Cardiac arrest (CA) remains a critical challenge in contemporary medicine, characterized by a dismally low survival rate and poor prognosis, and, therefore, has garnered global attention [1]. CA can be classified by the occurrence location into in-hospital CA (IHCA) and out-of-hospital CA (OHCA). Despite advancements in cardiopulmonary resuscitation techniques, global registry data indicate that the incidence and survival rates of CA have not significantly improved. The incidence of IHCA in the United States increased to 900 to 1000 per 100,000 hospitalized patients between 2008 and 2017, compared to 160 per 100,000 in the United Kingdom from 2011 to 2013 and 840 per 100,000 in China as of 2020 [2-4]. Meanwhile, the estimated averages of incidence of OHCA under emergency medical services (EMS) in North America, Asia, and Europe from 2010 to 2020 were 47.3, 45.9, and 40.6 per 100,000 people, respectively. The estimated averages of the survival rates of IHCA from 2010 to 2020 were 25% in the United States, 18% in the United Kingdom, and only 9.4% in China. For OHCA, the estimated averages of the survival rates during this same period were 10% to 12% in the United States, 8% in Europe, and just 3.6% in Asia [2,5]. These low survival rates also impose significant economic burdens on nations. According to relevant reviews, the cost-effectiveness threshold for CA ranged from US $20,000 to US $150,000 per quality-adjusted life year, with each life saved potentially reducing costs by US $19,000 to US $71,000 per case [6].

Although efforts to establish CA centers independently began in various regions of the United States as early as 2000 to 2010 [7] and Germany initiated CA center certification throughout the country [8] in August 2019 aiming to provide evidence-based, bundled care to improve CA survival rates, CA remains a formidable clinical challenge. If resuscitation is not timely, the patient may lose consciousness within approximately 10 seconds, with irreversible hypoxic-ischemic brain injury occurring within 4 minutes [9], and if resuscitation is delayed beyond 10 minutes, survival is practically impossible [10,11]. Thus, early prediction and identification of CA are critical factors in preventing death and poor outcomes and represent a major challenge that requires urgent clinical attention.

However, there is a scarcity of efficient, internationally recognized, and universally accepted assessment tools for early prediction and identification of CA risk and adverse outcomes. In recent years, with the rapid advancement of artificial intelligence (AI), many researchers have used machine learning (ML) to address clinical challenges. Commonly used ML approaches can be broadly categorized into supervised and unsupervised learning. In the context of supervised ML, clinical predictors can be incorporated into various models. In these models, their parameters are adjusted based on outcome variables to generate predictions regarding the probability of positive event occurrence [12]. It is now common to see ML being used to predict disease progression and even to diagnose and treat complex diseases effectively. For instance, in 2019, several authors, including Hatib et al [13] and Wijnberge et al [14], successfully predicted intraoperative hypotensive events using ML, leading to the clinical translation of these models into products that significantly enhanced patient safety during surgical anesthesia [15]. By 2023, some researchers had similarly affirmed the substantial potential of ML models in cancer detection, prognosis, and treatment, recognizing their exciting discoveries and contributions to advancing medical practice [16]. The aforementioned studies were based on supervised ML and the extensive use of interpretable clinical features to construct predictive models and simultaneously demonstrate the promising predictive performance of ML in clinical events across various fields. In this context, some researchers have also developed different ML models for risk prediction in CA. Recent reviews by Sem et al [17] and Chen et al [18] indicate that ML appears to exhibit high accuracy in both the management and risk prediction of CA. However, these reviews do not quantitatively synthesize the results of ML models, which significantly limits our ability to interpret the specific value of various ML models in CA applications and the selection of appropriate models. Therefore, we conducted this systematic review and meta-analysis to review the predictive performance of ML for the occurrence of CA, good neurological prognosis after CA, mortality, and the return of spontaneous circulation (ROSC) after CA to provide evidence-based guidance for the development and updating of simple prediction tools with high accuracy and direct access to results.

This study was conducted in adherence to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and prospectively registered with PROSPERO (International Prospective Register of Systematic Reviews; ID CRD42024518949). The detailed PRISMA checklist is presented in Multimedia Appendix 1.

Detailed inclusion and exclusion criteria were defined to screen the original studies relevant to our systematic review from the retrieved literature (Textbox 1).

A systematic search of the PubMed, Embase, Cochrane Library, and Web of Science databases was carried out from their inception to May 17, 2024. The search strategy involved controlled vocabulary and free-text terms, with no restrictions on geographical location or publication year. The detailed search strategy is presented in Multimedia Appendices 2-5.

The retrieved studies were imported into EndNote X9 (Clarivate Analytics). Their titles and abstracts were reviewed. After the exclusion of duplicates, the preliminary eligible original studies were selected and their full texts downloaded for determining the final inclusion. An electronic spreadsheet was prepared to extract the following information: first author, publication year, author’s country, study type, patient source, prediction events, data balance, location of CA occurrence, number of cases with study events, total number of cases, number of cases in the training and validation sets, method of validation set generation, missing data–handling methods, and types of models used. Study selection and data extraction were independently conducted by 2 researchers. Disagreements were discussed and resolved with a third author.

The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to assess the risk of bias in all the included original studies. PROBAST comprises several questions across 4 domains—participant, predictor, outcome, and statistical analysis—which reflect the overall risk of bias and applicability. These domains consist of 2, 3, 6, and 9 questions, respectively, each with 3 possible answers (Yes or Probably yes, No or Probably no, and No information). A domain was classified as high risk if any question was answered with No or Probably no. Conversely, a domain was regarded as low risk only if every question was answered with Yes or Probably yes. The overall risk of bias was assessed as low when all domains were deemed to be low risk. When at least one domain was high risk, the overall risk of bias was rated as high. In total, 2 authors independently assessed the risk of bias using PROBAST and cross-checked their findings. Any discrepancies were addressed by consulting with a third author to reach agreement.

The primary outcome measure was the C-index, which reflects the predictive ability of ML models for IHCA and OHCA. However, we found that the C-index might not have accurately described the predictive performance of ML for positive events, particularly in models built on severely imbalanced data, as these original studies often suffered from such imbalance. This limitation was evident in predicting the occurrence of IHCA and OHCA, the good cerebral performance category score 1 to 2 (CPC 1-2), mortality, and ROSC. Therefore, in addition to the C-index, our primary outcome measures encompassed sensitivity and specificity. Our secondary outcome was the frequency of variables used in the ML models.

A meta-analysis of the C-index, a measure of the general accuracy of ML models, was carried out. When the 95% CI and SE for the C-index were missing in some studies, the SE was estimated based on the study by Debray et al [19]. Due to the differences in the included variables and inconsistent parameters among the ML models, random-effects models were prioritized in the meta-analysis of the C-index. In addition, a meta-analysis on sensitivity and specificity was conducted through a bivariate mixed-effects model based on diagnostic 2 × 2 tables. However, most original studies did not report these tables. In such cases, we used the following methods to calculate the 2 × 2 tables: (1) calculation based on sensitivity, specificity, precision, and case numbers; and (2) calculation based on the best Youden index to extract sensitivity and specificity, followed by case number integration. Nevertheless, this method allowed for meta-analysis only when there were ≥4 models. For <4 models, we presented the range of sensitivity and specificity. Our meta-analysis was conducted in R (version 4.2.0; R Foundation for Statistical Computing).

A total of 1270 articles were obtained from databases, with 599 (47.17%) being duplicates. Among these 599 duplicates, 471 (78.6%) were found to be duplicates via software, and 128 (21.4%) were manually identified as duplicates. After the elimination of duplicates, 671 articles were screened by title and abstract, with 169 (25.2%) being selected for full-text review. After the exclusion of conference abstracts published in full text without peer review (19/169, 11.2%), studies with risk factor analyses but no complete ML models (22/169, 13%), studies lacking outcome indicators (28/169, 16.6%), and studies with severe statistical errors (7/169, 4.1%), a total of 93 articles were included finally. The detailed process is illustrated in Figure 1.

The 93 selected studies were published between 2011 and 2024, covering 14 countries, primarily South Korea, China, Japan, the United States, and Singapore. Among the 93 studies, there were 23 (25%) prospective cohort studies and 3 (3%) case-control studies, with the remainder (67/93, 72%) being retrospective cohort studies. Data for 26% (24/93) of the studies were sourced from multiple centers, whereas 37% (34/93) of the studies used data from registry databases and the rest (35/93, 38%) were single-center studies. In 30% (28/93) of the studies, the predicted outcome was the occurrence of CA. In 42 studies, the predicted outcome was the neurological prognosis of patients with CA, with 10 (24%) studies focused on patients with IHCA and the remainder (n=32, 76%) focused on patients with OHCA. In 27% (25/93) of the studies, the predicted outcome was CA mortality, and in 12% (11/93) of the studies, the predicted outcome was ROSC in patients with CA. The 93 studies collectively encompassed a total of 5,729,721 cases, including 1,737,085 OHCA cases and 3,992,636 IHCA cases. Regarding the predictive models constructed, 81% (75/93) of the studies had independent validation sets, but only 27% (25/93) used external validation, primarily using k-fold cross-validation and random-sampling internal validation methods. A total of 34% (32/93) of the studies described methods to prevent data overfitting, mainly through cross-validation. In total, 17 types of ML models were involved, with logistic regression (LR), random forest (RF), deep learning, and decision trees (DTs) being the most prominent. In addition, these studies validated several previously established scoring tools, including the Cardiac Arrest Neurological Prognosis score, distance scoring system, Emergency Department In-Hospital Cardiac Arrest Score, FACTOR score, Modified Early Warning Score, National Early Warning Score, National Early Warning Score 2, OHCA score, proposed scoring system, Rapid Emergency Medicine Score, Simplified Acute Physiology Score II, Cardiac Arrest Hospital Prognosis score, and ROSC after CA score. The details of the included studies are shown in Tables 1 and 2.

After our exclusion of previously established scoring tools, a quality assessment of 208 ML models, involving 17 types, was conducted. In total, 24% (50/208) of these models originated from case-control studies, which introduced a high risk of bias in study participant selection. Regarding predictive factors, 1% (2/208) of the models were linked to a high risk of bias owing to the use of outcome information. Regarding outcome assessment, as both CA and prognosis outcomes were clearly defined using standard definitions, no additional predictive factors were required, resulting in a low risk of bias in outcome assessment. The included ML models were primarily derived from large-sample statistical analyses; however, 9.1% (19/208) of the models were based on a very small number of cases, with an event per variable value of <10. In addition, inappropriate deletion methods were applied to address missing data in 65.4% (136/208) of the models, and only univariate analysis was used to screen for predictive factors in 22.6% (47/208) of the models, ultimately resulting in a high risk of bias for 76.9% (160/208) of the models in the domain of statistical analysis, as detailed in Figure 2.

Modeling variables were extracted and weighted for analysis from the 93 studies on ML models for predicting CA and CPC 1-2. Among the 28 studies on predicting CA, the variables with the highest weights were respiratory rate (n=22, 79%), blood pressure (n=20, 71%), age (n=19, 68%), temperature (n=19, 68%), oxygen saturation (n=15, 54%), and airway (n=9, 32%). Among the 42 studies on predicting CPC 1-2, the results showed that the modeling variables with the highest weights in the IHCA group were rhythm (shockable or nonshockable; 8/10, 80%), age (7/10, 70%), medication use (6/10, 60%), gender (5/10, 50%), and Glasgow Coma Scale (GCS; 5/10, 50%). The modeling variables with the highest weights in the OHCA group were age (25/32, 78%), rhythm (shockable or nonshockable; 24/32, 75%), medication use (18/32, 56%), ROSC (14/32, 44%), gender (12/32, 38%), no-flow time (resuscitation duration; 12/32, 38%), EMS transport (scene interval, arrival time, and response time; 12/32, 38%), defibrillation (11/32, 34%), and GCS (6/32, 19%). The detailed results of the modeling variables and weight analysis are presented in Tables S6 and S7 in Multimedia Appendix 20.

It was observed that ML has garnered widespread attention among numerous researchers in the management of CA, particularly focusing on early CA risk prediction in both in-hospital and out-of-hospital populations. Our systematic review and meta-analysis demonstrated a relatively favorable predictive value of ML in the validation set for forecasting CA risk, with a C-index of 0.89. Similarly, ML also appeared to exhibit a relatively favorable predictive value for neurological outcomes (CPC 1-2) and mortality in patients who had already experienced CA, with pooled C-indexes of 0.86 and 0.85, respectively. However, in predicting ROSC following CA, ML seemed to display a predictive value comparable to that of traditional scoring tools, with a pooled C-index of 0.77.

Currently, in clinical practice, classic early warning scoring tools, including the Cardiac Arrest Risk Triage (CART) score, Modified Early Warning Score, and VitalPAC Early Warning Score, are commonly used for predicting the occurrence of CA. A previous review by Churpek et al [112] found that, among these tools, the CART had the highest accuracy in predicting CA compared to the others. However, the CART had certain limitations, with an area under the curve of 0.83, sensitivity calculated at 0.61 based on the optimal Youden index, and a specificity of 0.84. Moreover, the CART has not been externally validated, and the included population is limited to ward inpatients. Therefore, whether the CART can dynamically monitor the occurrence of CA in real-time clinical events, improve rescue success rates, and enhance patient outcomes requires prospective validation using high-quality, large-sample external data. Our summarized results of the ML models reveal that ML has certain clinical predictive value in forecasting the occurrence of CA and demonstrates relatively favorable accuracy, with an overall C-index of 0.89, sensitivity of 0.83, and specificity of 0.93. Comparatively, this is superior to previous scoring tools, providing a certain clinical basis for the future establishment of more reliable early warning scoring systems for predicting CA.

In a recent review by Carrick et al [113], the accuracy of scoring tools for predicting survival or neurological outcomes following CA, such as the OHCA score, Cardiac Arrest Hospital Prognosis score, and Good Outcome Following Attempted Resuscitation score, was summarized. These 3 tools, which have undergone rigorous clinical validation, exhibited relatively high accuracy, with C-indexes of 0.79, 0.83, and 0.76, respectively. However, our summarized results of ML models suggested that ML seems to exhibit more favorable accuracy, with an overall C-index of 0.86, sensitivity of 0.72, and specificity of 0.79 for predicting favorable neurological outcomes. For predicting CA mortality, ML achieved an overall C-index of 0.85, sensitivity of 0.83, and specificity of 0.79.

Among various ROSC prediction models for CA that have been developed in the current clinical field, the ROSC after CA score developed by Gräsner et al [114] using data from 5471 patients with OHCA from the German Resuscitation Registry has attracted the most attention. It has been externally validated in several European and Asian countries, demonstrating relatively good accuracy, with an area under the curve of 0.736 in a recent large-scale external validation study [115]. However, our summarized results of ML models indicated that the overall C-index of ML was 0.77, with a sensitivity of 0.53 and specificity of 0.88. Comparatively, these ML models did not seem to significantly outperform traditional scoring tools in predicting ROSC outcomes for patients with CA.

In our review, the modeling variables of the discussed models primarily originated from common clinical features. It was found that variables such as respiratory rate, blood pressure, age, temperature, oxygen saturation, and airway were key predictors in existing ML models, and they also constitute critical variables in traditional scoring models [116,117]. Therefore, the impact of these variables on predicting the occurrence of CA is well established. However, these predictors differ to some extent from the findings of recent studies, such as the review by Andersen et al [3], which identified CA risk factors. The review by Andersen et al [3] suggested that age was more associated with post-CA prognosis and reduced survival rates, whereas a history of cardiac diseases such as myocardial infarction, arrhythmias, and heart failure was recognized as the most common risk factor for CA occurrence. Other potential risk factors included the use of certain medications, such as those that prolong the QT interval, opioids, and sedatives. Nonetheless, the review concurred that respiratory function and body temperature also had predictive significance for CA, with early interventions targeting these factors being crucial for achieving reversible outcomes [118].

In the models we reviewed that aimed to predict CPC 1-2 outcomes in patients with IHCA and OHCA, the modeling variables with the highest weight were age, rhythm (shockable or nonshockable), medication use, ROSC, gender, no-flow time (resuscitation duration), defibrillation, EMS transport (scene interval, arrival time, and response time), and GCS. When compared to the review by Sandroni et al [119], which highlighted the predictive value of GCS, biological markers (eg, neuron-specific enolase), and electrophysiological indicators (eg, somatosensory evoked potential) for favorable neurological outcomes, our findings show some differences. In addition, complex variables such as medical imaging might need consideration in clinical practice. In recent years, AI methods have been widely used in medical imaging for identifying disease progression and prognosis, demonstrating superior accuracy and cost-effectiveness compared to traditional clinical feature–based predictive models [120]. Therefore, in the prediction of CA occurrence and prognosis, the high-value variables identified in recent studies, such as electrocardiography [121] and ultrasound [122], are not reflected in traditional scoring tools. This raises the question of whether it is worth further validating these more complex variables and attempting to identify more efficient predictive factors to develop or update risk-scoring tools in the field of CA.

Our study reveals that ML methods appear to outperform traditional scoring tools in predicting the occurrence and progression of CA. Therefore, the development of simple auxiliary tools based on ML theory is recommended to facilitate rapid risk screening of CA for both in-hospital and out-of-hospital patients, enabling timely formulation of appropriate treatment strategies. These ML-based CA prediction models would be particularly beneficial for emergency departments and out-of-hospital response teams. Under the current circumstances, in which emergency departments worldwide are facing challenges of overcrowding, resource limitations, and a high influx of patients who are critically ill [123-125], relying solely on human assessment of CA risk based on various clinical data could pose significant challenges to the efficiency of CA treatment. Furthermore, the complexity and volume of clinical data, including patient demographics, laboratory results, imaging data, and textual notes from health care providers, are continuously increasing. Thus, using ML to analyze large datasets and handle complex variables, such as clinical images, seems to be a more feasible approach [126]. The development of simplified ML prediction tools or intelligent reading systems has the potential to mitigate risks such as treatment delays and poor prognoses in patients with CA in emergency departments. These tools could also enhance health care service quality, reduce human resource costs, and support the formulation of targeted therapeutic strategies. Similarly, ML models that incorporate real-time input of variables such as vital signs, electrocardiograms, and response times for out-of-hospital rescue scenarios can accurately predict positive CA events. This capability aids response teams in avoiding repeated and frequent evaluations, enabling timely decisions on whether to implement preventive therapeutic interventions to avert CA or, in cases of CA occurrence, whether to initiate extracorporeal membrane oxygenation cardiopulmonary resuscitation to improve survival rates [127,128].

In addition, our findings indicate that the balance of data significantly impacts the outcomes of ML model construction in CA-related studies. This effect is particularly pronounced in scenarios in which the outcome metrics exhibit severe imbalance. For instance, in predicting the occurrence of CA in our study, the rationale for the selected predictive factors remained challenging, and the accuracy of the constructed model was often influenced by the overwhelming proportion of negative events [129]. In such cases, the C-index hardly represented the actual outcome prediction accuracy of the model. Therefore, in our study, the sensitivity and specificity of ML models were also summarized [130]. The data balance in the studies we included was primarily addressed using oversampling, but these studies rarely considered validating models constructed from balanced data against imbalanced data. This raises certain doubts regarding the accuracy of existing models constructed based on balanced data when applied to real-world cases of CA. Our study reveals that CA events in the real world are often inherently imbalanced. Therefore, we recommended prioritizing the use of models constructed from real-world data or validating models constructed from balanced data against real-world data to ascertain their true effectiveness [131].

Although ML models demonstrated relatively satisfactory accuracy in predicting the occurrence and progression of CA in our study, several common challenges inherent to ML modeling should be acknowledged. For instance, compared to traditional scoring tools, ML models rely on general algorithms to generate desired outputs in response to specific input data, a process characterized by less explicit rules [132]. In addition, algorithmic biases may result in unrepresentative datasets [133], and the reliability of model validation remains a concern [134]. These issues underscore the ethical challenges associated with the application of AI in medicine, including result interpretability, algorithmic transparency, predictive fairness, and data privacy [135,136]. These potential ethical concerns are specifically reflected in a patient survey study on the prevention of CA occurrence and development conducted by Maris et al [137]. The study results indicate that, while AI-driven CA treatment decisions offer objective data, the absence of patient involvement and informed consent, along with the interpretability of the model, suggest that the overuse of AI technology may ultimately undermine patient trust in physicians. This, in turn, poses challenges to the current high-quality health care goal of patient-centered care [138] in the field of cardiovascular disease treatment, which is built on shared decision-making [139], respect for patient autonomy, and mutual trust. Therefore, in the high-risk and critical treatment of CA, it seems that physicians should continue to make final decisions in collaboration with ML models based on evidence-based clinical experience and the values of the patients.

Therefore, based on ethical considerations, the choice of different models during research remains challenging as model interpretability and accuracy are factors that need to be considered comprehensively during model construction [140]. Selecting models with higher interpretability, such as LR, Cox regression, or DTs, can facilitate better communication, interaction, and trust between health care providers and patients. However, these models may have limited predictive value for certain outcome events [141]. On the other hand, models whose interpretability is poorer, including neural networks, support vector machines (SVMs), and Extreme Gradient Boosting, often perform exceptionally well in predicting outcomes [142]. At this point, it may become necessary to grant patients and their families greater rights to information and autonomy, enabling their active participation in medical decision-making. In our research, LR was the most frequently used model type as it facilitated the development of predictive nomograms, which are simple and easily applicable tools.

Among the 17 ML models that we included, artificial neural networks, RF, and LR appeared to demonstrate relatively ideal predictive value in forecasting the occurrence of CA and were more frequently used by clinical researchers. In predicting neurological outcomes, our study found that LR remained the model most commonly selected by clinicians, followed by DT, RF, SVM, and others. If we aim to develop a simplified predictive scoring scale to assist in clinical practice, priority may be given to using LR for its development and subsequent updates. This preference arises because, according to our research findings, LR demonstrates relatively satisfactory accuracy and facilitates the construction of straightforward and practical predictive nomograms [143,144]. Furthermore, considering the interpretability of models is essential in real-world clinical practice. However, if the objective is to develop auxiliary applications for disease surveillance and prediction in clinical settings, alternative, more complex models may be considered. For example, models such as neural networks, SVM, and Extreme Gradient Boosting, which demonstrated higher accuracy in our study, could be appropriate choices. On the other hand, when dealing with image-based features, such as medical imaging or electrocardiograms, it may be necessary to use models with lower interpretability, such as those based on deep learning [145], rather than confining the analysis to commonly used clinical features with stronger interpretability.

In addition, we observed a minimal number of studies that constructed models based on artificial neural networks and ensemble learning, which exhibited highly favorable results, indicating that further validation of these models may be required in future research. Currently, in clinical practice, there is an increasing preference for using simple scoring tools based on interpretable clinical features. While opting for such tools may reduce the ethical dilemmas encountered in clinical settings, relying solely on traditional methods and highly interpretable clinical indicators, such as the Delphi method, during the development of these scoring tools seems to introduce significant bias into the constructed models. Therefore, we considered using multicenter real-world big data, using ML approaches, and incorporating a broader range of cases and clinical features to construct interpretable scoring tools and promote their application. Regarding the processing of clinical image features, our expectation lies in the development of intelligent reading tools based on deep learning methods. Nonetheless, in our study, there was limited research on deep learning based on medical imaging and ultrasounds, particularly in the field of CA, where such research remains relatively underexplored. Therefore, future research on CA should actively explore the integration of medical imaging and ultrasonography. In selecting datasets and algorithms for the development of AI prediction models, it is crucial to rigorously investigate and address the ethical shortcomings of AI applications in health care. Efforts should be made to minimize the influence of individual characteristics such as gender, race, skin color, and socioeconomic status, ensuring that population representation and sample size are carefully considered. Sufficient numbers and the quality of representative populations should be selected from diverse regions, ethnicities, and age groups to establish standardized big data models, thereby maximizing the potential of AI technologies [146,147]. It is essential to adopt a multifaceted, interdisciplinary approach; strengthen data protection systems to prevent the leakage of patient information; and conduct extensive reviews to avoid biases [148], ultimately preventing unfairness toward individuals or patient populations in the development of intelligent diagnostic or predictive tools for CA [149,150].

Our study has 3 strengths. First, it represents the first attempt to summarize evidence comparing ML models and scoring tools in predicting the occurrence and prognosis of CA, thereby providing an evidence-based foundation for the subsequent clinical update and development of new scoring tools or AI early warning systems in the field of CA. Second, our study encompassed 93 original studies with large sample sizes, covering 14 countries and involving 5,729,721 patients, significantly enhancing the strength of our evidence. Third, we conducted a detailed discussion of the accuracy of different models on balanced and imbalanced data. However, this study also had the following limitations. First, most of the original studies on the prediction of CA occurrence (28/93, 30%) constructed models based solely on imbalanced data without validating them on balanced data. Second, many model validation processes primarily used internal validation through random sampling, lacking external multicenter validation to examine their generalizability. Third, due to potential differences in the predictive performance of different models for outcome events, despite our in-depth discussion of various ML models and datasets, the limited number of studies on certain ML models restricted our ability to interpret the results of ML applications in CA more comprehensively. Fourth, due to the small number of included studies, we did not strictly distinguish between IHCA and OHCA populations when summarizing the predictors of CA. Fifth, as this review only included English-language original studies, there may be potential language bias.

Current traditional scoring tools have demonstrated relatively ideal efficacy in predicting the occurrence and prognosis of CA. On the basis of this review, ML appeared to offer greater advantages in predicting the occurrence of CA, neurological functional prognosis, and mortality outcomes. However, for predicting outcomes associated with ROSC after CA, ML models did not seem to significantly outperform traditional models. Therefore, in future studies on CA, researchers may explore the systematic updating of traditional scoring tools based on the superior performance of ML in specific outcomes. This approach would enable the implementation of AI-driven enhancements within complex and diverse clinical data, thereby assisting clinicians in monitoring and providing early warnings for multiple predictive factors. For outcomes that are still unpredictable, multicenter large-sample studies are warranted.