Abstract
Background
Post-cardiac arrest syndrome carries substantial mortality despite advances in resuscitation. We developed a machine learning model to predict 28-day mortality using comprehensive clinical data.
Methods
We analyzed data from 1122 cardiac arrest patients in the MIMIC-IV database. After applying exclusion criteria, 853 patients with complete 28-day outcome data were included. We extracted 99 variables across six domains and compared five machine learning algorithms. Lactate clearance was calculated as: (Day 1 lactate − Day 3 lactate)/Day 1 lactate × 100%. Model performance was evaluated using AUC-ROC, calibration metrics, and SHAP analysis.
Results
Among 853 patients (mean age 64.7 ± 16.3 years, 63.5% male), 327 (38.3%) died within 28 days. The XGBoost model achieved an AUC-ROC of 0.89 (95% CI: 0.86–0.92), outperforming APACHE III (AUC: 0.73). Lactate clearance rate emerged as the primary predictor (SHAP value: 0.24), followed by GCS score (0.21), SOFA score (0.18), age (0.16), and treatment intensity (0.14). Poor lactate clearance (0–25%) was associated with 67% mortality compared to 39% in moderate clearance (25–50%).
Conclusions
Our machine learning model demonstrated superior accuracy for 28-day mortality prediction in post-cardiac arrest patients. Dynamic lactate clearance and neurological assessment provide actionable clinical insights for risk stratification.
Keywords: Lactate clearance, Machine learning, Post-cardiac arrest syndrome, SOFA score, Mortality prediction
Introduction
Every year, approximately 350,000 people in the United States experience cardiac arrest. The survival rate of patients who experience this medical emergency outside of hospitals is approximately 10–15 percent, increasing to 15–25 percent when cardiac arrest occurs in a hospital setting.1, 2 Post-cardiac arrest syndrome has a complex pathophysiology, including brain injury, myocardial dysfunction, and ischemia–reperfusion responses, making outcome prediction challenging.3
Current prognostic tools rely on static scoring systems, such as SOFA and APACHE, which demonstrate modest performance with AUCs of 0.70–0.80.4, 5 These instruments have limitations including inability to capture dynamic physiological changes, failure to incorporate complex interactions between organ systems, and limited temporal resolution for clinical decision-making.
Machine learning algorithms can manage complex datasets effectively, identify nonlinear relationships, and incorporate temporal factors frequently overlooked by traditional methods. Despite promising results in sepsis detection and mortality prediction, their application in post-cardiac arrest syndrome remains limited.6, 7
We hypothesized that machine learning algorithms combining extensive clinical data would exhibit greater predictive precision than conventional scoring systems, with dynamic biomarkers such as lactate clearance serving as pivotal predictors.
Methods
Study design and ethics
This retrospective cohort study utilized MIMIC-IV version 2.2, containing deidentified clinical data from ICU patients at Beth Israel Deaconess Medical Center (2008–2019). The Institutional Review Board waived informed consent requirements based on retrospective data analysis with deidentified participants. The study adhered to STROBE guidelines for observational studies (Supplementary Table S1).
Study setting
Beth Israel Deaconess Medical Center is a major academic tertiary care hospital affiliated with Harvard Medical School, located in Boston, Massachusetts. The hospital serves a diverse urban population and maintains comprehensive ICU services including medical, surgical, and cardiac intensive care units. ICU admission decisions follow institutional protocols based on severity of illness and need for organ support. Neurological assessment using the Glasgow Coma Scale is routinely performed and documented in the electronic health record system.
Patient selection
We identified patients with cardiac arrest using ICD-10 codes (I46.x, Z87.74) from the MIMIC-IV database. The database includes both in-hospital cardiac arrest (IHCA) and out-of-hospital cardiac arrest (OHCA) patients who were subsequently admitted to the ICU. Inclusion criteria were: age ≥18 years at ICU admission, primary or secondary cardiac arrest diagnosis, ICU length of stay ≥24 h, and available key demographic and clinical variables. Exclusion criteria included duplicate ICU admissions (first admission retained), missing primary outcome data, ICU stay <24 h, and missing >50% of core predictive variables. Of 1122 initially identified patients, 853 met all criteria and were included in the final analysis (Fig. 1). In the MIMIC-IV database, ICD-10 codes do not reliably differentiate out-of-hospital cardiac arrest (OHCA) from in-hospital cardiac arrest (IHCA). Only procedure-related cardiac arrest codes (e.g., I97.120, I97.121, I97.710, I97.711) can be confidently classified as in-hospital events; 43 patients (5.0%) met these criteria. Therefore, the study cohort represents a mixed cardiac arrest population, and arrest location was not used as a primary modeling variable.
Fig. 1.
Patient selection flowchart showing the derivation of the study cohort from the MIMIC-IV database. Of 1122 patients with cardiac arrest-related diagnoses, 853 patients with complete 28-day outcome data were included in the final analysis (survivors: n = 526, 61.7%; non-survivors: n = 327, 38.3%). Exclusion criteria included missing 28-day outcome data (n = 185), ICU stay < 24 h (n = 54), and missing key variables (n = 30).
Variable extraction
We extracted 99 clinical variables from six domains. Demographics included age, gender, race, admission source, insurance status, weight, and primary diagnosis. Comorbidities encompassed hypertension, diabetes, coronary artery disease, chronic kidney disease, stroke history, chronic obstructive pulmonary disease, and active malignancy. Treatment interventions included extracorporeal membrane oxygenation (ECMO), continuous renal replacement therapy (CRRT), intra-aortic balloon pump, mechanical ventilation, percutaneous coronary intervention, targeted temperature management, PICC placement, and bilateral pneumonia treatment. Pharmacological interventions consisted of vasoactive medications, antiarrhythmics, corticosteroids, and sodium bicarbonate. Laboratory parameters comprised 63 variables with serial measurements over the first three days. Clinical severity scores included GCS (assessed within the first 24 h of ICU admission, excluding periods of deep sedation when possible), SOFA, Charlson Comorbidity Index, APACHE III, LODS, OASIS, and SIRS criteria.
Lactate clearance was calculated using the formula: (Day 1 lactate − Day 3 lactate)/Day 1 lactate × 100%. This dynamic measure captures metabolic recovery trajectory and tissue perfusion adequacy over the initial 72 h of ICU care. Excellent lactate clearance was defined as >75% reduction, good as 50–75%, moderate as 25–50%, poor as 0–25%, and worsening as ≤0% (indicating lactate increase). The primary outcome was 28-day all-cause mortality from ICU admission.
Statistical analysis
Continuous variables are expressed as mean ± standard deviation. Categorical variables are presented as frequencies and percentages. Univariate comparisons used chi-square tests for categorical variables and t-tests for continuous variables. For machine learning modeling, data preprocessing included K-nearest neighbors imputation (k = 5), z-score normalization, and categorical encoding. We used an 80/20 train-test split with stratified sampling and compared five algorithms: Logistic Regression with L2 regularization, Random Forest with 100 trees, XGBoost, LightGBM, and Support Vector Machine with RBF kernel. Hyperparameter optimization used 5-fold cross-validation with grid search. Model performance was evaluated using AUC-ROC, AUC-PR, and calibration plots. Feature importance was assessed using SHAP (SHapley Additive exPlanations) values.8 Statistical significance was set at P < 0.05. Analyses were performed using Python 3.9 and R 4.2.0.
Results
Patient characteristics
From 1122 patients with cardiac arrest-related diagnoses in the MIMIC-IV database, 853 patients with complete 28-day outcome data were included in the final analysis (Fig. 1). The study population had a mean age of 64.7 ± 16.3 years, with 63.5% male and 55.7% White. Within the 28-day follow-up period, 327 deaths occurred, yielding a mortality rate of 38.3%.
Non-survivors were older (68.2 ± 15.9 versus 62.6 ± 16.3 years, p < 0.001), had lower GCS scores (5.23 ± 1.58 versus 11.71 ± 1.93, p < 0.001), higher SOFA scores (8.61 ± 2.62 versus 5.91 ± 2.99, p < 0.001), higher APACHE III scores (85.0 ± 27.9 versus 59.4 ± 29.3, p < 0.001), and elevated initial lactate concentrations (4.74 ± 3.54 versus 3.32 ± 2.62 mmol/L, p < 0.001) compared to survivors (Table 1).
Table 1.
Baseline characteristics stratified by 28-day survival status.
| Variable | Overall (n = 853) | Survivors (n = 526) | Non-survivors (n = 327) | P value |
|---|---|---|---|---|
| Demographics | ||||
| Age, years | 64.7 ± 16.3 | 62.6 ± 16.3 | 68.2 ± 15.9 | <0.001 |
| Male, n (%) | 542 (63.5) | 349 (66.3) | 193 (59.0) | 0.04 |
| White race, n (%) | 475 (55.7) | 303 (57.6) | 172 (52.6) | 0.17 |
| Procedure-related IHCA, n (%) | 43 (5.0) | 31 (5.9) | 12 (3.7) | 0.15 |
| Clinical scores | ||||
| GCS score | 9.22 ± 3.63 | 11.71 ± 1.93 | 5.23 ± 1.58 | <0.001 |
| SOFA score | 6.94 ± 3.14 | 5.91 ± 2.99 | 8.61 ± 2.62 | <0.001 |
| APACHE III score | 69.2 ± 31.4 | 59.4 ± 29.3 | 85.0 ± 27.9 | <0.001 |
| Laboratory values (Day 1) | ||||
| Lactate, mmol/L | 3.87 ± 3.09 | 3.32 ± 2.62 | 4.74 ± 3.54 | <0.001 |
| Creatinine, mg/dL | 1.62 ± 1.40 | 1.48 ± 1.26 | 1.85 ± 1.57 | <0.001 |
| Treatment interventions | ||||
| Mechanical ventilation, n (%) | 706 (82.8) | 416 (79.1) | 290 (88.7) | <0.001 |
| CRRT, n (%) | 82 (9.6) | 32 (6.1) | 50 (15.3) | <0.001 |
| ECMO, n (%) | 5 (0.6) | 3 (0.6) | 2 (0.6) | 1.00 |
| Vasoactive drugs, n (%) | 226 (26.5) | 146 (27.8) | 80 (24.5) | 0.33 |
Data are presented as mean ± SD or n (%). P values were calculated using independent t-test for continuous variables and chi-square test for categorical variables.
Abbreviations: GCS, Glasgow Coma Scale; SOFA, Sequential Organ Failure Assessment; APACHE, Acute Physiology and Chronic Health Evaluation; CRRT, continuous renal replacement therapy; ECMO, extracorporeal membrane oxygenation.
Mechanical ventilation was used in 82.8% of patients, continuous renal replacement therapy in 9.6%, ECMO in 0.6% (5 patients), and vasoactive medications in 26.5%. Mortality rates were significantly higher among patients receiving mechanical ventilation (88.7% versus 79.1% in survivors, p < 0.001) and CRRT (15.3% versus 6.1%, p < 0.001).
Lactate clearance patterns
Among the 312 patients with complete lactate data for clearance calculation, patterns were significantly associated with outcomes9, 10 (Fig. 4). Patients with moderate clearance (25–50%) had the lowest mortality at 39% (n = 54), while those with poor clearance (0–25%) demonstrated the highest mortality at 67% (n = 33). Patients with worsening lactate levels (≤0%), indicating ongoing metabolic deterioration, had 53% mortality (n = 59). Excellent clearance (>75%) was associated with 42% mortality (n = 57), and good clearance (50–75%) with 47% mortality (n = 109). This demonstrates a clear relationship between lactate clearance patterns and survival outcomes.
Fig. 4.
Kaplan-Meier survival curves stratified by lactate clearance rate groups over the 28-day follow-up period (n = 312 patients with complete lactate data). Patients were categorized based on lactate clearance rate calculated from Day 1 to Day 3: Excellent (>75%, n = 57, 42% mortality), Good (50–75%, n = 109, 47% mortality), Moderate (25–50%, n = 54, 39% mortality), Poor (0–25%, n = 33, 67% mortality), and Worsening (≤0%, n = 59, 53% mortality). Poor lactate clearance was associated with significantly higher mortality risk (p < 0.001 by log-rank test).
Machine learning performance
The XGBoost model achieved the highest performance with an AUC-ROC of 0.89 (95% CI: 0.86–0.92), followed by LightGBM (0.87), Random Forest (0.85), Logistic Regression (0.80), and Support Vector Machine (0.78). The traditional APACHE III scoring system achieved an AUC of 0.73 (Fig. 2, Table 2). The superior performance of gradient boosting methods (XGBoost, LightGBM) can be attributed to their ability to capture complex nonlinear interactions between clinical variables and their iterative approach to building decision trees that focus on previously misclassified cases.
Fig. 2.
Receiver operating characteristic (ROC) curves comparing the performance of different machine learning algorithms and traditional APACHE III scoring system for predicting 28-day mortality. The XGBoost model achieved the highest area under the curve (AUC = 0.89), followed by LightGBM (AUC = 0.87), Random Forest (AUC = 0.85), and APACHE III (AUC = 0.73). The diagonal dashed line represents random performance (AUC = 0.50).
Table 2.
Performance comparison of machine learning models for 28-day mortality prediction.
| Algorithm | AUC-ROC (95% CI) | AUC-PR (95% CI) | Sens | Spec | F1 |
|---|---|---|---|---|---|
| XGBoost | 0.89 (0.86–0.92) | 0.84 (0.80–0.88) | 0.82 | 0.84 | 0.77 |
| LightGBM | 0.87 (0.84–0.90) | 0.81 (0.77–0.85) | 0.79 | 0.82 | 0.74 |
| Random Forest | 0.85 (0.82–0.88) | 0.78 (0.74–0.82) | 0.76 | 0.80 | 0.71 |
| Logistic Regression | 0.80 (0.77–0.83) | 0.72 (0.68–0.76) | 0.72 | 0.76 | 0.67 |
| Support Vector Machine | 0.78 (0.75–0.81) | 0.69 (0.65–0.73) | 0.70 | 0.74 | 0.64 |
| APACHE III (Traditional) | 0.73 (0.69–0.77) | 0.62 (0.58–0.66) | 0.65 | 0.70 | 0.58 |
Bold indicates the best-performing model. Abbreviations: AUC-ROC, area under the receiver operating characteristic curve; AUC-PR, area under the precision-recall curve; Sens, sensitivity; Spec, specificity; CI, confidence interval.
The XGBoost model demonstrated sensitivity of 0.82 (95% CI: 0.78–0.86), specificity of 0.84 (95% CI: 0.81–0.87), positive predictive value of 0.72 (95% CI: 0.68–0.76), negative predictive value of 0.90 (95% CI: 0.87–0.93), and F1-Score of 0.77. Calibration was excellent with Hosmer-Lemeshow p = 0.31 (Fig. 5).
Fig. 5.
Calibration plot showing the agreement between predicted probabilities and observed frequencies of 28-day mortality for the XGBoost model. The model demonstrates excellent calibration (Hosmer-Lemeshow p = 0.31), with predicted probabilities closely following the ideal calibration line (diagonal). Error bars represent 95% confidence intervals for each decile of predicted risk.
Feature importance
SHAP analysis identified lactate clearance rate as the most influential predictor (SHAP value: 0.24), where poor clearance substantially increased mortality risk (Fig. 3). GCS score ranked second (0.21), with lower neurological function scores consistently predicting worse outcomes. SOFA score followed (0.18), with higher scores predicting increased mortality. Age contributed significantly (0.16), with risk increasing progressively above 70 years. Treatment intensity (0.14), a composite score reflecting life support interventions, and first-day creatinine (0.12) completed the top predictors. Dynamic features contributed 31% of total predictive power, static features 42%, and treatment variables 27%.
Fig. 3.
Feature importance ranking based on SHAP (SHapley Additive exPlanations) values from the XGBoost model. The top 15 most influential predictive features are shown. Lactate clearance rate was the most important predictor (SHAP value: 0.24), followed by GCS score (0.21), SOFA score (0.18), age (0.16), treatment intensity (0.14), and first-day creatinine (0.12). Higher SHAP values indicate greater contribution to mortality prediction.
Model validation
Internal validation demonstrated robust performance with minimal overfitting (training AUC: 0.91; validation AUC: 0.89). Decision curve analysis showed superior net benefit compared to traditional scoring across clinically relevant thresholds (0.3–0.8). The model correctly reclassified 23% of patients compared to APACHE III, with a net reclassification improvement of 0.18 (95% CI: 0.12–0.24, p < 0.001).
Discussion
We developed a machine learning model for 28-day mortality prediction in post-cardiac arrest patients that significantly outperformed traditional clinical scoring systems. The XGBoost model achieved an AUC-ROC of 0.89 with excellent calibration. Lactate clearance emerged as the strongest predictor, followed by GCS score, highlighting the importance of both metabolic recovery markers and neurological assessment for prognostication.
Our findings align with and extend prior research. Traditional scoring systems typically achieve AUC 0.70–0.75 in cardiac arrest populations,11, 12 consistent with our APACHE III performance (AUC: 0.73). Recent machine learning studies in general ICU populations report AUC of 0.80–0.87,13, 14 whereas our cardiac arrest-specific model achieved an AUC of 0.89, suggesting that disease-specific models offer superior performance.
The gradient boosting algorithms (XGBoost, LightGBM) outperformed other approaches due to their capacity for handling heterogeneous data types, automatic feature interaction detection, and robustness to missing values. Random Forest, while effective, showed slightly lower performance potentially due to its inability to optimize across trees sequentially. Support Vector Machines, despite their theoretical advantages in high-dimensional spaces, were limited by sensitivity to feature scaling and difficulty in capturing complex variable interactions inherent in clinical data.
Lactate clearance as the top predictor supports extensive literature on lactate kinetics in critical illness.15, 16 Our finding that poor clearance (0–25%) is associated with 67% mortality compared to 39% in moderate clearance (25–50%) provides clinically actionable thresholds. The prominence of GCS score as the second most important predictor underscores the critical role of neurological assessment in post-cardiac arrest prognosis. The significant difference in GCS scores between survivors (11.71 ± 1.93) and non-survivors (5.23 ± 1.58) demonstrates the prognostic value of early neurological evaluation.17
The high negative predictive value (0.90) supports use of this model as a screening tool for identifying low-risk patients who may benefit from less-intensive monitoring. Lactate clearance provides actionable targets for intervention, while GCS assessment guides neurological prognostication strategies. The interpretability provided by SHAP analysis addresses critical concerns about clinical AI adoption, enabling clinicians to understand prediction drivers and facilitate integration into decision-making workflows.18, 19 The low ECMO mortality (40%) in our cohort compared to historical series (50–70%) may reflect evolving selection criteria and improved protocols, although the small sample size (n = 5) limits conclusions.20
Limitations
This study has several limitations. As a retrospective analysis from a single academic center, generalizability may be limited. The MIMIC database represents a tertiary academic institution in the United States, and external validation using independent datasets from different healthcare settings is necessary before clinical implementation. Unobserved confounders including detailed neurological assessments, imaging findings, and care limitation preferences may influence outcomes but were not captured. Functional neurological status, essential for cardiac arrest survivors, was not incorporated. Model drift from changing clinical practices remains a concern for prolonged implementation. Additionally, the MIMIC-IV database does not allow reliable discrimination between OHCA and IHCA based on ICD coding alone. Although procedure-related cardiac arrests (5.0%) can be identified as definite IHCA, the majority of cases could not be definitively classified. As OHCA and IHCA differ in pathophysiology and prognosis, this heterogeneity may affect generalizability and warrants validation in datasets with explicit arrest location information.
Conclusion
Our machine learning model significantly outperformed traditional scoring systems for mortality prediction in post-cardiac arrest patients. Dynamic lactate clearance was identified as the most significant prognostic factor (AUC-ROC: 0.89), followed by GCS score, SOFA score, age, and treatment intensity. The clear stratification of mortality risk across lactate clearance categories (39–67% mortality range) demonstrates the clinical utility of this dynamic biomarker. While external validation and prospective evaluation are needed, our findings demonstrate the potential of artificial intelligence to enhance prognostic accuracy and clinical decision-making in post-cardiac arrest syndrome.
Declaration of Generative AI and AI-assisted technologies in the writing process
During the preparation of this work, the authors used Deepseek for language editing and manuscript formatting. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.
CRediT authorship contribution statement
Danxia Chen: Writing – original draft, Investigation, Formal analysis, Data curation. Guode Li: Writing – original draft, Formal analysis, Data curation. Junlin Huang: Resources, Investigation, Data curation. Weifeng Yuan: Supervision, Resources, Software. Qinqin Shu: Data curation, Formal analysis. Yan Li: Writing – review & editing, Writing – original draft, Supervision, Project administration, Formal analysis, Data curation, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Footnotes
Supplementary material to this article can be found online at https://doi.org/10.1016/j.resplu.2026.101223.
Contributor Information
Weifeng Yuan, Email: walfred1986@sina.com.
Yan Li, Email: criticalcare@163.com.
Appendix A. Supplementary material
The following are the Supplementary material to this article:
References
- 1.Yan S., Gan Y., Jiang N., et al. The global survival rate among adult out-of-hospital cardiac arrest patients who received cardiopulmonary resuscitation: a systematic review and meta-analysis. Crit Care. 2020;24(1):61. doi: 10.1186/s13054-020-2773-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Girotra S., Nallamothu B.K., Spertus J.A., et al. Trends in survival after in-hospital cardiac arrest. N Engl J Med. 2012;367(20):1912–1920. doi: 10.1056/NEJMoa1109148. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Nolan J.P., Neumar R.W., Adrie C., et al. Post-cardiac arrest syndrome: epidemiology, pathophysiology, treatment, and prognostication. Resuscitation. 2008;79(3):350–379. doi: 10.1016/j.resuscitation.2008.09.017. [DOI] [PubMed] [Google Scholar]
- 4.Haywood K., Whitehead L., Nadkarni V.M., et al. COSCA (Core Outcome Set for Cardiac Arrest) in adults: an advisory statement from the International Liaison Committee on Resuscitation. Circulation. 2018;137(22):e783–e801. doi: 10.1161/CIR.0000000000000562. [DOI] [PubMed] [Google Scholar]
- 5.Geocadin R.G., Callaway C.W., Fink E.L., et al. Standards for studies of neurological prognostication in comatose survivors of cardiac arrest: a scientific statement from the American Heart Association. Circulation. 2019;140(9):e517–e542. doi: 10.1161/CIR.0000000000000702. [DOI] [PubMed] [Google Scholar]
- 6.Hyland S.L., Faltys M., Hüser M., et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med. 2020;26(3):364–373. doi: 10.1038/s41591-020-0789-4. [DOI] [PubMed] [Google Scholar]
- 7.Meyer A., Zverinski D., Pfahringer B., et al. Machine learning for real-time prediction of complications in critical care: a retrospective study. Lancet Respir Med. 2018;6(12):905–914. doi: 10.1016/S2213-2600(18)30300-X. [DOI] [PubMed] [Google Scholar]
- 8.Lundberg S.M., Lee S.I. A unified approach to interpreting model predictions. Adv Neural Inf Proces Syst. 2017;30:4765–4774. [Google Scholar]
- 9.Donnino M.W., Andersen L.W., Giberson T., et al. Initial lactate and lactate change in post-cardiac arrest: a multicenter validation study. Crit Care Med. 2014;42(8):1804–1811. doi: 10.1097/CCM.0000000000000332. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Hayashida K., Suzuki M., Yonemoto N., et al. Early lactate clearance is associated with improved outcomes in patients with postcardiac arrest syndrome: a prospective, multicenter observational study. Crit Care Med. 2017;45(6):e559–e566. doi: 10.1097/CCM.0000000000002307. [DOI] [PubMed] [Google Scholar]
- 11.Richardson D.K., Zive D.M., Daya M., et al. The impact of early do not resuscitate (DNR) orders on patient care and outcomes following resuscitation from out of hospital cardiac arrest. Resuscitation. 2013;84(4):483–487. doi: 10.1016/j.resuscitation.2012.08.327. [DOI] [PubMed] [Google Scholar]
- 12.Beylin M.E., Perman S.M., Abella B.S., et al. Higher mean arterial pressure with or without vasoactive agents is associated with increased survival and better neurological outcomes in comatose survivors of cardiac arrest. Intensive Care Med. 2013;39(11):1981–1988. doi: 10.1007/s00134-013-3075-9. [DOI] [PubMed] [Google Scholar]
- 13.Johnson A.E.W., Bulgarelli L., Shen L., et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1. doi: 10.1038/s41597-022-01899-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.von Elm E., Altman D.G., Egger M., et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370(9596):1453–1457. doi: 10.1016/S0140-6736(07)61602-X. [DOI] [PubMed] [Google Scholar]
- 15.Vincent J.L., Silva Q.E., A, Couto L Jr, Taccone FS. The value of blood lactate kinetics in critically ill patients: a systematic review. Crit Care. 2016;20(1):257. doi: 10.1186/s13054-016-1403-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Bakker J., Nijsten M.W., Jansen T.C. Clinical use of lactate monitoring in critically ill patients. Ann Intensive Care. 2013;3(1):12. doi: 10.1186/2110-5820-3-12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Ajam K., Gold L.S., Beck S.S., et al. Reliability of the Cerebral Performance Category to classify neurological status among survivors of ventricular fibrillation arrest: a cohort study. Scand J Trauma Resusc Emerg Med. 2011;19:38. doi: 10.1186/1757-7241-19-38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Adrie C., Cariou A., Mourvillier B., et al. Predicting survival with good neurological recovery at hospital admission after successful resuscitation of out-of-hospital cardiac arrest: the OHCA score. Eur Heart J. 2006;27(23):2840–2845. doi: 10.1093/eurheartj/ehl335. [DOI] [PubMed] [Google Scholar]
- 19.Maupain C., Bougouin W., Lamhaut L., et al. The CAHP (Cardiac Arrest Hospital Prognosis) score: a tool for risk stratification after out-of-hospital cardiac arrest. Eur Heart J. 2016;37(42):3222–3228. doi: 10.1093/eurheartj/ehv556. [DOI] [PubMed] [Google Scholar]
- 20.Stub D., Bernard S., Pellegrino V., et al. Refractory cardiac arrest treated with mechanical CPR, hypothermia, ECMO and early reperfusion (the CHEER trial) Resuscitation. 2015;86:88–94. doi: 10.1016/j.resuscitation.2014.09.010. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.