Our review questions are as follows:

This review will not involve direct contact with human participants. All clinical data (ie, laboratory test results and vital signs) used in ML model construction will be presented as aggregated and deidentified data from previously published studies. This protocol was registered in the Open Science Framework on November 16, 2024 [43].

The focus of this review is to synthesize the clinical features used to construct ML models and evaluate the performance of these models in predicting IHCA.

The inclusion and exclusion criteria are as follows. We will include only peer-reviewed, English-language studies published between April 2009 and April 2024 that used ML algorithms to predict IHCA. The selection of 2009 as the starting year for our search was based on significant advancements in ML applications within health care around that time. In 2009, a prominent national report indicated that only a small proportion of US hospitals had adopted a “basic” EHR system [44]. In addition, the Health Information Technology for Economic and Clinical Health Act provided incentives for health care providers who used EHRs. This period marked the beginning of more widespread adoption of EHRs and the integration of advanced computational methods, which facilitated the application of ML algorithms in clinical settings. These technological advancements enabled more accurate and efficient prediction models for IHCA. Therefore, studies published after 2009 are more likely to reflect the current state of ML applications in predicting IHCA [45,46]. We will exclude the following types of articles:

Regarding the patient population, only studies involving adults aged ≥18 years who have experienced IHCA will be included. There will be no restriction based on sex, gender, geographic location, race, or types of ML techniques used. We will use the Utstein Out-of-Hospital Cardiac Arrest Resuscitation Registry Template definition of IHCA [1], which refers to the delivery of defibrillator shocks or chest compressions to a patient admitted to an inpatient bed. Studies that do not adhere to this definition or that did not collect data from hospitals or medical centers will be excluded. In addition, studies that include cases of death or sudden death without resuscitation or OHCA and animal studies will be excluded.

Regarding study outcomes and ML metrics, some studies list composite outcomes, including IHCA, admission to intensive care units, and patient deterioration. We will include studies that report ML metrics for predicting IHCA, even when they are part of a composite outcome, as long as the prediction of IHCA is clearly specified and distinguishable from other outcomes. This approach will allow us to capture a broader range of clinically relevant studies while maintaining focus on IHCA prediction.

Four reputable databases were searched: PubMed (biomedical research), Web of Science (a global citation database in the sciences, arts, and humanities), IEEE Xplore (for technology and engineering research), and Embase (medical research). The following 2 sets of search terms—“in-hospital cardiac arrest” and “machine learning”—were entered into the databases connected using AND after several trials to select different text words.

Before protocol registration on the Open Science Framework in November 2024, we conducted a preliminary exploratory search to refine the search terms, databases, and feasibility of conducting this review. While this process involved screening a small set of articles for pilot-testing purposes, the formal and comprehensive literature search used for study inclusion and data extraction was carried out after protocol registration. This preregistration phase did not influence the core research objectives, inclusion and exclusion criteria, or data extraction domains. An interprofessional informationist team member worked with disciplinary faculty and created the base search to refine key concepts. The informationist completed a pilot search in PubMed, which a team member screened to determine retrieval accuracy for key articles on IHCA and ML. For example, we initially selected the following search terms and keywords using the MeSH (Medical Subject Headings) database:

(“In-hospital cardiac arrest” [text word] OR “In-hospital cardiac arrest” [tiab] OR “cardiopulmonary resuscitation” [text word] OR “heart arrest” [text word] OR “sudden cardiac death” [text word] OR asystole [text word] OR heart arrest [MeSH] OR asystole [MeSH] OR cardiopulmonary arrest [MeSH] OR arrest, cardiopulmonary [MeSH]) AND (“Artificial intelligence” [text word] OR “artificial intelligence” [tiab] OR Artificial intelligence [MeSH] OR Intelligence, Artificial [MeSH] OR Computational Intelligence [MeSH] OR Intelligence, Computational [MeSH] OR Machine Intelligence [MeSH] OR Machine Learning [MeSH] OR Deep Learning [MeSH] OR Supervised Machine Learning + [MeSH] OR Unsupervised Machine Learning [MeSH] ) 2009-2024

Each trial was conducted using carefully selected keywords involving Boolean operators and MeSH options along with other text words with or without the assigned years (2009-2024), which produced a set number of articles. These articles were reviewed by a team member with expertise in IHCA and familiarity with ML techniques. A discussion was then initiated between the informationist and the team member to improve search terms for targeted peer-reviewed articles in PubMed. This process was iterated 7 times, producing a range of published articles from 95 to 2004. On the basis of the search sensitivity, revisions were made to the search query, which was then translated from PubMed into 3 additional databases—Web of Science, Embase, and IEEE Xplore—for hand searching. A second research librarian will conduct a full Peer Review of Electronic Search Strategies review of the search and will recommend revisions.

All search results will be downloaded and exported into Rayyan (Qatar Computing Research Institute) for deduplication. Automated tools in Rayyan will be used to identify and remove duplicates. Once duplicates are removed, the titles and abstracts will be screened according to the exclusion criteria, and the articles that do not meet the criteria will be excluded. The full texts of the remaining articles will then be reviewed to determine eligibility, with reasons for exclusion recorded and reported in the final scoping review. Two reviewers will independently conduct the screening process using the Rayyan application. Any disagreement between the reviewers will be resolved through discussion, with input from experts in resuscitation and data sciences. The results of the search (ie, the included articles) will be presented in a PRISMA flow diagram.

Our review aims and questions will guide the types of data elements for extraction. The targeted data elements will be identified and extracted from the eligible full texts of the articles by 2 team members who will record them as variables in a developed template using Microsoft Excel. We will have several sheets within a Microsoft Excel file to complete data charting (eg, demographics and biomarkers). The Microsoft Excel file can be considered our tool and will be revised as necessary during the extraction phase. The extracted data will be reviewed by the third team member for accuracy. To resolve any conflict, a discussion among all these team members will take place to reach consensus. A list of data elements for future extraction is presented in Figure 1, which includes study characteristics, ML metrics, medical diagnosis, inclusion and exclusion criteria, vital signs, population characteristics, laboratory test results, and definitions and number of outcomes. We specifically pay attention to 2 important criteria, including representativeness of the population by examining the sample sizes in subgroups (eg, men and women) of the total population and in the construction of ML models (eg, training and testing phases). The physiological differences between men and women that are represented by their signs and symptoms will also be evaluated. We will also determine the interpretability of the clinical data through the utility of different tools, including Shapley additive explanations.

The vital sign metrics comprise heart rate or pulse rate, systolic blood pressure, diastolic blood pressure, respiratory rate, temperature, and oxygen saturation. These metrics were selected after reviewing the literature [27,47,48] and a discussion among 2 investigators in this scoping review with medical knowledge and expertise (MA and KB). For example, oxygen saturation is typically not included in its traditional format [49], or certain vital sign metrics were not selected in some studies (eg, diastolic blood pressure) [47,48]. We decided to incorporate all vital sign metrics to comprehensively evaluate the quality and quantity of the clinical data.

Descriptive statistics, including numbers, percentages, means, and SDs, will be reported to quantify the selected variables, and a comprehensive narrative will provide a detailed synthesis of variables across the articles. We will extract clinical features used in ML models for predicting IHCA, including patient demographics (eg, age and sex) and medical histories (eg, diagnoses), vital signs, laboratory test results, and all other available data. We will also gather information regarding the data sources (eg, EHRs and registries) and the methods used to handle missing or incomplete data. We will evaluate the number of features included in the models and the sample sizes used for constructing ML models. The number and type of all ML models (ie, traditional vs deep neural network) will be presented in a graph. The metrics of ML model performance will be presented in a table including the area under the curve, sensitivity, specificity, negative predictive value, positive predictive value, accuracy, precision, and F₁-score (Figure 2). We will compare the performance metrics of different ML models to understand how data quality influences model outcomes. Subgroup analyses will also be conducted to evaluate the effects of different types of clinical data (eg, vital signs vs laboratory test results) on predictive performance. To ensure alignment with our review objectives, the synthesis of extracted data will be directly mapped to the review questions. We will group and summarize findings on the geographical distribution of studies, data quality and quantity, shared clinical predictors, and performance metrics using descriptive statistics (eg, frequency, percentages, means, and SDs, described previously) and a narrative approach. This dual synthesis strategy will allow us to highlight patterns in ML model inputs and outputs, methodological heterogeneity, and gaps in reporting practices. Figure 1 outlines the core variables identified for extraction and synthesis, reinforcing the linkage between data elements and our research questions.

We will conduct a comprehensive evaluation of the clinical features extracted from EHRs for each phase of ML model development, including training, validation, and testing. This evaluation will focus on assessing population diversity, such as age, sex, race and ethnicity, and comorbidities, to ensure inclusivity and representation in model development. Recognizing that EHR data often contain time-series information, we will thoroughly analyze the temporal characteristics of clinical features, particularly in relation to the timing of IHCA onset. This includes exploring patterns, trends, and critical time points that could enhance the predictive power of ML models.

Furthermore, we will identify and discuss the significance of commonly reported predictors across studies, such as vital signs, laboratory test results, and comorbid conditions, and critically evaluate their clinical implications. This will provide insights into the relevance of these predictors in the context of IHCA and their potential utility in real-world clinical settings.

Special emphasis will be placed on enhancing the interpretability of ML models, addressing the black box challenge often associated with these approaches. Techniques such as feature importance analysis, visualization of decision pathways, and integration of domain knowledge will be explored to improve the transparency and usability of ML predictions. The discussion will also cover the metrics and methodologies for optimizing model performance, including sensitivity, specificity, precision, recall, and calibration, with a focus on their implications for clinical decision-making and patient outcomes. Through this synthesis, we aim to support the standardization of reported clinical features and methodologies in the development of predictive ML models for IHCA. This standardization will facilitate comparability across studies; enhance reproducibility; and, ultimately, contribute to the broader adoption of ML tools in clinical practice for early detection and prevention of IHCA.

This study will have several strengths and limitations. Among its strengths, it will provide a comprehensive analysis of clinical features extracted from EHR data considering population diversity and temporal dynamics. This ensures that the findings are inclusive and applicable across varied patient populations. The focus on temporal characteristics is another strength as it addresses the critical aspect of timing in predicting IHCA, enabling better anticipation of adverse events. In addition, this study places significant emphasis on interpretability, tackling the black box nature of ML models and ensuring that results are transparent and usable by clinicians, fostering trust and adoption in health care. Furthermore, the synthesis of predictors and features across studies will support standardization in reporting practices, enhancing reproducibility and comparability in ML model development. The emphasis on clinical relevance will bridge the gap between technical advancements and practical implementation, whereas considerations of population diversity will promote equity and reduce biases in predictive models.

However, this study will also face certain limitations. The quality, completeness, and standardization of EHR data, which can vary across institutions, may impact the analysis and model performance. Despite efforts to include diverse populations, biases in available datasets could limit the generalizability of findings, especially for underrepresented groups. While interpretability is emphasized, some ML models may remain opaque due to their inherent complexity. In addition, a scoping review has inherent limitations that stem from its methodology and objectives, which focus on mapping existing literature rather than providing definitive answers to specific research questions. One primary limitation is the lack of depth in analysis as scoping reviews typically do not involve a critical appraisal of the quality of the included studies. This means that studies with varying levels of rigor may be included, potentially affecting the reliability of the findings. In addition, the heterogeneity of evidence, including diverse study designs, populations, and outcomes, can make it challenging to synthesize findings into cohesive conclusions. Unlike meta-analysis, scoping reviews do not perform formal meta-analyses or provide pooled effect estimates, limiting their ability to quantify relationships or provide definitive conclusions. The broad scope of this scoping review may also introduce selection bias as the inclusion of studies depends on the search strategy, availability of data, and reviewer judgment. Moreover, this scoping review will rely heavily on published and accessible literature, potentially overlooking unpublished studies, gray literature, or ongoing research, leading to publication bias. Despite these challenges, this study provides valuable insights and a balanced contribution to the development and evaluation of ML models for predicting IHCA.

Our scoping review will add to the body of IHCA literature by synthesizing different aspects of ML studies to predict IHCA. This will be our first step to comprehend the complexity of ML studies and promote realistic options for the clinical decision-making process.