Long-term outcomes of out-of-hospital cardiac arrest care at regionalized centers

Abstract

Background:

It is unknown whether regionalization of post-arrest care via inter-facility transfer to cardiac arrest receiving centers (CARCs) reduces mortality. We sought to evaluate whether treatment at a CARC, whether via direct transport or early interfacility transfer, is independently associated with long-term outcome.

Methods:

We performed a retrospective cohort study including adults resuscitated from out-of-hospital cardiac arrest in Southwestern Pennsylvania and neighboring Ohio, West Virginia and Maryland, which includes approximately 5.7 million residents in urban, suburban and rural counties. Patients were treated by one of 78 ground emergency medical services agencies or two air medical transport agencies between January 1^st, 2010 and November 30^th, 2014. Our primary exposures of interest were inter-facility transfer to a CARC within 24 hours of arrest, or any treatment at a CARC regardless of transfer status. Our primary outcome was vital status assessed through December 31^st, 2014 using National Death Index records. We used unadjusted and adjusted survival analyses to test the independent association of CARC care, whether through direct or interfacility transport, on mortality.

Results:

Overall, 5,217 cases were observed for 3629 person-years with 3,865 total deaths. Most patients (82%) were treated at 42 non-CARCs with median annual volume of 17 (interquartile range 1 – 53 cases per center annually) while 18% were cared for at CARCs receiving at least 1 inter-facility transfer per month. In adjusted models, treatment at a CARC was independently associated with reduced hazard of death compared to treatment at a non-CARC (adjusted HR 0.84, 95%CI 0.74 – 0.94). These effects were unchanged when analysis was restricted to patients brought from the scene to the treating hospital. No other hospital characteristic, including total out-of-hospital cardiac arrest patient volume or cardiac catheterization capabilities, independently predicted outcome.

Conclusion:

Both early inter-facility transfer to a CARC and direct transport to a CARC from the scene are independently associated with reduced mortality.

Introduction

Sudden cardiac arrest (SCA) is the most common cause of death in high-income countries.¹ Most patients who are transported to the hospital after achieving return of spontaneous circulation (ROSC) do not survive to discharge, although outcome varies both regionally and between hospitals.^2–5 Because of this variability in outcomes and the complexity of SCA care, both the American Heart Association and National Academies of Medicine suggest development of regional systems of care.^{6, 7} Although hospital-level characteristics such as case volume or advanced procedural capabilities are inconsistently associated with in-hospital mortality,^{4, 8–12} small studies suggest long-term benefit from care at specialty centers that provide a comprehensive bundle of SCA care.^{13, 14} Similarly, allowing emergency medical services (EMS) to bypass nearby hospitals in favor of transport directly to specialized centers has been associated improved outcomes in several regions.^{15, 16}

Multiple strategies have been employed to ensure appropriate triage and access to specialized care. Trauma systems designate tiers of hospitals based on available resources and case volume, and EMS or initial receiving hospitals direct transport of patients to these centers based on severity of injury.^{17, 18} Current treatment of acute stroke similarly employs many primary stroke centers that provide rapid access to time-sensitive early interventions, followed by transfer of specific patients to comprehensive stroke centers for complex ongoing specialty care.¹⁹ Traditionally, SCA patients are delivered to the nearest facility after ROSC. Aside from single center reports, it is unknown whether subsequent transfer to a receiving center for continued acute and post-acute care affects SCA patient outcomes.

We examined initial treatment and long-term survival after SCA in a geographic region with multiple health systems to test the hypotheses that treatment at a center receiving a high volume of inter-facility SCA referrals, whether through direct transport from the scene by EMS providers or early inter-facility transfer within 24 hours of arrest, is independently associated with reduced long-term mortality.

Methods

Study Setting

Southwestern Pennsylvania and neighboring Ohio, West Virginia and Maryland includes approximately 5.7 million residents in urban, suburban and rural counties over 43.3 thousand square miles. The 78 ambulance services included in this study perform >95% of EMS transports in the region. Ground transports are performed with a combination of basic and advanced life support ambulances staffed with Emergency Medical Technicians and paramedics, who operate within a single regional system and use statewide treatment protocols. Cardiac arrest responses in the region are dispatched as a single tier advanced life support (ALS) response staffed at the paramedic level and may be assisted by first responders at the emergency medical responder level. EMS medical directors from four major health networks provide individual medical oversight of these agencies: UPMC, Allegheny Health Network, St. Clair Hospital, and Excela Health. Two air medical transport agencies are responsible for interfacility transports of critically ill patients among 46 short-term acute care hospitals in the catchment area for these health networks (Figure 1).

Figure 1:

Cardiac arrest scene locations and included acute care hospitals in Southwestern Pennsylvania and surrounding Ohio, West Virginia and Maryland.

We assembled a multi-institution collaboration to generate a regional cohort of EMS-transported patients resuscitated from out-of-hospital SCA. The University of Pittsburgh served as the coordinating center. All aspects of the present study were approved with a waiver of informed consent by the Institutional Review Boards of the University of Pittsburgh, Allegheny Health Network, St. Clair Hospital and Excela Health.

Data Sources

Co-investigators at each site queried the medical records of the EMS agencies under their medical command from January 1, 2010 to November 30, 2014 for cases of out-of-hospital SCA. During the study period, each of these agencies used a common National EMS Information Systems (NEMSIS)-compliant web-based prehospital electronic health record (emsCharts, Warrendale, PA) (see Supplemental Appendix for details). During the study period, the system was NEMSIS version 2.2 compliant. emsCharts has custom reporting software that allows identification of cases and reporting of data based on multiple fields in the record. Once the electronic query is built, the contents of specified fields are pulled and aggregated automatically, without the need for manual chart abstraction (see Supplemental Appendix for details). We queried emsCharts to identify likely cases of SCA with transport to a hospital during the study period, which we considered to be any patients with a medical category of “cardiac arrest” or documentation of any of the following in the EMS record: chest compressions, defibrillation, or automated defibrillator use. Each site abstracted identifiable data from emsCharts including patient name, address, social security number, age, race and sex, which we used to query the National Death Index (NDI) for patients’ vital status through December 31^st, 2014. The NDI is a comprehensive database of vital statistics obtained from state death records that has been used extensively to determine vital status.^{13, 20–23} We linked NDI results to patient records using probabilistic linkage (Supplemental Appendix),¹³ then de-identified and aggregated each sites’ data for analysis.

Inclusion and Exclusion

We included patients treated and transported by EMS after resuscitation from SCA from January 1, 2010 to November 30, 2014. We excluded patients with age <16 years; those where resuscitation was not attempted because of obvious signs of death in the field; those transported with ongoing cardiopulmonary resuscitation (CPR) and without ROSC at the time of transfer of care to the initial receiving hospital; those transported to or from hospitals outside Southwestern Pennsylvania; and, those with no receiving hospital listed (i.e. termination of resuscitation in the field). We implemented these exclusions both through the initial emsCharts query and then again during data cleaning and aggregation. We included patients regardless of arrest etiology (traumatic vs “medical”) for several reasons. First, out-of-hospital data are not sufficient to define arrest etiology in many cases (e.g. a low velocity motor vehicle crash may occur secondary to a “medical” cardiac arrest, while syncope and ground level fall may lead to arrest from cervical spine fracture that is not recognized until advanced imaging is performed. Second, the historical dichotomization of medical vs traumatic arrests does not parallel observed pathophysiological differences across distinct arrest etiology. “Medical” etiologies like subarachnoid hemorrhage, heroin overdose and acute myocardial infarction may in fact differ more from each other than exsanguination from gastrointestinal hemorrhage or spontaneous pneumothorax do from traumatic etiology arrests. As a post hoc analysis, we repeated all statistical procedures excluding arrests presumed to be trauma to confirm the stability of our findings.

Exposure, covariates and outcomes

Our primary exposures of interest were 1) inter-facility transfer within 24 hours of presentation to the initial receiving hospital and 2) final treatment at a facility receiving a high volume of inter-facility SCA transfers regardless of route of presentation (transfer by EMS directly from the scene or inter-facility transfer). We linked EMS records for scene to initial hospital transports and for inter-facility transports using probabilistic linkage on first and last name, date of birth, social security number, and hospital (i.e. the initial receiving hospital for a scene chart matched the referring hospital for inter-facility chart). We linked charts matching with a probability ≥0.75, which corresponded to near-perfect matching on at least 4 of 5 criteria. We then manually inspected matches with probabilities ≥0.5 to link these charts based on full available data.

We additionally abstracted clinical covariates of interest including age, sex, race, and past medical history. EMS providers classified race from a drop-down list of option. For past medical history classification, we used text-based search strategies of free text medical history documentation to categorize the presence or absence of: automatic implantable cardioverter-defibrillator or pacemaker in situ; atrial fibrillation; coronary artery disease or past myocardial infarction; congestive heart failure; prior cardiac arrest; cardiac conduction disease; structural heart disease; venous thromboembolism; vascular disease; alcohol abuse; tobacco abuse; other recreational drug abuse; anemia; any malignancy; hypertension; hyperlipidemia; diabetes; cirrhosis; neurological disease or dementia; psychiatric disease; renal disease or dialysis dependence; pulmonary or respiratory tract disease; gastrointestinal disease; other past medical history not classified as one of the above; unknown past medical history; or, no past medical history. We treated each of these as a binary predictor. Arrest-specific covariates included arrest etiology, which we dichotomized as traumatic or medical; witnessed collapse; provision of bystander CPR; shockable arrest rhythm at any time during resuscitation; number of epinephrine doses administered during CPR; EMS dispatch to CPR interval; CPR to ROSC interval; and, transport time from scene to first receiving hospital.

We considered the treating hospital to be the last hospital at which the patient received care (i.e. receiving hospital from the scene for non-transferred patients and receiving hospital from inter-facility transport for transferred patients). We inspected overall SCA case volume per center for natural breakpoints in the data, and categorized hospitals as low volume centers if they treated fewer than 100 OHCA patients over the 4-year study period, moderate volume centers if they treated 100 to 399 patients, and high volume centers if they treated at least 400 patients. Next, we evaluated only the number of patients received by each center through inter-facility transfer, and categorized hospitals as Cardiac Arrest Receiving Centers (CARCs) if they received at least one patient via inter-facility transfer every quarter during the study period. As a sensitivity analysis, we varied this cutoff as described below, and compared our results. This designation was developed based on observed referral patterns in the present cohort, and was not known to EMS agencies at the time of patient care.

We defined additional hospital level characteristics using 2010 Centers for Medicare and Medicaid Services Hospital Cost Report Information System (HCRIS), a publically available hospital-level database with detailed information on structural data on all U.S. hospitals. These included teaching status (using the resident-to-bed ratio, classifying hospitals as nonteaching if they had no resident trainees, small teaching if the ratio was more than zero and less than 0.2, and large teaching if the ratio was 0.2 or greater), total number of licensed beds, and total number of licensed intensive care unit beds, which we tested both as continuous predictors and categorized by quartile. Using 2010 Centers for Medicare and Medicaid Services MedPar data we classified hospitals according to their annual cardiac catheterization procedure volume (≥160 vs <160 cases/year) and annual volume of mechanically ventilated patients (≥400 vs <400 adjusted mechanical ventilation cases annually). We used ICD-9-CM procedure codes to identify cases of cardiac catheterization (37.22, 37.23, 88.53, 88.54, 88.55, 00.66, 36.04, 36.06, 36.07) and mechanical ventilation (96.70, 96.71, and 96.72).

Our primary outcome of interest was vital status through December 31^st, 2014, which we identified through NDI search results.

Statistical analysis

We used descriptive statistics to summarize overall population characteristics. We used survival analysis to estimate the effects of inter-facility transfer to or initial treatment at a CARC on hazard of death after up to 5 years of observation. Our a priori analysis plan was to use hierarchical Cox proportional hazards models to control for patient- and hospital-level characteristics in unadjusted and adjusted regression. However, the relationship between transfer and outcome was complex and time varying. Therefore, we used piecewise exponential survival regression, splitting the data into blocks 0 to 1, 1 to 3, 3 to 7, and >7 days post-arrest. We selected these time epochs based on biological plausibility and previous research.^24–26 Briefly, we believed that during days 0 to 1 and 1 to 3 patients would be at high risk for rearrest or early withdrawal of life sustaining treatment for perceived poor neurological prognosis;^{24, 25} day 3 to 7 to be the period during which delayed neurological prognostication is recommended by consensus guidelines;^{26, 27} and, beyond day 7 to be the post-acute period during which a majority of subjects would have revealed their neurological trajectory but still be at risk for medical sequelae of OHCA, critical illness and pre-existing comorbidities.^{13, 24, 25}

We first estimated univariable patient- and hospital-level associations of predictors with mortality (adjusted only for time epoch), using clustered sandwich estimators to account for correlation between patients within the same hospital. For variables missing not at random, we included a predictor level of “missing” in our analysis. We assessed pairwise correlations of independent predictors to ensure there was no severe multicollinearity. Then, we used a backwards stepwise approach as further exploration of our cohort, sequentially eliminating predictors with an adjusted association with outcome at a threshold of P >0.1 to build a final adjusted regression model. In this final model, we explored and included significant interaction terms between predictors and time epoch. Next, we calculated coefficients for the effect of each treating hospital on outcome both with and without adjustment for patient-level characteristics. To describe between-hospital variation in outcome, we computed all pairwise comparisons of these coefficients to calculate exact unadjusted and adjusted median hazards ratios (MHR).²⁸

We performed a number of sensitivity and secondary analyses to confirm the robustness of our main results. First, we compared various piecewise splits of our time variable, up to the extreme case splitting at every death to obtain perfectly constant hazard functions. Second, to minimize the potential of referral bias in patients undergoing inter-facility transfer, we tested the independent effect of CARC status on outcome restricting analysis to the cohort of non-transferred patients. Third, we varied our case volume cutoffs defining overall SCA case volume and CARC status, defining CARCs variably as receiving between 4 and 48 total transfers during the study period. Fourth, we tested for the possibility of systematically biased EMS-level transport of SCA patients from the scene to the initial treating hospital based on CARC status. To do this, we constructed a propensity score predicting transport to the closest hospital versus EMS bypass of the closest hospital. This score accounted for patient-level characteristics (transfer status; age; sex; race; medical vs traumatic arrest etiology; shockable rhythm; witnessed status; bystander CPR; dispatch to CPR interval; epinephrine doses administered; and the 26 binary past medical history variables listed in Table 1), transport time, as well as distance to, density of and CARC status of the hospitals surrounding the arrest scene. We included transfer status in this model to account for potential survival bias whereby transferred patients might be expected to be healthier (since they survived transfer) and referral bias, whereby transferred patients might be expected to be sicker (since they were deemed to exceed the capabilities of the referring hospital). We used this model instead of simply adjusting for bypass of the closest hospital in our initial regression model because our region’s CARCs are located in areas with a high density of surrounding hospitals. This results in patients transported to CARCs having greater opportunity to have bypassed the nearest facility by chance due to overall hospital density, without necessarily implying systematically biased selection of initial hospital by CARC status. We then built a propensity score-adjusted piecewise exponential survival model to test the independent effect of CARC status and transfer status on outcome. Finally, we performed a series of post hoc analyses to confirm our findings were robust. We removed traumatic arrests from the cohort and repeated the adjusted analyses described above. Next, to account for potential secular trends, we tested for an association of calendar quarter or year with outcome. We assessed the fit of survival models by examining plots of deviance and Cox Snell residuals. We used Stata Version 14.2 (StataCorp, College Station, TX) for all statistical analyses and ArcGIS 10.4 (ESRI, Redlands, CA) for all geospatial analyses.

Table 1:

Baseline population characteristics, stratified by treatment at a Cardiac Arrest Receiving Center (CARC) or non-CARC.

Characteristic	Overall cohort (n = 5,217)	Treated at CARC (n = 920)	Treated at non-CARC (n = 4,297)
Transferred for care	390 (7)	350 (38)	40 (1)
Age, years	65 ± 18	61 ± 18	66 ± 18
Female sex	2,046 (39)	363 (39)	1,683 (39)
Race
White	3,743 (72)	456 (50)	3,287 (77)
Black	426 (8)	75 (8)	351 (8)
Other	45 (1)	8 (1)	37 (1)
Unknown	1,003 (19)	381 (41)	622 (14)
Non-traumatic arrest	5,002 (96)	819 (89)	4,183 (97)
Shockable rhythm	2,488 (48)	526 (57)	1,962 (46)
Witnessed status
Unwitnessed	1,425 (27)	339 (37)	1,086 (25)
Witnessed	2,493 (48)	511 (56)	1,982 (46)
Missing	1,299 (25)	70 (9)	1,229 (29)
Bystander CPR
None	150 (3)	33 (4)	117 (3)
Layperson	1,194 (23)	231 (25)	963 (22)
Professional	1,239 (24)	246 (27)	993 (23)
Unknown	2,634 (51)	410 (45)	2,224 (52)
Dispatch to CPR interval
< 10 min	257 (5)	68 (7)	189 (4)
10–19 min	347 (7)	100 (11)	247 (6)
≥20 min	502 (10)	136 (15)	366 (9)
Missing	4,111 (79)	616 (70)	3,495 (81)
Epinephrine doses, mg
1 to 2	2,357 (21)	241 (26)	794 (18)
3 to 4	2,867 (25)	238 (26)	1,329 (31)
5 to 6	1,565 (14)	112 (13)	847 (20)
>6	702 (6)	46 (5)	380 (9)
None documented	3,975 (35)	273 (30)	947 (22)
Time from scene to receiving hospital
< 10 min	5,423 (50)	335 (47)	2,268 (53)
≥ 10 min	4,791 (45)	369 (51)	1,811 (42)
Missing	544 (5)	16 (2)	190 (4)
Documented past medical history
AICD/PPM	316 (6)	50 (5)	269 (6)
Atrial fibrillation	388 (7)	73 (8)	315 (7)
CAD/MI	1,077 (21)	168 (18)	909 (21)
Congestive heart failure	673 (13)	226 (13)	557 (13)
Cardiac arrest	58 (1)	11 (1)	47 (1)
Cardiac conduction disease	77 (1)	19 (2)	58 (1)
Structural heart disease	114 (2)	20 (2)	94 (2)
Venous thromboembolism	102 (2)	78 (2)	24 (3)
Vascular disease	194 (4)	26 (3)	168 (4)
Alcohol abuse	117 (2)	26 (3)	91 (2)
Tobacco abuse	55 (1)	11 (1)	44 (1)
Other drug abuse	213 (4)	46 (5)	167 (4)
Anemia	204 (4)	30 (3)	174 (4)
Malignancy	559 (11)	86 (9)	473 (11)
Diabetes	1,198 (23)	199 (22)	999 (23)
Hypertension	1,856 (36)	306 (33)	1,550 (36)
Hyperlipidemia	493 (9)	96 (10)	397 (9)
Cirrhosis	96 (2)	27 (3)	69 (2)
Neurological disease	968 (19)	131 (14)	837 (19)
Psychiatric disease	629 (12)	105 (11)	524 (12)
Renal disease/dialysis	500 (10)	98 (11)	402 (9)
Pulmonary/respiratory disease	1,102 (21)	191 (21)	911 (21)
Gastrointestinal disease	506 (10)	93 (11)	413 (10)
Other past medical history	1,799 (35)	280 (30)	1,519 (35)
Unknown past medical history	559 (11)	126 (14)	433 (10)
No past medical history	601 (12)	120 (13)	481 (11)

Categorical data are presented as raw number (percent), and continuous data are presented as mean ± standard deviation.

Abbreviations: CARC – Cardiac arrest receiving center; CPR – Cardiopulmonary resuscitation; AICD/PPM – Automatic implantable cardioverter/defibrillator or permanent pacemaker; CAD/MI – Coronary artery disease or myocardial infarction

Results

Our initial query identified 7,887 cases of EMS-transported OHCA during the study period. Of these, we excluded 2,670 (Figure 2), leaving 5,217 cases and 3,629 person-years of observation. There were 3,865 deaths (74% of subjects) with the remaining subjects surviving until December 31^st 2014. Overall, median survival time was one day (interquartile range (IQR) 1 – 1309 days), and 2,659 (51%) of patients died on this day. Day one mortality differed significantly by CARC status (240/920 (26%, 95%CI 23% – 29%) for CARCs vs 2,419/4,297 (56%, 95%CI 55% – 58%) for non-CARCs; P <0.001). Conditional on surviving through day 1, overall median survival was 1,593 days (IQR non-calculable). Overall, 390 (7.5%) patients were transferred to a CARC within 24 hours of arrest. No patient underwent multiple inter-facility transfers. Mean age was 65 (standard deviation (SD) 18) years and 2,046 (39%) were female (Table 1). Of all patients, 4,297 (82%) were treated at 42 non-CARCs with median annual volume of 17 (IQR 1 – 53 cases per center annually), 485 (9%) were treated at 3 lower-volume CARCs with average annual volume 41±30 cases per center and 435 (8%) were treated at a single high-volume CARC with an average of 109 cases per year.

Figure 2:

Overall patient cohort and exclusions

In unadjusted analysis of patient-level characteristics, transfer to a CARC was associated with a reduction in the hazard of death (unadjusted hazard ratio (HR) 0.72, 95% confidence interval (CI) 0.63 – 0.81), as were multiple patient-level factors (Supplemental Table 1). Treatment at a CARC was associated with a 27% reduction in the hazard of death compared to treatment at a non-CARC (Table 2). When we compared the three lower-volume CARCs to the single higher-volume CARC, there was no difference in survival (P = 0.49 for the unadjusted HR). Treatment at a center with a high volume of mechanical ventilation was associated with lower hazard of death (unadjusted HR 0.77, 95%CI 0.62 – 0.96) (Table 2). Overall OHCA case volume, hospital size and teaching status were not associated with outcome. The unadjusted MHR for hospital effect was 1.52 (IQR 1.23 – 1.97).

Table 2:

Hospital characteristics and unadjusted associations with patient-level outcome

Hospital-level characteristic	Treating hospitals (n = 46)	Unadjusted HR (95%CI)
CARC status (vs non-CARC)	4 (9)	0.73 (0.59 – 0.91)
OHCA case volume
Low volume (<25 cases/year)	28 (61)	Ref
Moderate volume (25 – 99 cases/year)	15 (34)	1.02 (0.83 – 1.25)
High volume (≥100 cases/year)	3 (7)	0.89 (0.68 – 1.17)
Total licensed beds*
Small (<157)	22 (50)	Ref
Moderate (157 – 409)	18 (41)	1.03 (0.85 – 1.26)
Large (≥410)	4 (9)	0.80 (0.61 – 1.04)
Total intensive care beds*
Small (<14)	14 (38)	Ref
Moderate (14–28)	15 (41)	1.03 (0.85 – 1.25)
Large (≥29)	8 (22)	0.81 (0.64 – 1.03)
Academic status*
Non-teaching hospital	27 (64)	Ref
Small teaching hospital	6 (14)	1.03 (0.78 – 1.35)
Large teaching hospital	9 (21)	0.88 (0.68 – 1.15)
High volume mechanical ventilation center*	6 (14)	0.77 (0.62 – 0.96)
High volume cardiac catheterization center*	19 (43)	0.97 (0.80 – 1.18)

^*Total number reported for several variables is less than 46 because CMS data were not available for all hospitals

Abbreviations: HR – Hazard ratio; OHCA – Out of hospital cardiac arrest; CARC – Cardiac arrest receiving center.

In adjusted models, treatment at a CARC was independently associated with a reduced hazard of death compared to treatment at a non-CARC (adjusted HR 0.84, 95%CI 0.74 – 0.94) (Table 3). These effects were unchanged when analysis was restricted only to non-transferred patients brought from the scene to the treating hospital (Table 3). When we compared the three lower-volume CARCs to the single higher-volume CARC, there was no difference in survival (P = 0.72 for the adjusted HR). Transfer status was independently associated with lower hazard of death (adjusted HR 0.31, 95%CI 0.20 – 0.48). There was a significant interaction between transfer status and time, with the effect of transfer driven by reduced mortality in the day 0 to 1 epoch. By contrast, transferred patients had increased hazard of death in each subsequent epoch (Table 3). Other than CARC status, no other hospital level predictors independently predicted outcome. After adjustment for patient-level characteristics, the MHR for hospital effect was 1.27 (IQR 1.12 – 1.49).

Table 3:

Multivariable survival models for both overall and limited to non-transferred patients transported directly from the OHCA scene to the treating hospital

Characteristic	Overall cohort (n = 5,217) Adjusted HR (95% CI)	Non-transfers only (n = 4,827) Adjusted HR (95% CI)
Transferred to CARC	0.31 (0.20 – 0.48)	--
Final treating hospital CARC	0.84 (0.74 – 0.94)	0.84 (0.75 – 0.94)
Time epoch
Day 0 to 1	Ref	Ref
Days 1 to 3	0.28 (0.24 – 0.32)	0.28 (0.24 – 0.32)
Days 3 to 7	0.10 (0.08 – 0.13)	0.10 (0.08 – 0.13)
Days 7+	0.001 (0.001 – 0.001)	0.001 (0.001 – 0.001)
Epoch-transfer interaction
Transfer × Day 1 to 3	4.42 (3.38 – 5.79)	--
Transfer × Day 3 to 7	9.13 (6.48 – 12.86)	--
Transfer × Day 7+	3.52 (1.73 – 7.18)	--
Age	1.01 (1.00 – 1.01)	1.01 (1.00 – 1.01)
Female sex	0.91 (0.88 – 0.95)	0.90 (0.87 – 0.94)
Non-traumatic arrest	0.68 (0.61 – 0.77)	0.67 (0.59 – 0.77)
Witnessed arrest
Unwitnessed	Ref	Ref
Witnessed	0.84 (0.77 – 0.91)	0.84 (0.77 – 0.91)
Missing	0.70 (0.60 – 0.82)	0.71 (0.60 – 0.83)
Dispatch to CPR interval
< 10 min	Ref	Ref
10–19 min	1.10 (0.89 – 1.34)	1.08 (0.88 – 1.30)
≥20 min	1.27 (1.06 – 1.51)	1.23 (1.05 – 1.45)
Missing	1.50 (1.29 – 1.74)	1.47 (1.28 – 1.70)
Epinephrine doses, mg
None documented	Ref	Ref
1 to 2	2.21 (1.93 – 2.54)	2.29 (1.99 – 2.63)
3 to 4	2.64 (2.28 – 3.08)	2.68 (2.28 – 3.14)
5 to 6	2.90 (2.47 – 3.41)	2.93 (2.47 – 3.47)
>6	2.95 (2.48 – 3.50)	2.91 (2.45 – 3.45)
Time from scene to receiving hospital
< 10 min	Ref	Ref
≥ 10 min	0.90 (0.84 – 0.96)	0.90 (0.84 – 0.96)
Missing	0.83 (0.69 – 1.00)	0.82 (0.69 – 0.97)
Documented past medical history
Congestive heart failure	1.07 (1.00 – 1.14)	1.06 (0.99 – 1.14)
Venous thromboembolism	1.19 (1.08 – 1.31)	1.19 (1.11– 1.28)
Malignancy	1.07 (0.99 – 1.15)	1.07 (0.99 – 1.15)
Hypertension	1.08 (1.03 – 1.13)	1.09 (1.04 – 1.14)
Neurological disease	1.06 (1.00 – 1.12)	1.05 (0.99 – 1.10)
Renal disease/dialysis	1.14 (1.07 – 1.21)	1.12 (1.05 – 1.19)
Pulmonary/respiratory disease	1.09 (1.04 – 1.14)	1.08 (1.03 – 1.13)
Other past medical history	1.12 (1.06 – 1.19)	1.12 (1.06 – 1.19)
Unknown past medical history	0.92 (0.85 – 1.00)	0.92 (0.85 – 0.99)

Abbreviations: HR – Hazard ratio; CARC – Cardiac arrest receiving center; CPR – Cardiopulmonary resuscitation.

Results did not change in sensitivity analyses comparing various time epoch splits. Sensitivity analyses exploring various breakpoints in our definition of overall SCA case volume and CARC categorization yielded similar estimates (data not shown). In our propensity-score adjusted analysis accounting for the possibility of ground EMS-level referral bias, our results also did not change. There was no meaningful change in our results when traumatic arrests were excluded, and no association with calendar quarter of treatment with outcome.

Discussion

The findings that early inter-facility transfer to a CARC or direct transport to a CARC from the scene are independently associated with improved long-term survival support calls to create regionalized systems for post-arrest care.^{6, 7} These results are consistent with improved short-term outcomes observed when EMS bypass of non-CARCs was instituted in some regions.^{15, 16} A regional system of many primary hospitals with a smaller number of CARCs parallels modern care for trauma and acute stroke, and may be advantageous in many regions where direct transport to CARCs by EMS is infeasible because of geographic or other constraints. Although survival bias and referral bias may contribute to the protective effect of inter-facility transfer, a CARC treatment effect persisted when analysis was restricted only to non-transferred patients, supporting the existence of a true outcome difference between centers. Given the prevalence of out-of-hospital SCA, the potential for improved outcomes through optimization of regionalized care is considerable.

Overall, the MHR for hospital effect revealed moderate unexplained variability in outcome. Treatment at one randomly selected hospital compared to treatment at another randomly selected hospital resulted in a 27% difference in the long-term hazard for mortality, even after accounting for measurable sources of variation. Despite this, no hospital level characteristic except for CARC status was an independent predictor of outcome. Although some prior studies have associated hospital characteristics with short-term survival after SCA, only the availability of cardiovascular interventions has been consistently associated with better outcomes.^{4, 8–12} Our work does not address what features of CARC care explain the increase in survival. Moreover, some arrest etiologies may require subspecialty care beyond general CARC capabilities. We do not, for example, propose that arrests due to major trauma be transported to a CARC that is not a Level 1 trauma center. Other populations may benefit from extracorporeal membrane oxygenation, toxicological specialty care, or other critical care interventions.

The temporal pattern of mortality suggests several potential mechanisms. CARC care is associated with a substantial mortality reduction in the first day after OHCA. Large multicenter analyses demonstrate most post-arrest deaths in the first day are due to withdrawal of life-sustaining therapy for perceived poor neurological prognosis.²⁵ Because neurological prognostication is imprecise early after cardiac arrest, early inter-facility transfer may prevent premature assumption of poor outcome and mortality. Delayed neurological prognostication may also explain the increase in late mortality associated with CARC care. A smaller but considerable proportion of early deaths are also due to multisystem organ failure or rearrest.²⁵ CARCs may provide robust systems of general critical care or coronary revascularization, thereby also reducing early mortality. Aggressive early care including targeted temperature, ventilator, and hemodynamic management may also reduce the risk of secondary neurological injury, further reducing mortality.²⁹

It is likely that organized systems of care rather than any single hospital-level characteristics result in optimal post-arrest care. This concept was recently formalized by the Institute of Medicine (IOM),⁶ which envisioned tiers of hospitals stratified by ability to care for acute coronary syndrome, hemodynamic instability, recurrent arrhythmias and post-anoxic coma. Clinically, these characteristics may also define patients likely to benefit from transfer to a CARC. The awake patient without anoxic brain injury, for example, likely derives little benefit from a center capable of providing neurocritical care, while the hemodynamically stable patient with a respiratory cause of SCA does not require advanced cardiac care capabilities. In our study, granular hospital-level data were not available for many included centers, although all CARCs regardless of volume met IOM’s description of a Level 1 center. Sufficient post-arrest patient level characteristics were not available to explore the differential benefit of transfer or CARC care in specific patient sub-populations. Data sets including this detailed patient-level information might reveal subgroups of patients most likely to benefit from transfer to a CARC.

Limitations

Our study has several important limitations. Because of the observational design of our work, we report the association between CARC and survival but cannot assert causality. However, we believe that our work represents the highest possible level of evidence short of a randomized trial, and is supportive of conducting future interventional research in this area. Because clinical data were only available from the prehospital record, some information, including post-arrest neurological assessment, were not available. This is one potential source of bias, although the direction of this effect is difficult to determine. It may be that the sickest patients are most likely to be transferred to specialized care from primary or secondary hospitals, biasing our results towards the null, or that these patients experience rapid limitation of care at these outlying facilities, increasing the estimated effect size. Also, in accordance with our a priori analysis plan, we did not adjust for multiple hypothesis testing. The P values for our main effect size are sufficiently small that they would remain statistically significant despite a conservative adjustment (e.g. Bonferroni correction), but the P values presented in the table do not account for the number of hypotheses tested.

We were unable to obtain data from several EMS agencies in the region that receive medical direction from outside the four regional health systems. The agencies we included represent the majority of services, including all of the area’s large services, and are responsible for >95% of EMS responses. Non-included services follow the same statewide protocols and represent a minority of transports, making it unlikely that their exclusion introduced systematic bias to our results. It may also be that some patients included in the analysis transiently regained pulses prior to hospital arrival without sustained return of spontaneous circulation. This may have artificially increased estimates of early post-arrest mortality, although 26% of patients in our study survived at least 30 days, consistent with prior outcome estimates in similar multicenter cohorts suggesting that this was not a major source of bias.⁸

The generalizability of our findings to other geographic regions is also uncertain. Record linkage to vital statistics, or even hospital outcome data, requires access to identifiable patient information, making it difficult to develop a national cohort. Regional data likely represent the highest level of available evidence short of an interventional trial. The overall distribution of demographic and arrest-specific characteristics in our cohort as well as their association with patient outcomes is like those described in other large OHCA cohorts.^{30, 31} Similarly, our region is not an outlier in terms of hospitals, beds or intensive care beds per capita (Supplemental Figure 1). However, whether our region’s four CARCs perform similarly to CARCs in other regions is unknown. That said, the CARCs in our study are hubs of the region’s major competitive health networks, improving the generalizability of our results. Finally, while we report associations with long-term survival after OHCA, quality of life and degree of functional recovery for these survivors is unknown. Survivors of critical illness in general,^32–34 and anoxic brain injury specifically,^{35, 36} may experience significant functional, cognitive and psychological limitations. In the United States, the National Death Index aggregates data from death certificates, which can be used to determine individuals’ vital status. Although some international studies have used governmental databases to estimate long-term functional status,³⁷ no such data source exists in America making this sort of linkage impossible. That said, any future prospective trial of post-arrest care regionalization must evaluate not only for vital status but also patient-centered outcomes including measures of quality of life and functional recovery.

Conclusion

In conclusion, linking regional data derived from 78 EMS services, 46 hospitals, and four health networks in a defined geographic region to vital statistics demonstrated a reduction in long-term mortality associated with both transfer to a CARC and initial CARC care among non-transferred patients. Further studies may explore whether the observed associations with reduced mortality lead to improved long-term functional outcomes and replicate our findings in other geographic regions.