Predicting return of spontaneous circulation in out-of-hospital cardiac arrest resuscitated in the emergency department: a single center retrospective study

Abstract

Background and aim

Out-of-hospital cardiac arrest (OHCA) is a major global health burden, with survival rates remaining critically low despite advancements in emergency care. Early identification of patients likely to achieve return of spontaneous circulation (ROSC) is vital for optimizing resuscitation strategies and improving clinical outcomes. This study aimed to develop and validate a multivariable predictive model for ROSC in patients with out-of-hospital cardiac arrest (OHCA) who received resuscitation in the emergency department, using readily available clinical variables.

Methods

We retrospectively analyzed clinical records of 902 OHCA patients who received resuscitative care in the emergency department of Hefei Second People’s Hospital between January 2021 and June 2024. Key variables including age, sex, number of defibrillations, total epinephrine dosage, dopamine (inotrope) administration, and CPR duration were extracted. Multivariable logistic regression was used to identify independent predictors of ROSC. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis (DCA).

Results

Among the 890 included patients, ROSC was achieved in 20.3%. Independent predictors of ROSC included number of defibrillations (OR: 1.40, 95% CI: 1.08–1.81), total epinephrine dose (OR: 0.59, 95% CI: 0.54–0.66), dopamine dose (OR: 1.01, 95% CI: 1.01–1.02), and CPR duration ≥ 30 min (OR: 0.18, 95% CI: 0.12–0.29). The model demonstrated good discriminative ability with an AUC of 0.833. Calibration and DCA supported the model’s clinical utility.

Conclusion

We developed a robust and interpretable predictive model for ROSC using real-world data from OHCA patients. While not intended for real-time clinical decision-making, the model can support retrospective evaluation and risk stratification, offering insights to improve resuscitation strategies and future research.

Clinical trial registration

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12873-025-01442-2.

Introduction

Out-of-hospital cardiac arrest (OHCA) remains a significant global public health challenge, with survival rates remaining critically low despite advancements in emergency medical services. In the United States alone, approximately 326,000 cases of OHCA are reported annually, with survival to hospital discharge rates estimated at around 10% [1]. The primary challenge lies not only in the immediate resuscitation efforts but also in accurately identifying patients who are likely to achieve return of spontaneous circulation (ROSC), an essential initial step toward survival, though not necessarily indicative of long-term favorable outcomes.

Timely recognition of patients with a higher probability of achieving ROSC can support real-time clinical decision-making, optimize resource allocation, and potentially improve survival rates. Several clinical factors have been explored in relation to ROSC, such as defibrillation, epinephrine use, inotropic or vasopressor support, and CPR duration. However, the predictive value of these factors varies across studies, and not all have been consistently identified as independent predictors [2–5]. However, the predictive value of these variables may differ across populations and healthcare settings due to variations in emergency response systems and patient characteristics.

Recent studies have underscored the need to develop predictive models tailored to specific populations. For example, Shojaie et al. highlighted the utility of machine learning approaches in uncovering non-clinical predictors of OHCA, such as sociodemographic and environmental factors, that may not be captured in conventional models [6]. Moreover, the 2024 update to the Utstein-style guidelines for OHCA registries emphasizes the importance of standardized data collection to enhance the accuracy and applicability of predictive models across clinical settings [7].

Against this background, the present study aims to identify key clinical predictors of ROSC and develop a multivariable logistic regression model to estimate ROSC probability in OHCA patients. Using real-world clinical data from 902 patients treated at the Emergency Department of Hefei Second People’s Hospital between January 2021 and June 2024, we sought to develop an interpretable model from routinely recorded variables that supports retrospective evaluation and early post-ROSC risk stratification, rather than real-time use during active CPR.

Patients and methods

Study design and setting

This retrospective cohort study was conducted at the Emergency Department of Hefei Second People’s Hospital, a tertiary care center in Anhui Province, China. We included patients who experienced out-of-hospital cardiac arrest (OHCA) and received cardiopulmonary resuscitation (CPR) either in the prehospital setting by emergency medical services (EMS) or upon arrival at the emergency department between January 1, 2021, and June 30, 2024. Patients for whom field termination of resuscitation (TOR) was declared and who were not transported to the hospital were not included in this dataset, as prehospital TOR is not routinely practiced in our regional EMS system. The study protocol was approved by the institutional ethics committee, and patient data were anonymized prior to analysis.

Study population

A total of 902 patients were screened. Inclusion criteria were: (1) age ≥ 18 years, (2) OHCA with documented CPR initiated in the prehospital or ED setting, and (3) availability of complete clinical data regarding resuscitation interventions and outcomes. We excluded patients who arrived at the hospital without any signs of life and those with incomplete data for key variables (defibrillation count, epinephrine dosage, dopamine administration, or CPR duration. Only patients who received active resuscitation efforts in the ED were included in the analysis.

Data collection and variables

Clinical and demographic data were extracted from the hospital’s electronic medical record system. The following variables were collected: age, sex, initial cardiac rhythm, number of defibrillation attempts, total epinephrine dosage (mg), dopamine dosage (mg), CPR duration (minutes), bystander CPR status, and the clinical outcome of return of spontaneous circulation (ROSC).

ROSC was defined as the sustained restoration of spontaneous circulation, with a palpable pulse and effective perfusion lasting at least 20 min, as documented in prehospital or ED records. For patients who achieved ROSC in the prehospital setting, medical records were reviewed to confirm that circulation was sustained for at least 20 min, either entirely in the field or continuously across the prehospital and ED phases. Patients with transient ROSC lasting less than 20 min were classified as ROSC-negative.

Statistical analysis

Data were analyzed using Python (v3.11) with packages including pandas, numpy, seaborn, and statsmodels. Continuous variables were summarized as median with interquartile range (IQR) and compared using the Mann–Whitney U test for non-normally distributed data (e.g., CPR duration and age), and as mean ± standard deviation (SD) with Student’s t-test for normally distributed variables. Categorical variables were summarized as counts and percentages and compared using the chi-square test or Fisher’s exact test, as appropriate.

Univariable logistic regression was first performed to identify potential predictors of ROSC. Variables with a significance level of P < 0.10 in univariable analysis were included in the multivariable logistic regression model to identify independent predictors. The final model included number of defibrillation attempts, number of epinephrine doses administered, dopamine usage (mg), and CPR duration ≥ 30 min (dichotomized as a binary variable).

While this approach is widely used in clinical research to streamline model development, we recognize that it may exclude variables with biological plausibility or clinical importance. Penalized regression techniques such as LASSO were considered during the design phase but were not implemented in the final model to prioritize interpretability and ease of clinical translation. Our goal was to develop an interpretable model from routinely recorded variables that supports retrospective evaluation and early post-ROSC risk stratification, rather than real-time use during active CPR.

CPR duration was initially examined as a continuous variable in univariable analysis, and visualized descriptively (see Figs. 2 and 3). However, for multivariable modeling, we dichotomized it using a 30-minute cutoff, which is a clinically relevant threshold associated with significantly reduced odds of favorable outcomes in prior literature [10, 11]. A sensitivity analysis using CPR duration as a continuous variable yielded similar findings, confirming the robustness of the observed association.

Fig. 2.

CPR Duration vs. Epinephrine Dose Stratified by ROSC Outcome. This scatter plot visualizes the relationship between CPR duration and epinephrine dose, colored by ROSC outcome. Patients who did not achieve ROSC (red) tended to receive higher epinephrine doses and had longer CPR durations compared to those who achieved ROSC (green). Axis intervals were optimized to improve interpretability, and visual transparency was applied to reduce overplotting

Fig. 3.

Distribution of CPR Duration by ROSC Outcome. This histogram compares the distribution of CPR duration in patients who achieved ROSC versus those who did not. A rightward shift in CPR duration among non-ROSC patients suggests that prolonged resuscitation is associated with lower ROSC rates

Model validation

To assess the generalizability of the model and minimize overfitting, we performed internal validation using 5-fold cross-validation. The dataset was randomly partitioned into five equal subsets, and the model was iteratively trained on four subsets and tested on the remaining one. The mean cross-validated area under the ROC curve (AUC) was 0.829 (95% CI: 0.796–0.858), demonstrating strong predictive performance. Calibration and decision curve analysis (DCA) were performed on the full dataset to assess model calibration and clinical utility. A two-tailed P-value < 0.05 was considered statistically significant.

Model development and evaluation

A multivariable logistic regression model was constructed to predict the likelihood of ROSC. Model performance was evaluated using the AUC to assess discrimination. Calibration of the model was evaluated graphically, and clinical usefulness was assessed using decision curve analysis (DCA). Statistical significance was defined as a two-tailed P-value < 0.05.

Results

Study population

A total of 902 patients with out-of-hospital cardiac arrest (OHCA) were evaluated. After excluding 12 cases with incomplete data on key variables, 890 patients were included in the final analysis. The median age of the cohort was 62 years (IQR: 52–74), and 59% were male. Return of spontaneous circulation (ROSC) was achieved in 181 patients, corresponding to a ROSC rate of 20.3%.

Baseline characteristics

Baseline characteristics stratified by ROSC status are summarized in Table 1. Patients who achieved ROSC had, on average, shorter CPR durations and lower epinephrine dosages compared to those who did not. The proportion of patients receiving defibrillation or inotropic support also differed between groups.

Table 1.

Baseline characteristics stratified by ROSC outcome

Variable	ROSC (n = 181)	No ROSC (n = 709)	P-value
Sex			0.120¹
Male	68 (60.7%)	356 (68.3%)
Female	44 (39.3%)	165 (31.7%)
Age, median (IQR)	63 (53–76)	61 (51–74)	0.637²
No-flow time, median (IQR)	28 (15–40)	30 (20–45)	0.030²
Bystander CPR			< 0.001¹
Yes	8 (7.1%)	5 (1.0%)
No	104 (92.9%)	516 (99.0%)
Comorbidities			0.048¹
Yes	37 (33.0%)	131 (25.1%)
No	75 (67.0%)	390 (74.9%)
Defibrillation Count			0.009¹
0	93 (83.0%)	481 (92.4%)
1	10 (8.9%)	20 (3.8%)
≥2	9 (8.1%)	20 (3.8%)
Number of Epinephrine Doses			< 0.001¹
≤2	35 (31.2%)	24 (4.6%)
>2–4	21 (18.8%)	28 (5.4%)
>4–6	50 (44.6%)	394 (75.6%)
>6–8	5 (4.5%)	44 (8.4%)
>8	1 (0.9%)	31 (6.0%)
Inotrope Use			< 0.001¹
Yes	36 (32.1%)	43 (8.3%)
No	76 (67.9%)	478 (91.7%)
CPR Duration (min)			< 0.001²
<30	34 (30.4%)	22 (4.2%)
≥30	78 (69.6%)	499 (95.8%)

¹ Chi-square test

² Mann–Whitney U test

*Note: Median and IQR used for non-normally distributed variables (age, CPR duration, and no-flow time)*

Univariable and multivariable analysis

In univariable logistic regression analysis, the following variables were significantly associated with ROSC: number of defibrillations, number of epinephrine doses administered, dopamine dose (as a proxy for inotropic support), and CPR duration ≥ 30 min. These variables were subsequently included in a multivariable logistic regression model.

Multivariable analysis identified four independent predictors of ROSC (Table 2):

Table 2.

Multivariable logistic regression for ROSC prediction

Variable	Odds Ratio (OR)	95% CI	P-value	Interpretation
Defibrillation Count	1.40	1.08–1.81	0.012	Higher count increases ROSC odds
Epinephrine Dose (mg)	0.59	0.54–0.69	< 0.001	Higher dose associated with lower ROSC odds
Dopamine Dose (mg)	1.01	1.01–1.02	< 0.001	Higher dose slightly increases ROSC odds
CPR Duration ≥ 30 min	0.18	0.12–0.29	< 0.001	Prolonged CPR strongly lowers ROSC odds

-Defibrillation count: OR = 1.40; 95% CI: 1.08–1.81; P = 0.012

-Epinephrine dose (mg): OR = 0.59; 95% CI: 0.54–0.69; P < 0.001

-Dopamine dose (mg): OR = 1.01; 95% CI: 1.01–1.02; P < 0.001

-CPR duration ≥30 min: OR = 0.18; 95% CI: 0.12–0.29; P < 0.001

Model performance

The logistic regression model demonstrated strong discriminatory performance, with an area under the receiver operating characteristic (ROC) curve of 0.833 (Fig. 1). Calibration and decision curve analyses supported the model’s accuracy and clinical applicability. Calibration and decision curve analyses (Supplementary Figs. 1 and 2) supported the model’s accuracy and clinical utility. Internal validation using 5-fold cross-validation yielded a mean AUC of 0.829 (95% CI: 0.796–0.858), indicating robust performance and good generalizability across data partitions.

Fig. 1.

Cross-Validated Receiver Operating Characteristic (ROC) Curve. The ROC curve illustrates the discriminative performance of the logistic regression model in predicting ROSC among OHCA patients. The model achieved a mean cross-validated AUC of 0.929 (± 0.018), with fold-wise curves shown. This performance supports generalizability and robust classification accuracy

Exploratory visualization

Additional exploratory analyses were performed to visualize the relationships between CPR duration, epinephrine dosage, and ROSC outcome. A scatter plot illustrating the distribution of CPR duration and epinephrine dose, stratified by ROSC status, is presented in Fig. 2. Patients who failed to achieve ROSC tended to receive higher doses of epinephrine and undergo longer durations of CPR. A histogram of CPR duration distribution by ROSC status is shown in Fig. 3, highlighting a rightward shift in CPR time among non-ROSC patients.

Discussion

In this retrospective cohort study, we developed and validated a multivariable logistic regression model to predict the ROSC in patients with OHCA. Utilizing data from 890 patients treated at a tertiary care hospital, we identified four independent predictors of ROSC: number of defibrillation attempts, total epinephrine dosage, dopamine dosage, and prolonged CPR duration (≥ 30 min). The model demonstrated excellent discriminative performance with an AUC of 0.833, and its clinical utility was supported by calibration and DCA.

Our results reinforce well-established resuscitation principles. Early and frequent defibrillation was positively associated with ROSC, consistent with prior studies and current resuscitation guidelines emphasizing early rhythm recognition and shock delivery in cases of ventricular fibrillation or pulseless ventricular tachycardia [8, 9]. Conversely, longer CPR durations were significantly associated with lower odds of ROSC, likely reflecting delayed circulation restoration and prolonged myocardial ischemia [10, 11]. Although CPR duration was dichotomized using a 30-minute cutoff for model simplicity and clinical relevance, we acknowledge that this approach may limit resolution in capturing non-linear trends. We conducted sensitivity analyses treating CPR duration as a continuous variable, which confirmed the inverse association with ROSC. The 30-minute threshold is widely referenced in cardiac arrest literature and reflects a pragmatic clinical decision point for ongoing resuscitative efforts [9–11]. This association is further supported by the inverse relationship between the number of epinephrine doses administered and ROSC, likely reflecting longer resuscitation times and more severe underlying clinical scenarios.

The inverse association between cumulative epinephrine administration and ROSC adds to a growing body of literature questioning the optimal timing and dosage of vasopressors during cardiac arrest. While epinephrine has demonstrated short-term benefits in achieving ROSC, recent trials have raised concerns about its impact on neurologic outcomes and long-term survival [12]. This supports the need for careful titration and real-time clinical judgment regarding vasopressor use during prolonged resuscitation.

We also observed a positive association between dopamine dosage and ROSC. While dopamine may enhance myocardial perfusion post-ROSC, its use likely reflects more aggressive resuscitative efforts overall, making it difficult to disentangle causality from clinical context. Although epinephrine administration is a cornerstone of cardiac arrest protocols, our findings—along with those from large trials such as PARAMEDIC2—suggest that while epinephrine may increase ROSC rates, it does not necessarily improve long-term survival or neurological outcomes [13].

Our model offers several advantages: it is based on routinely available clinical variables and employs an interpretable logistic regression approach. It is suited for retrospective evaluation and early post-ROSC risk stratification once key data are available, rather than for real-time decisions during CPR. Unlike machine learning algorithms, which may be perceived as “black boxes,” our model maintains transparency while still achieving high predictive performance [6, 14]. Additionally, DCA provided a nuanced assessment of clinical net benefit, reinforcing the model’s value for guiding decisions under uncertainty [15, 16].

Despite its strengths, this study has limitations. First, the retrospective design introduces potential biases related to selection and data quality. Second, this was a single-center study, which may limit generalizability to different patient populations or EMS systems. The median age of our cohort was 62 years, which is relatively younger compared to some other OHCA populations. This may reflect regional demographics or differences in EMS access and response time. Additionally, the proportion of patients who received bystander CPR was notably low. This could be attributed to limited public awareness, lack of formal CPR training programs, or low availability of bystanders in witnessed arrests, and highlights an area for future public health intervention. Third, we did not assess long-term or neurologically intact survival, which are essential metrics in cardiac arrest research. Finally, we acknowledge that our variable selection method, based on univariable screening, may have excluded some clinically important predictors. While more advanced techniques like LASSO regression offer advantages in variable selection, we opted for a conventional logistic regression approach to preserve transparency and enhance clinical usability. While internal validation via 5-fold cross-validation supports the robustness of our findings, external validation in independent cohorts is needed to confirm the model’s generalizability across diverse healthcare settings.

Conclusion

We developed an interpretable model for ROSC using routinely recorded variables. Defibrillation frequency, epinephrine dose count, dopamine use, and CPR duration ≥ 30 min were independently associated with ROSC. The model demonstrated good discrimination and calibration. It is intended for retrospective evaluation and early post-ROSC risk stratification, not for real-time decisions during CPR. External validation is needed before wider clinical application.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

OHCA

Out–of–Hospital Cardiac Arrest

ROSC

Return of Spontaneous Circulation

CPR

Cardiopulmonary Resuscitation

AUC

Area Under the Curve

DCA

Decision Curve Analysis

OR

Odds Ratio

CI

Confidence Interval

IQR

Interquartile Range

SD

Standard Deviation

EMS

Emergency Medical Services

Author contributions

ZS contributed to the literature review, patient recruitment, and data collection. GW conducted the statistical analysis, interpreted the results, and drafted the manuscript. ZS, GW participated in clinical evaluation, data verification, and manuscript editing. ZS assisted in data collection and resuscitation protocol review. LH designed and supervised the study, developed the intervention protocol, and critically revised the manuscript for important intellectual content. All authors have read and approved the final version of the manuscript.

Funding

This study was supported by the Hefei Municipal Health and Medical Application Research Project (Grant No. Hwk2021yb011).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request and subject to ethical and privacy restrictions.

Declarations

Ethics approval and consent to participate

The study was approved by the Ethics Committee of Hefei Second People’s Hospital. Written informed consent was obtained from all participants or their legal representatives prior to inclusion, in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.