistent across models, indicating that while the prevalence of some clinical predictors varied across data subsets, the strength of association with SCA was stable.

Development of prediction models in the training dataset:

We used backwards stepwise logistic regression in the training dataset to build prediction models, retaining variables with p<0.20. Five separate models were constructed: model 0 (reference model, with LVEF only); model 1 (clinical variables only from dataset 1); model 2 (clinical variables and ECG predictors from dataset 2); model 3 (clinical variables and echocardiogram predictors from dataset 3); and model 4 (clinical, ECG, and echocardiogram predictors from dataset 4). Primary analysis focused on Models 1 and 4 (Figure 1), with Models 2 and 3 as secondary models. ROC curves (C statistics) were used to evaluate model discrimination, i.e., the ability of the model to separate individuals who experienced SCA from those who did not.(24) We also performed a sensitivity analysis by restricting cases and controls to individuals with established coronary disease.

Internal validation:

We applied the fixed beta coefficients from the logistic regression models developed in the training datasets to the validation datasets using the “score” statement in SAS. We did this separately for Models 1 through 4. For each model, the C statistic was interpreted as the overall performance of the prediction model as applied to the respective validation dataset.

Risk score development:

The risk score (VFRisk) was developed using variables retained in Model 4 (clinical, ECG, and echo predictors). Points were assigned based on the odds ratio for each predictor from the training dataset, rounded to the nearest tenth. Secondary risk scores were developed from Model 2 (clinical plus ECG) and Model 3 (clinical plus echocardiogram) in the same manner. ORs associated with a 1-unit increase in the risk score, and by risk score quartile, were calculated in the training and validation datasets separately to evaluate the risk score’s utility. Finally, in the combined training and validation datasets we evaluated the consistency of risk score performance in population subgroups by testing for interaction between the VFRisk score and age, sex, and LVEF, and by calculating odds ratios by VFRisk quartiles stratified by age, sex, and LVEF. We applied the fixed beta coefficients from the training logistic regression models to the external validation datasets using the “score” statement in SAS and used C statistics to evaluate the performance of the prediction model in the external validation dataset. We did this separately for Models 1 and 4. VFRisk points were then assigned to each individual in the external validation dataset based on their clinical, ECG, and echocardiogram profile. Finally, in the external validation population we calculated the OR associated with a 1-unit increase in the risk score, and ORs by risk score quartile.

Results

Demographics of SCA

A total of 1,374 SCA cases (33% of total cases; 77% male, mean age 64 ±15 years) ascertained in the Oregon SUDS from 2002 to 2019 presented with VF/VT and met criteria for analysis. A total of 1,600 control subjects (67%male, mean age 65 ± 13 years) were enrolled during the same period and met criteria for inclusion in this analysis. The full case-control dataset was divided into a training dataset with 919 cases and 1074 controls, and an internal validation dataset with 455 cases and 526 controls (Figure 1). In both the training and validation datasets, case and control mean ages were similar and cases were more likely to be male (Table 1).

Table 1.

Characteristics of cases (SCA presenting with VF/VT) and controls in the training and internal validation datasets.

	Training Dataset*				Internal Validation Dataset*
	N=721				N=355
	Cases	Controls		p-value	Cases	Controls	p-value
	N=230	N=491			N=115	N=240
DEMOGRAPHICS
Age, yrs (mean ± SD)	68 ± 13	68 ± 11		0.70	68 ± 13	67 ± 12	0.96
Male	74%	66%		0.03	66%	65%	0.90
Race / ethnicity				0.37			0.33
White	83%	88%			85%	87%
Black	10%	8%			12%	7%
Asian	3%	2%			1%	2%
Hispanic	1%	1%			2%	3%
Other	3%	1%			0%	1%
CLINICAL HISTORY
Cardiovascular Risk Factors
Current smoker	17%	15%	0.67		27%	20%	0.14
Diabetes	49%	34%	<0.001		53%	35%	0.002
Hypertension	83%	81%	0.43		83%	81%	0.76
Body mass index (mean ± SD)	31.1 ± 9.3	29.9 ± 7.0	0.10		30.3 ± 9.5	30.2 ± 6.7	0.90
Prevalent Cardiovascular
Disease
Heart failure	60%	25%	<0.001		63%	22%	<0.001
Stroke	25%	11%	<0.001		16%	13%	0.56
Atrial fibrillation	51%	29%	<0.001		43%	26%	0.001
MI	56%	38%	<0.001		54%	43%	0.06
Comorbidities
COPD	22%	12%	0.001		26%	17%	0.047
Chronic renal insufficiency	41%	23%	<0.001		35%	17%	<0.001
Seizure disorder	5%	2%	0.01		3%	2%	0.43
Syncope	14%	6%	<0.001		8%	5%	0.38
ECG PARAMETERS
Heart rate >75 bpm	46%	31%	<0.001		55%	33%	<0.001
QTc >=460 female, >=450 male	63%	40%	<0.001		70%	42%	<0.001
Tpeak-end (lead V5) > 89	53%	42%	0.01		48%	43%	0.44
Delayed intrinsicoid deflection	44%	21%	<0.001		46%	24%	<0.001
LVH by EKG	19%	11%	0.004		24%	14%	0.02
QRSd > 110 ms	43%	24%	<0.001		43%	26%	0.002
QRST angle >90	45%	24%	<0.001		53%	30%	<0.001
Delayed QRS transition	46%	29%	<0.001		41%	28%	0.02
Left bundle branch block	12%	3%	<0.001		13%	7%	0.06
PR interval prolonged	24%	16%	0.01		17%	19%	0.73
Left atrial enlargement	29%	19%	0.005		39%	23%	0.004
ECHOCARDIOGRAM
PARAMETERS
LV ejection fraction (mean ± SD)	47 ± 16	55 ± 12	<0.001		47 ± 16	54 ± 13	<0.001
LVEF categories			<0.001				<0.001
EF ≥55	40%	65%			39%	65%
EF 36–54	33%	24%			32%	22%
EF≤35	27%	11%			29%	13%
LVH by echo	40%	15%	<0.001		44%	23%	<0.001
Mitral regurgitation	14%	5%	<0.001		17%	6%	0.001
Aortic stenosis	7%	6%	0.49		8%	6%	0.37

^*Values shown are percentages unless otherwise stated. Dataset 4, complete data for clinical, ECG, and echocardiogram variables. COPD = chronic obstructive pulmonary disorder; ECG = electrocardiogram; LVEF = left ventricular ejection fraction; LVH = left ventricular hypertrophy; MI = myocardial infarction; SCA = sudden cardiac arrest; VF = ventricular fibrillation; VT = ventricular tachycardia.

Case and control characteristics: training dataset

Several cardiovascular risk factors, prevalent cardiac disease, and non-cardiac comorbidities were more common among cases than controls, including current smoking, diabetes, obesity, prevalent heart failure, and a history of stroke, atrial fibrillation, seizures, and syncope (Table 1). In the complete data subset, all abnormal ECG findings were significantly more prevalent among cases, as were all abnormal echo findings, except for aortic stenosis (Table 1).

Multivariable prediction models

Supplemental Table 1 shows side-by-side results from models 1 through 4, including ORs for all variables retained (P < 0.20) in the backward stepwise logistic regression models, and significant variables (P < 0.05) in bold font. Model 1 retained all clinical predictors except chronic renal insufficiency. Model 4 with clinical, ECG, and echocardiogram variables was largely consistent with results from Models 2 and 3, except that LVEF was no longer significantly associated with SCD when accounting for both clinical and ECG markers, and LVH by ECG was no longer significant when the echocardiogram variables were included. Several clinical variables were dropped (current smoking, hypertension, body mass index) when ECG or echocardiogram variables were included.

Table 2 presents the C statistics from the multivariable prediction models in the training and validation datasets. In all datasets, Model 4 with variables from the ECG and echo had significantly higher C statistics than Model 0 (LVEF only) and Model 1 (clinical variables only). Figure 2 shows the ROC curves from all models in the training dataset. Model 1 (clinical variables only) did not perform as well (C=0.688, 95% CI 0.665–0.712) when compared to Model 4 with variables from both ECG and echo (C=0.808, 95% CI 0.774–0.827). Model 2 with variables from the ECG (C=0.788, 95% CI 0.762–0.815) and Model 3 with variables from the echo (C=0.776, 95% CI 0.743–0.809) performed substantially better than Model 1, but not as well as Model 4. All models, however, performed substantially better than the model with LVEF alone (C=0.638, 95% CI 0.598–0.678) (Figure 2).

Table 2.

C statistics evaluating the performance of prediction models in the training and validation datasets

	C Statistic (95% CI)
	Training dataset	Internal Validation dataset	External Validation dataset
Model 0: Ejection fraction only*	0.638 (0.598 – 0.678)	0.643 (0.586 – 0.700)	0.701 (0.641 – 0.760)
Model 1: Clinical variables only†	0.688 (0.665 – 0.712)	0.644 (0.609 – 0.678)	0.679 (0.644 – 0.715)
Model 4: Clinical + ECG + Echo*	0.808 (0.774 – 0.842)	0.776 (0.725 – 0.827)	0.782 (0.718 – 0.846)

^*In Dataset 4 (training n=721, internal validation n=355, external validation n=255)

^†In Dataset 1 (training n=1993, internal validation n=981, external validation n=1011)

Figure 2. ROC curves for the different VF/VT prediction models in the training dataset based on the type of variables included.

Prediction models shown include (model 0) left ventricular ejection fraction (EF) as the only predictor; (model 1) clinical (medical history) predictors only; (model 2) clinical plus ECG predictors; (model 3) clinical plus echocardiogram predictors; and (model 4) clinical, ECG, and echocardiogram predictors. N shown in legend is the number of subjects in the dataset for each model. Echo, echocardiogram; ROC, receiver operating curve; other abbreviations as in Figure 1.

The VFRisk final prediction model (Model 4) included 13 components retained in the backward stepwise logistic regression models at p<0.20 (Table 3): 8 clinical predictors (history of diabetes, heart failure, stroke, atrial fibrillation, myocardial infarction, COPD, seizure disorder, and syncope); 4 ECG predictors (resting heart rate, QTc interval, Tpeak-Tend interval, and delayed intrinsicoid deflection); and 1 echocardiogram marker (LVH).

Table 3.

Independent predictors of SCA and points for VFRisk risk score in Model 4, training dataset.

	Univariate odds ratio (95% CI)	Adjusted odds ratio* (95% CI)	Points for Risk Score
Cardiovascular risk factors
Diabetes	1.84 (1.34 – 2.53)	1.40 (0.95 – 2.06)	1.4
Prevalent cardiovascular disease
MI	2.02 (1.47 – 2.78)	1.44 (0.99 – 2.10)	1.4
Atrial fibrillation	2.54 (1.84 – 3.51)	1.77 (1.20 – 2.61)	1.8
Stroke	2.78 (1.84 – 4.21)	1.65 (1.00 – 2.72)	1.7
Heart failure	4.31 (3.09 – 6.02)	2.14 (1.43 – 3.21)	2.1
Comorbidities
COPD	1.96 (1.30 – 2.96)	1.46 (0.89 – 2.38)	1.5
Seizure disorder	3.03 (1.20 – 7.65)	3.49 (1.24 – 9.86)	3.5
Syncope	2.77 (1.63 – 4.71)	2.80 (1.48 – 5.28)	2.8
ECG parameters
Heart rate >75 bpm	1.87 (1.36 – 2.59)	1.48 (1.00 – 2.18)	1.5
QTc ≥460 ms female, ≥450 ms male	2.58 (1.86 – 3.59)	1.71 (1.16 – 2.53)	1.7
Tpeak-end (lead V5) ≥90 ms	1.55 (1.11 – 2.16)	1.66 (1.11 – 2.46)	1.7
Delayed intrinsicoid deflection ≥50 ms	2.92 (2.07 – 4.14)	2.11 (1.41 – 3.17)	2.1
Echo parameters
LVH by echocardiogram	3.69 (2.49 – 5.47)	2.39 (1.51 – 3.79)	2.4

^*Odds ratios are adjusted for all other variables in the model.

Internal validation of prediction models

When the fixed coefficients from the prediction models in the training dataset were applied to the validation dataset, the C statistics indicated very good discrimination (Table 2; C statistic 0.776 for Cohort 4).

Risk Score (VFRisk)

The points used to construct the VFRisk score are shown in Table 3. A total of 13 components resulted in a median score of 6.2 (min 0 – max 18.1) in the training dataset and 6.4 (min 0 – max 18.5) in the internal validation dataset. The median risk score among cases in the training and validation datasets was 8.8 and 9.1, respectively, while among controls it was 4.9 in the training dataset and 5.2 in the validation dataset (Table 4).

Table 4.

Risk prediction according to VFRisk point score in the training and validation datasets.

	Training n=721	Internal Validation n=355	External Validation n=255
VFRisk points, median (min - max)
All subjects	6.2 (0 – 18.1)	6.4 (0 – 18.5)	6.6 (0 – 17.5)
Cases	8.8 (0 – 18.1)	9.1 (0 – 18.5)	9.5 (0 – 17.5)
Controls	4.9 (0 – 15.3)	5.2 (0 – 14.6)	5.7 (0 – 15.9)
Odds Ratio (95% CI), per 1-unit increase in VFRisk score	1.35 (1.28 – 1.42)	1.31 (1.21 – 1.40)	1.33 (1.22 – 1.45)
Odds Ratio (95% CI), by VFRisk quartile
Quartile 1 (0 – 4.6 points) (Reference)	1.0	1.0	1.0
Quartile 2 (4.7 – 7.2 points)	1.3 (0.7 – 2.4)	3.5 (1.4 – 8.8)	2.5 (0.9 – 7.1)
Quartile 3 (7.3 – 10.0 points)	4.6 (2.7 – 8.0)	5.7 (2.3 – 14.0)	2.4 (0.9 – 6.7)
Quartile 4 (≥10.1 points)	12.1 (7.0 – 20.8)	16.6 (7.0 – 39.3)	14.9 (5.6 – 39.9)

External validation

Characteristics of the external validation population are shown in Supplementary Table 2. The prediction models performed well in the external validation dataset and the C statistic indicated very good discrimination (Table 2) (C statistic 0.782 for cohort 4). Median risk scores in the external validation dataset were similar to those in the training dataset (9.5 among cases and 5.7 among controls) (Table 4). Figure 3 shows the distribution of VFRisk among cases and controls in the training, internal validation, and external validation datasets. The performance of VFRisk was consistent across training and validation datasets. A 1-unit increase in the risk score was associated with a 35% increase in odds of VF/VT (training dataset OR: 1.35, 95% CI: 1.28 to 1.42) (Table 4), with similar results in the internal validation dataset (OR: 1.31, 95% CI: 1.21 to 1.40). ORs by VFRisk quartile, with the lowest quartile as reference, were also consistent in the training and validation datasets, with the highest risk score quartile associated with a 12- to 16-fold higher odds of VF than the lowest quartile (Table 4). ORs for a unit change in VFRisk were similar in the external validation dataset (OR: 1.33, 95% CI: 1.22 to 1.45). Results of the sensitivity analysis performed by restricting cases and controls to individuals with documented coronary artery disease were consistent for both internal and external validation (data not shown).

Figure 3. Distribution of VFRisk Score in Cases and Control.

Histogram showing distribution of (A) VFRisk score among cases and (B) controls in the training, internal validation, and external validation datasets.

Performance of VFRisk in population subgroups

Based on the consistent results in the training and validation datasets, we evaluated the performance of VFRisk by population subgroup in the combined Oregon SUDS dataset (Figure 4). Models with interaction terms between VFRisk score and age, sex, and LVEF revealed no significant interaction, indicating similar performance across subgroups. ORs increased across quartiles (with the lowest quartile as reference) for all subsets by age, sex, and LVEF.

Figure 4. Performance of the VFRisk Score in Distinct Population Subgroups in the combined Training and Internal Validation Dataset.

Odds ratios and 95% CIs by VFRisk quartile, by age, left ventricular (LV) ejection fraction (EF), and sex, in the training dataset.

Discussion

A clinical algorithm for prediction of VF/VT (VFRisk) was developed from individuals who suffered out-of hospital SCA with documented VF/pulseless VT (Central Illustration) and control subjects of whom the majority (67%) had significant coronary artery disease. VFRisk, constructed with 13 clinical, ECG, and echocardiographic variables had very good discrimination in the training dataset as well as the internal validation dataset. VFRisk was also externally validated successfully, in a geographically distinct population. Secondary risk prediction models that included clinical variables combined with either ECG or echocardiographic risk markers also performed well. All of these clinical algorithms substantially outperformed use of LVEF ≤35% as a risk stratification tool and performed equally well across all sub-categories of LVEF.

To our knowledge, VFRisk is the first clinical algorithm assembled for prediction of SCA presenting with documented VF/VT (ie, potentially avertable SCD) in a population representing individuals at intermediate risk of SCA. Previously published SCD risk scores have identified variables in addition to LVEF from cohort studies that contribute to improvements in risk stratification.(10,12) However, these studies lacked information regarding presenting rhythm and therefore could not specifically predict risk of ventricular tachyarrhythmias potentially treatable by the ICD. (9,10,12,17) Individuals who suffer SCA can manifest with rhythms potentially treatable with an electric shock (ie, VT or VF) or rhythms that are nonshockable (ie, pulseless electrical activity or asystole).(2) Effectiveness of pacing or shocking therapies delivered from primary prevention ICDs is largely dependent on the patient manifesting with VF/VT.(25) It follows that a focus on prediction of VF/VT is likely to be useful for primary prevention of SCD using the ICD. Also, in existing cohort studies few participants had LVEF <50% at baseline and the contribution of severely reduced LVEF could not be incorporated in the prediction models.(12) Other ad hoc analyses of clinical trial data were restricted to LVEF <35% and therefore not able to address risk prediction across the LVEF spectrum.(11,13)

Although collectively the composition of VFRisk is unique, several component variables have been reported in existing risk scores including diabetes, heart failure, myocardial infarction, stroke, resting heart rate, and QTc interval. (9,10,12,26) On the other hand, all of the remaining components of the VFRisk are novel as incorporated in a risk score. All of these clinical and ECG variables have been reported as having individual associations with increased risk of SCD.(14–16, 19, 27, 28) The inclusion of echocardiographic LVH in a clinical risk score for SCD is also new. (21)

The final prediction model that included clinical, ECG and echocardiographic variables had the best performance (C-statistic 0.808). With a C-statistic of 0.788 the model with clinical and ECG variables was a close second. However, all 3 secondary risk prediction models developed with simple clinical variables only, or with addition of ECG or echocardiography variables performed better than use of the LVEF alone (Figure 2). The VFRisk score performed equally well in 3 important and separate subpopulations: age younger or older than 65 years, sex, and LVEF less than or greater than 35% (Figure 4).

Currently there are no available clinical predictors for most individuals who suffer SCD with LVEF >35%, comprising at least two-thirds of all SCD cases. This risk score for VF/VT comprised of noninvasive, broadly available, and relatively inexpensive clinical markers performed as well among individuals with LVEF >35%, indicating its potential for developing risk stratification methodology for SCD among individuals with LVEF >35%. VFRisk could facilitate the design of randomized clinical trials of the primary prevention ICD among patients with LVEF >35%. Furthermore, among those with LVEF <35%, there is a need to refine candidate selection for the primary prevention ICD by balancing high risk of VF/VT with lowest risk of noncardiac mortality from comorbid conditions. (29) VFRisk could also enable optimization of candidate selection for primary prevention ICD among patients with LVEF≤35%.

Study Limitations:

The identification of feasible numbers of out-of-hospital SCA cases manifesting with VF/VT mandates the use of a prospective, population-based design. Some inherent limitations of this type of study design should be considered while interpreting these findings. Due to the absence of established cardiac conditions or warning signs, a sizeable proportion of community residents may not have approached health care providers for evaluation before their unexpected SCA event. Others may have had variable types of clinical evaluations and treatments performed. Therefore, information regarding clinical profile and tests incorporated in the clinical scores was not uniformly available for all SCA cases. In addition, some individuals excluded from this analysis may have initially presented with VF/VT which then deteriorated into a nonshockable rhythm (asystole or PEA) by the time EMS recorded their first monitored rhythm, but given the setting of SCA, these are not identifiable. Because there is almost complete certainty that patients with VF or pulseless VT as a first monitored rhythm had a shockable (avertable) presenting rhythm, we proceeded with the “cleanest” dataset that we could identify (ie, individuals whose EMS-documented rhythm was VF/VT). Finally, whereas this study has identified and validated a novel clinical algorithm that predicts SCA presenting with a shockable rhythm, further evaluation in randomized clinical trials will be needed before potential clinical application.

Conclusions:

VFRisk, a clinical algorithm composed of widely available and noninvasively obtained variables, successfully predicted occurrence of SCA manifesting as documented VF/VT across the LVEF spectrum. The targeted identification of high-risk candidates with potentially treatable SCA could enhance the effectiveness of sudden death prevention especially among individuals with mid-range or preserved LVEF.

Clinical Perspectives.

Competency in Medical Knowledge:

SCD remains a major public health problem with a critical need to enhance risk prediction and primary prevention. Patients who experience SCA can present with VF/VT, PEA, or asystole. The only lethal rhythms amenable to treatment with the implantable cardioverter-defibrillator are VF/VT, with the potential to avert SCD. A clinical algorithm focused on prediction of VF/VT could have implications for enhancement of risk stratification and primary prevention of SCD.

Translational Outlook 1:

We developed a novel and user-friendly clinical prediction algorithm (VFRisk) for documented VF/VT constructed with 13 clinical, ECG, and echocardiographic variables. VFRisk performed well and was successfully validated in internal and external datasets.

Translational Outlook 2:

VFRisk substantially outperformed left ventricular ejection fraction (LVEF) ≤ 35% and performed well across the LVEF spectrum. Further evaluation in randomized clinical trials is warranted before potential clinical application.

Abbreviations List

SCD

Sudden cardiac death

VF

Ventricular fibrillation

VT

Ventricular tachycardia

SCA

Sudden cardiac arrest

LVEF

Left ventricular ejection fraction

PEA

Pulseless electrical activity

SUDS

Sudden Unexpected Death Study

PRESTO

Prediction of Sudden Death in Multi-ethnic Communities

CAD

Coronary artery disease

EMS

Emergency medical services