Predicting Out-of-Hospital Cardiac Arrest in the General Population Using Electronic Health Records

Abstract

BACKGROUND:

The majority of out-of-hospital cardiac arrests (OHCAs) occur among individuals in the general population, for whom there is no established strategy to identify risk. In this study, we assess the use of electronic health record (EHR) data to identify OHCA in the general population and define salient factors contributing to OHCA risk.

METHODS:

The analytical cohort included 2366 individuals with OHCA and 23 660 age- and sex-matched controls receiving health care at the University of Washington. Comorbidities, electrocardiographic measures, vital signs, and medication prescription were abstracted from the EHR. The primary outcome was OHCA. Secondary outcomes included shockable and nonshockable OHCA. Model performance including area under the receiver operating characteristic curve and positive predictive value were assessed and adjusted for observed rate of OHCA across the health system.

RESULTS:

There were significant differences in demographic characteristics, vital signs, electrocardiographic measures, comorbidities, and medication distribution between individuals with OHCA and controls. In external validation, discrimination in machine learning models (area under the receiver operating characteristic curve 0.80–0.85) was superior to a baseline model with conventional cardiovascular risk factors (area under the receiver operating characteristic curve 0.66). At a specificity threshold of 99%, correcting for baseline OHCA incidence across the health system, positive predictive value was 2.5% to 3.1% in machine learning models compared with 0.8% for the baseline model. Longer corrected QT interval, substance abuse disorder, fluid and electrolyte disorder, alcohol abuse, and higher heart rate were identified as salient predictors of OHCA risk across all machine learning models. Established cardiovascular risk factors retained predictive importance for shockable OHCA, but demographic characteristics (minority race, single marital status) and noncardiovascular comorbidities (substance abuse disorder) also contributed to risk prediction. For nonshockable OHCA, a range of salient predictors, including comorbidities, habits, vital signs, demographic characteristics, and electrocardiographic measures, were identified.

CONCLUSIONS:

In a population-based case–control study, machine learning models incorporating readily available EHR data showed reasonable discrimination and risk enrichment for OHCA in the general population. Salient factors associated with OCHA risk were myriad across the cardiovascular and noncardiovascular spectrum. Public health and tailored strategies for OHCA prediction and prevention will require incorporation of this complexity.

Sudden cardiac death (SCD), defined as an unexpected death with rapid loss of spontaneous circulation and cardiac mechanical activity, is a major public health burden, estimated to account for 15% to 20% of global mortality.¹ Out-of-hospital cardiac arrest (OHCA) is the most common context for SCD. Despite advances in resuscitation science and care delivery,² survival after OHCA remains poor, with a median survival of ≈10%.^3,4 Mitigating the global burden of SCD requires effective upstream prevention, including accurate prediction of SCD and identification of modifiable risk.⁵

There are several challenges to the prediction and prevention of SCD. First, the majority of SCD occurs in the general population, in which individual risk is low (≈0.05% to 0.1% per year).¹ Traditional observational epidemiological approaches are generally underpowered in this setting. Second, contemporary prevention strategies for SCD, such as the use of an implantable cardioverter defibrillator,⁶ are restricted to patients at high absolute risk, who comprise a small minority of the global burden of SCD. Third, given the frequent out-of-hospital context of SCD, adjudication is challenging and often uses epidemiological definitions,⁷ with inference regarding timing and mechanism of death.

In this study, we link a population-based emergency medical services (EMS) OHCA registry with a longitudinal electronic health record of ≈1.5 million patients. Using a case–control study design, we train and test machine learning (ML) models for the prediction of OHCA and further define the most salient factors contributing to OHCA risk prediction.

METHODS

Study Cohort

The study cohort comprised individuals who sought care at the University of Washington (UW) between 2010 and 2021. UW Medicine is a large academic health system comprising 3 hospitals providing care for 1.5 million unique patients over the study period. UW Medicine is located in Seattle, King County, which is a metropolitan region with a population of 2.3 million. Cardiac arrest cases were identified within the UW Medicine health system by cross-referencing all health encounters within the study period against a comprehensive registry of OHCA in King County and Seattle. The EMS system of Greater King County and Seattle maintain a comprehensive registry of all EMS-treated cardiac arrests, with information regarding arrest rhythm, presentation, treatment, and outcome (Methods in the Supplemental Material).^8,9

OHCA cases with antecedent health care within the UW Medicine system were identified by matching name (first, last) and date of birth. OHCA cases were considered if the date of arrest within the EMS registry was between January 1, 2013, and December 31, 2021. Exclusion criteria included a traumatic mechanism of OHCA, age <18 years or >90 years at the time of first health encounter, and a history of cardiac arrest, cardiopulmonary resuscitation, or presence of an implantable cardioverter defibrillator at the time of first health encounter (details regarding exclusion criteria are provided in the Methods in the Supplemental Material). We identified 2366 OHCA cases in this manner. The control population was selected from the remainder of individuals with health encounters within the UW Medicine system over the study period (additional details regarding identification of controls are provided in the Methods in the Supplemental Material). Controls were matched to individuals with OHCA by age (by decade) and sex at a 1:10 case:control ratio, yielding 23 660 controls (Figure 1).

Figure 1. Overview of study design and electronic health record measures used in risk prediction models.

Overview of the analytical cohort, which identified 2366 individuals with out-of-hospital cardiac arrest (OHCA) who had received antecedent health care at the University of Washington over the study period. Highlighted are the exclusion criteria as well as the categories of measures included in machine learning models (sociodemographic variables, vital signs, electrocardiographic measures, clinical comorbidities, and medications). Further details regarding these categories are available in the Methods in the Supplemental Material and in Table 1.

Data collection and analyses for the study were approved by the review boards of the University of Washington and Public Health–Seattle and King County. The data that support the findings of this study are available from the corresponding author upon reasonable request.

Ascertainment of Covariates

The UW Medicine enterprise data warehouse includes all electronic patient records from medical sites across the UW Medicine system, gathering data from >60 sources, including laboratory results, microbiology reports, demographic data, diagnosis codes, medication prescriptions, and any electronic note text. Covariates of interest extracted from the enterprise data warehouse included sociodemographic information, vital signs, comorbidities, habits, electrocardiographic measures, health care engagement metrics, and medication use (Methods in the Supplemental Material and Tables S1 and S2).^10,11

The window of covariate ascertainment for cases was the first qualifying health visit until the day before the date of cardiac arrest. For controls, the window of ascertainment was first qualifying health visit until the first instance of death (adjudicated using documentation within the enterprise data warehouse and the Washington State Death Index), date of last known health care contact documented within the enterprise data warehouse, or December 31, 2021.

There was missingness, as anticipated, for electrocardiographic and vital signs covariates (Table S3), and these values were imputed (Methods in the Supplemental Material). We assessed the fidelity of imputation by examining association of imputed and nonimputed variables with known clinical outcomes (Table S4). Imputed variables demonstrated the same significant association with established outcomes as nonimputed variables.

Statistical Analysis

Descriptive statistics, including comparison of demographic and comorbidity comparisons for OHCA cases and controls, were performed using t tests or χ² tests as appropriate. Descriptive statistics were also performed to compare shockable and nonshockable OHCA. We selected 3 ML approaches for the prediction of OHCA: Ridge regression, LASSO (least absolute shrinkage and selection operator) regression, and gradient boosted trees (XGBoost). LASSO and Ridge models were selected because they yield computationally efficient generalized linear models incorporating high-dimensional data. To capture potential salient interactions between high-dimensional covariates, we also used a flexible nonlinear model (XGBoost). All covariates were included in the models including follow-up time (see Methods 1.4 in the Supplemental Material for comprehensive list of covariates included in all models). Data were split by subject into 70% training and 30% testing. Internal validation was performed on the training set using cross-validation. The testing set was used as external validation. This data split was selected to minimize risk of overfitting and obtain accurate estimates of model performance. Stratified sampling was used to create balanced data sets. In the Ridge and LASSO regression models, cross-validation was used to tune the regularization measure. The Ridge and LASSO models used the imputed data sets, whereas XGBoost was used without imputation because it automatically handles data with missingness. The ML models were compared with a baseline model comprising previously published covariates associated with SCD in the general population^12,13: smoking, diabetes, body mass index, mean systolic blood pressure, mean diastolic blood pressure, antihypertensive medication use, and lipid-lowering medication use.

For each model, performance was assessed using area under the receiver operating characteristic curve (AUC). The 95% CIs for AUC were constructed by means of 1000 bootstraps. Model AUCs (each ML model versus baseline) were compared using the DeLong test for 2 unpaired receiver operating characteristic curves with Holm-Bonferroni correction for multiple comparisons. To evaluate the effect of an unbalanced case:control ratio, F1 scores and area under the precision recall curve were calculated. We then calculated model sensitivity, positive predictive value (PPV), and negative predictive value for 3 a priori specificity thresholds (0.90, 0.95, and 0.99), selected on the basis of evaluating the use of models as screening tools. Given the influence of disease incidence on model performance, metrics were adjusted for the observed rate of OHCA in the entire UW Medicine health system (Methods in the Supplemental Material).¹ To understand the relative importance of covariates contributing to prediction, covariate importance was derived for each model. For Ridge and LASSO, covariates were scaled in a harmonized manner, and the absolute values of coefficients were used to assess rank importance. For XGBoost, covariate gain, which is a quantitative measure of covariate contribution to outcome prediction in a tree-based model, was evaluated. In a sensitivity analysis, covariate importance was evaluated in the nonimputed cohort (ie, complete case cohort). In secondary analysis, model performance and covariate importance were evaluated for shockable and nonshockable OHCA. Statistical analysis was conducted using R or SAS (version 9.4). A 2-sided P<0.05 was considered to indicate statistical significance.

RESULTS

Study Population

Baseline characteristics of the study cohort are shown in Table 1. There were 2366 individuals with OHCA (mean age, 61±16 years; 64% male) and 23 660 individuals without OHCA (mean age, 61±16 years; 64% male). Median follow-up was 2 years (median [range], 729 [50–2167] days). There were significant differences in demographic characteristics, health care engagement, vital signs, electrocardiographic measures, clinical comorbidities, and medication distribution. People with OHCA were less likely to be White, to be married, and to have private insurance compared with controls (P≤0.0001 for all comparisons). People with OHCA had a higher mean heart rate, systolic blood pressure, and respiratory rate, and lower mean oxygen saturation (P≤0.002 for all comparisons). Individuals with OHCA had a longer mean PR interval, wider QRS duration, and longer corrected QT interval (P≤0.005 for all comparisons). There were several comorbidities and medications enriched in those with OHCA across the cardiometabolic, gastrointestinal, renal, substance abuse, endocrine, neurological, and infectious spectrum (Table 1; Table S5). Differences in baseline covariates without imputation were similar (Table S6).

Table 1. Baseline Characteristics of the Study Cohort.

Characteristics	OHCA (n=2366)	Control (n=23 660)	P value
Sociodemographic covariates
Age, y	61±16	61±16	0.29
Male sex	1504 (64)	15 040 (64)	1.0
Race and ethnicity
White	1441 (61)	15 723 (66)	<0.0001
Black	370 (16)	1260 (5)
Asian	214 (9)	1829 (8)
Native Hawaiian or Pacific Islander	43 (2)	139 (1)
American Indian or Alaskan Native	72 (3)
Other*	226 (10)	4428 (19)
Marital status
Married	766 (33)	10 971 (46)	<0.0001
Divorced	261 (11)	1368 (84)
Single	892 (38)	4917 (21)
Widowed	173 (7)	1230 (5)
Other	274 (12)	5174 (22)
Insurance type
Public	1591 (69)	10 598 (47)	<0.0001
Private	530 (23)	10 438 (46)
Other	117 (5)	1325 (6)
Self	47 (2)	378 (2)
Unknown	81(3)	921 (4)
Primary language
English	2093 (88)	21 503 (91)	<0.0001
Spanish	37 (2)	534 (2)
Other	236 (10)	1623 (7)
Health care access covariates
No-show appointments	1 (0–6)	0 (0–1)	<0.0001
Total hospitalizations	5 (0–5)	0 (0–1)	<0.0001
Emergency department visits	0 (0–3)	0 (0–0)	<0.0001
Vital sign covariates
Heart rate, bpm
Minimum	72±16	70±15	<0.0001
Maximum	91±19	83±17	<0.0001
Mean	81±14	76±13	<0.001
SBP, mm Hg
Minimum	119±22	121±20	<0.0001
Maximum	144±24	139±22	<0.0001
Mean	131±18	130±18	0.002
DBP, mm Hg
Minimum	69±15	72±14	<0.0001
Maximum	86±15	83±13	<0.0001
Mean	78±12	78±11	0.92
Respiratory rate, breaths per minute
Minimum	16±6	15±5	<0.0001
Maximum	20±12	18±8	<0.0001
Mean	18±6	16±5	<0.0001
Spo2, %
Minimum	95±7	96±7	<0.0001
Maximum	98±3	98±2	<0.0001
Mean	96±4	97±2	<0.0001
Weight, kg
Minimum	81±24	80±21	0.37
Maximum	89±29	85±24	<0.0001
Mean	85±25	83±22	<0.0001
Electrocardiographic covariates
PR interval, ms
Minimum	157±39	160±37	0.001
Maximum	171±46	164±44	<0.0001
Mean	164±40	162±39	0.005
QRS duration, ms
Minimum	94±25	91±22	<0.0001
Maximum	101±30	94±24	<0.0001
Mean	97±27	92±22	<0.0001
QT corrected, ms
Minimum	443±38	427±40	<0.0001
Maximum	467±51	434±41	<0.0001
Mean	454±38	430±38	<0.0001
Clinical comorbidities
Hypertension	1439 (61)	8965 (38)	<0.0001
Diabetes	770 (33)	3981 (17)	<0.0001
Hyperlipidemia	1219 (52)	9254 (39)	<0.0001
Heart failure	602 (25)	22 029 (7)	<0.0001
Coronary heart disease	680 (28)	2916 (12)	<0.0001
Valvular heart disease	453 (19)	1662 (7)	<0.0001
Stroke or TIA	452 (19)	1843 (8)	<0.0001
Myocardial infarction	387 (16)	1131 (5)	<0.0001
Peripheral arterial disease	472 (20)	2183 (9)	<0.0001
Cerebral atherosclerosis	552 (23)	2669 (11)	<0.0001
Chronic kidney disease	665 (28)	2235 (9)	<0.0001
Renal failure	550 (23)	1985 (8)	<0.0001
Substance abuse disorder	614 (26)	1120 (5)	<0.0001
Depression	801 (34)	3964 (17)	<0.0001
Obstructive sleep apnea	254 (11)	1102 (5)	<0.0001
Cardiac arrhythmias	750 (32)	2466 (10)	<0.0001
Ventricular tachycardia	121 (5)	295 (1)	<0.0001
Atrial fibrillation	413 (17)	2056 (9)	<0.0001
Palpitations	466 (20)	1709 (7)	<0.0001
Syncope	339 (14)	1238 (5)	<0.0001
Cirrhosis	472 (20)	2085 (9)	<0.0001
Gastrointestinal bleeding	96 (4)	341 (1)	<0.0001
Peptic ulcer without bleeding	80 (3)	293 (1)	<0.0001
Smoking	1050 (44)	4798 (20)	<0.0001
Pulmonary disease	803 (34)	3815 (16)	<0.0001
Obesity	559 (24)	4427 (19)	<0.0001
Weight loss	428 (18)	1628 (7)	<0.0001
Anemia	734 (31)	2907 (12)	<0.0001
HIV	52 (2)	233 (1)	<0.0001
Alcohol abuse	528 (22)	1270 (5)	<0.0001
Autoimmune conditions	103 (4)	878 (4)	<0.0001
Coagulopathy	369 (16)	1551 (7)	<0.0001
Dementia	180 (8)	970 (4)	<0.0001
Hypothyroidism	314 (13)	2344 (10)	<0.0001
Neurological disorders	313 (13)	1280 (5)	<0.0001
Paralysis	226 (10)	733 (3)	<0.0001
Epilepsy	287 (12)	827 (4)	<0.0001
Psychoses	604 (26)	1760 (7)	<0.0001
Pulmonary circulation disease	245 (10)	671 (3)	<0.0001
Any malignancy	336 (14)	3656 (15)	<0.0001
Fluid and electrolyte disorders	1081 (46)	3799 (16)	<0.0001

Values are mean±SD, n (%), or median (interquartile range), as appropriate. bpm indicates beats per minute; DBP, diastolic blood pressure; HIV, human immunodeficiency virus; OHCA, out-of-hospital cardiac arrest; SBP, systolic blood pressure; and TIA, transient ischemic attack.

^*Other reflects any race or ethnicity not reflected in the above, including individuals identifying as ‘mixed race or ethnicity’.

In a secondary analysis, we compared characteristics between individuals with shockable and nonshockable OCHA (Table S7). There were no differences in the prevalence of several cardiovascular risk factors or comorbidities between the 2 groups, including hypertension, heart failure, diabetes, atrial fibrillation, obesity, and cerebral atherosclerosis. Coronary heart disease, hyperlipidemia, and myocardial infarction were more prevalent in shockable OHCA compared with nonshockable OHCA (P≤0.03 for all comparisons), whereas chronic kidney disease, substance abuse, depression, cirrhosis, smoking, pulmonary disease, anemia, dementia, malignancy, and fluid and electrolyte disorders were enriched in those with nonshockable OHCA compared with shockable OHCA (P≤0.01 for all comparisons; Table S7).

Prediction of OHCA

We examined model performance of a baseline model inclusive of established predictors of OHCA and ML-based models for prediction of OHCA (Table 2; Figure 2). In external validation, discrimination of ML models (AUC 0.80–0.85) was superior to the baseline model (0.66; Bonferroni-Holm adjusted P<0.001 for AUC comparison of each ML model versus baseline). In external validation, the estimated PPV of ML models corrected for observed OHCA rate across the entire health system was 2.5% to 3.1% for ML models and 0.8% for the baseline model (Table 2; Table S8).

Table 2. Model Performance for Prediction of Out-of-Hospital Cardiac Arrest.

Variable	Internal validation (n=18 218)				External validation (n=7808)
	Baseline	Ridge	Lasso	XGBoost	Baseline	Ridge	Lasso	XGBoost
AUC (95% CI)	0.68 (0.66–0.69)	0.82 (0.81–0.83)	0.82 (0.81–0.83)	0.95 (0.94–0.95)	0.66 (0.64–0.68)	0.80 (0.79–0.82)	0.81 (0.79–0.82)	0.85 (0.84–0.87)
Sensitivity*	0.04	0.15	0.15	0.51	0.04	0.14	0.14	0.18
PR-AUC	0.18	0.40	0.39	0.75	0.18	0.36	0.36	0.42
PR-AUC, adjusted†	0.004	0.02	0.02	0.31	0.004	0.02	0.02	0.02
PPV*	0.27	0.61	0.60	0.84	0.31	0.58	0.59	0.64
PPV, adjusted†	0.007	0.027	0.027	0.084	0.008	0.025	0.026	0.031
NPV*	0.91	0.92	0.92	0.95	0.91	0.92	0.92	0.92
NPV, adjusted†	0.99	0.99	0.99	0.99	0.99	0.99	0.99	0.99

AUC indicates area under the receiver operating characteristic curve; LASSO, least absolute shrinkage and selection operator; NPV, negative predictive value; PPV, positive predictive value; PR-AUC, area under the precision recall curve; and XGB, XG Boost.

^*At a specificity threshold of 0.99.

^†Denoted model measures are sensitive to underlying disease incidence. Adjustment reflects estimated performance when deployed across the entire University of Washington health system (see Methods, Supplemental Material).

Figure 2. Internal and external validation receiver operating characteristic curves for prediction of out-of-hospital cardiac arrest.

Receiver operating characteristic (ROC) curves for the baseline model (purple) and machine learning models including ridge regression (red), LASSO (least absolute shrinkage and selection operator) regression (green), and XGBoost (XGB; blue). The left panel represents internal validation (70% training set) and the right panel represents external validation (30% holdout testing set; see Methods).

We next examined covariate importance in the ML models (Figure 3; Table S9). Salient categories of covariates that contributed to an increased risk of OHCA included sociodemographic factors (Black race, single marital status, insurance status), vital signs (higher heart rate, lower oxygen saturation, higher systolic and diastolic blood pressure), habits (alcohol abuse, substance abuse disorder), electrocardiographic pathology (longer corrected QT interval), comorbidities (fluid and electrolyte disorder, diabetes, obesity), and health care engagement (greater number of no-show visits). There was significant agreement in terms of covariate importance across all models (Figure 3; Table S9), with longer corrected QT interval, substance abuse disorder, alcohol abuse, fluid and electrolyte disorder, and higher heart rate identified as predictors in all 3 models. In a sensitivity analysis, covariate importance was generally similar when evaluated within the nonimputed data (Table S10).

Figure 3. Covariate importance consensus among machine learning models predicting out-of-hospital cardiac arrest.

Covariates contributing most to the prediction of out-of-hospital cardiac arrest. Covariates identified in multiple models are shown in the overlap of the Venn diagram. For regression-based models (LASSO [least absolute shrinkage and selection operator], Ridge), the directionality of association with out-of-hospital cardiac arrest is highlighted with italicized covariates associated with a lower risk of out-of-hospital cardiac arrest. For tree-based models, directionality is not salient, and thus directionality is not indicated. BP indicates blood pressure; and XGB, XGBoost.

Prediction of Shockable Versus Nonshockable OHCA

Model performance was also evaluated in the prediction of shockable and nonshockable OHCA (Tables S11 and S12). Similar to the prediction of overall OHCA, discrimination in external validation was improved with ML models (shockable AUC 0.77–0.80; nonshockable AUC 0.82–0.87) compared with the baseline model (shockable AUC 0.61; nonshockable AUC 0.68; Bonferroni-Holm adjusted P<0.001, comparison of each ML model versus baseline). In external validation, at a specificity threshold of 0.99 and with correction for the observed OHCA rate across the health system, the PPV of the ML model was superior to the baseline model for both shockable OHCA (corrected PPV 1.7%–1.8% for ML models versus 0.5% for the baseline model) and nonshockable OHCA (corrected PPV 2.6%–3.7% for ML models versus 0.9% for the baseline model; Tables S11 and S12).

For shockable OHCA, established cardiovascular risk factors retained predictive importance (myocardial infarction, longer corrected QT interval, hyperlipidemia, heart failure, higher mean heart rate, higher systolic or diastolic blood pressure, antianginal and antihypertensive medication use), but there were other covariates of significance, including race (Native Hawaiian/Pacific Islander, Alaskan Native or Native American, Black), substance abuse disorder, single marital status, and no-show visits (Table S13). For nonshockable OHCA, there were a range of salient predictors, including comorbidities (fluid and electrolyte disorders, psychoses, diabetes), habits (substance abuse disorder, smoking, alcohol abuse), vital signs (lower oxygen saturation, higher heart rate), demographic characteristics (minority race, single marital status, and insurance type), and electrocardiographic measures (longer corrected QT interval; Table S14).

DISCUSSION

In a large US health system with longitudinal electronic health record data linked to EMS-adjudicated OHCA, we show that predictors of OHCA in the general population are myriad. Sociodemographic factors (self-identified Black race, single marital status, insurance status), vital signs (higher heart rate, lower oxygen saturation, higher blood pressure), electrocardiographic measures (longer corrected QT interval), cardiovascular risk factors or disease surrogates (diabetes, antianginal medication use), and noncardiovascular comorbidities (fluid and electrolyte disorders, substance and alcohol abuse disorder, psychosis) contributed the most to OHCA prediction. ML models using electronic health record data demonstrated reasonable discrimination and modest PPV, outperforming conventional clinical models of OHCA prediction.

The study has 3 important findings. First, OHCA risk is complex, and underlies a range of sociodemographic, cardiovascular, and noncardiovascular factors. Previous studies of OHCA pathogenesis, centered in primarily White populations, have suggested that coronary heart disease underlies OHCA in 75% to 80% of patients overall,^14–16 with women having a lower prevalence of coronary disease (45% to 50%).^17,18 A recent autopsy study of OHCA decedents identified coronary heart disease as pathogenic in 32%.¹⁹ In this study, we show that established coronary disease was present in a minority (28%) of individuals with OHCA, with significant contribution of noncardiovascular comorbidities to OHCA risk prediction, including fluid and electrolyte disorders, as well as alcohol and substance abuse disorder. Black race,²⁰ longer QT interval,^20–22 higher mean heart rate,^23,24 and single marital status²⁵ have been previously identified as risk factors for SCD in the general population, and were also identified in this study as important contributors to OHCA risk. Taken together, these findings highlight the evolving epidemiology of OHCA and support future efforts to understand the ways in which elements beyond cardiovascular factors underlie OHCA risk. Strategies to mitigate the global burden of sudden death will require identification of modifiable risk reflective of this shifting epidemiology.

Second, classifying OHCA as shockable or nonshockable, although practical during acute resuscitation, is reductive and underestimates the complexity of OHCA phenotypes. Shockable OHCA, attributed to lethal ventricular arrhythmias, has historically been linked to cardiovascular risk factors, whereas nonshockable OHCA has been linked to noncardiovascular factors.¹ Here, we show that several cardiovascular risk factors and comorbidities were equally prevalent in shockable and nonshockable OHCA. Likewise, we found that some noncardiovascular risk factors, including substance abuse disorder, contributed substantially to shockable OHCA risk prediction, suggesting that mitigating OHCA burden will similarly require attention to a broader spectrum of risk factors beyond cardiovascular risk.

Third, ML methods that include a broad range of possible predictors yield good discrimination (AUC 0.80–0.85) and modest PPV for OHCA in the general population. Previous studies in the general population centered on cardiovascular risk factors^12,13 or complex electrocardiographic measures^20,26 demonstrated numerically lower discrimination of OHCA risk, similar to the baseline conventional model in our study. When considering actionable clinical deployment of an OHCA risk model in the general population, indices such as PPV are contingent on underlying incidence of disease, which, in the instance of OHCA, is estimated at ≈0.1% per year.¹ In this study, when we adjust model performance metrics for the observed rate of OHCA across the health system, the PPV for OHCA was 2.5% to 3.1% over a median follow-up of 2 years, or ≈1.3% to 1.6% per year. Put plainly, ML models derived in this study enriched OHCA risk ≈13- to 16-fold: from 1 in 1000/year to 1 to 2 in 100/year. Such OHCA models could provide a foundation for a staged screening approach,²⁷ wherein individuals identified as high risk would be candidates for further targeted evaluation linked to patient-level factors driving OHCA risk. In their optimal deployment, such risk scores could be incorporated as back-end core data services within an electronic health record with dynamic updating of risk after each health encounter coupled with best practice advisory flagging of individuals above a given risk threshold.

Limitations

There are limitations to this study. First, these data reflect care within a large quaternary-care health system, for which underlying comorbidity distribution and health care engagement may differ from other communities, for which these results may not generalize. Validation in other study communities is warranted. Second, the findings of the study were not oriented toward causal inference, and the pathogenic relationship of salient covariates to OHCA is not established. Third, given the real-world nature of the data, there was missingness requiring imputation for variables of interest. We demonstrated the fidelity of our imputation approach (Table S4) and found similar performance in tree-based models that did not require imputation. Fourth, patients with OHCA who were not attended to by EMS were not included in this study and could have led to misclassification of controls. Fifth, comorbidities were assessed using previously validated billing codes, but misclassification is possible. Sixth, model performance measures were calculated with explicit correction for the observed rate of OHCA in the study health system. These results represent an approximation of performance in this setting and warrant prospective validation. Seventh, as anticipated in a general population cohort, echocardiographic data were only available in a small minority of individuals. Given that routine echocardiography in the general population is not scalable or consistent with standard of care, we elected not to include this measure in predictive models. The role of echocardiographic measures or other unstructured data elements such as radiographic findings in predictive models of OHCA in the general population warrants future investigation. Finally, medication use was on the basis of prescription and retained covariate importance in models. Incorporation of medication compliance or usage could have complementary value to risk prediction.

Conclusions

Use of longitudinal electronic health record data in the general population highlights the sociodemographic, cardiovascular, and noncardiovascular complexity underlying OHCA. Effective reduction in the global burden of OHCA will require public health strategies reflective of this complexity. ML approaches incorporating longitudinal health record data in the general population may serve as a reasonable screening tool to discriminate OHCA risk and serve as a foundation for subsequent evaluation.

Supplementary Material

Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/CIRCULATIONAHA.124.069105.

Methods

Tables S1–S14

Clinical Perspective.

What Is New?

• There are no established approaches or guidelines for out-of-hospital cardiac arrest (OHCA) prevention in the general population.

• Linkage of a comprehensive emergency medical services OHCA registry with a large health system allowed for evaluation of OHCA in the general population, including its prediction and salient factors associated with risk.

• Machine learning models incorporating readily available electronic health record data in the general population enriched OHCA risk ≈15-fold, from ≈1 in 1000/year to 1 to 2 in 100/year after accounting for the observed rate of OHCA across the health system.

• Health system–level factors associated with OHCA were myriad, including sociodemographic, cardiovascular, and noncardiovascular factors.

What Are the Clinical Implications?

• Public health efforts to mitigate the substantial public health burden of OHCA will require attention to the contemporary epidemiology of OHCA risk, reflecting both cardiovascular and noncardiovascular risk factors.

• Machine learning approaches integrating readily available electronic health record data could serve as a reasonable screening paradigm for OHCA risk and a foundation for further targeted screening and intervention on the basis of patient-level factors driving OHCA risk.

Nonstandard Abbreviations and Acronyms

AUC

area under the receiver operating characteristic curve

EMS

emergency medical services

LASSO

least absolute shrinkage and selection operator

ML

machine learning

OHCA

out-of-hospital cardiac arrest

PPV

positive predictive value

SCD

sudden cardiac death

UW

University of Washington