R E S U S C I T A T I O N P L U S 15 (2023) 100435

Available online at www.sciencedirect.com

Resuscitation Plus
journal homepage: www.elsevier.com/locate/resuscitation-plus

AI and machine learning in resuscitation:

Ongoing research, new concepts, and key
challenges

Yohei Okada a,b,*, Mayli Mertens c,d, Nan Liu a, Sean Shao Wei Lam a,
Marcus Eng Hock Ong a,e

Abstract
Aim: Artificial intelligence (AI) and machine learning (ML) are important areas of computer science that have recently attracted attention for their
application to medicine. However, as techniques continue to advance and become more complex, it is increasingly challenging for clinicians to stay
abreast of the latest research. This overview aims to translate research concepts and potential concerns to healthcare professionals interested in
applying AI and ML to resuscitation research but who are not experts in the field.
Main text: We present various research including prediction models using structured and unstructured data, exploring treatment heterogeneity, rein-
forcement learning, language processing, and large-scale language models. These studies potentially offer valuable insights for optimizing treatment
strategies and clinical workflows. However, implementing AI and ML in clinical settings presents its own set of challenges. The availability of high-
quality and reliable data is crucial for developing accurate ML models. A rigorous validation process and the integration of ML into clinical practice is
essential for practical implementation. We furthermore highlight the potential risks associated with self-fulfilling prophecies and feedback loops,
emphasizing the importance of transparency, interpretability, and trustworthiness in AI and ML models. These issues need to be addressed in order
to establish reliable and trustworthy AI and ML models.
Conclusion: In this article, we overview concepts and examples of AI and ML research in the resuscitation field. Moving forward, appropriate under-
standing of ML and collaboration with relevant experts will be essential for researchers and clinicians to overcome the challenges and harness the full
potential of AI and ML in resuscitation.
Keywords: Prediction model, Natural language processing, Heterogeneity, Self-fulfilling prophecy, Feedback loop, Large language model,
Emergency medicine

strategies. However, as techniques continue to advance and

Introduction become more complex, it is increasingly challenging for clinicians
to stay abreast of the latest research involving AI and ML techniques
Artificial intelligence (AI) and Machine learning (ML) are important in the resuscitation field.
areas of computer science that have recently attracted attention for This review aims to introduce recent AI and ML research to
their combined application to medicine. AI refers to technology in healthcare professionals interested in applying ML to resuscitation
which computer systems have the ability to think and learn like research but who are not experts in the field. We reviewed the rele-
humans and to automatically perform tasks that humans would nor- vant literatures searched as described in the Supplementary file to
mally perform such as cognition driven decision-making.1 ML is used introduce prediction models, natural language processing (including
to develop algorithms and models that can learn from and make pre- large language models, LLM), consideration of treatment hetero-
dictions or recommend decisions based on large datasets.1 In resus- geneity, and optimization of medical practice and resource manage-
citation medicine, AI and ML hold the potential to revolutionize ment by reinforcement learning. We also discuss the limitations and
patient care by providing decision support and optimizing treatment challenges of implementing AI and ML tools in actual clinical settings.

* Corresponding author at: Health Services and Systems Research, Duke-NUS Medical School, National University of Singapore, Singapore.
E-mail addresses: [email protected] (Y. Okada), [email protected] (M. Mertens), [email protected] (N. Liu),
[email protected] (S.S.W. Lam), [email protected] (M.E.H. Ong).
https://doi.org/10.1016/j.resplu.2023.100435
Received 19 June 2023; Received in revised form 9 July 2023; Accepted 14 July 2023

2666-5204/Ó 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.
org/licenses/by-nc-nd/4.0/).
2 R E S U S C I T A T I O N P L U S 15 (2023) 100435

We aim to facilitate discussion on the potential for further research ture, and laboratory data, to identify patterns that may suggest a
and enhance communication between clinicians, resuscitation patient’s condition is deteriorating. As a result, EWS can alert health-
researchers and AI and ML experts. care providers to intervene before a cardiac arrest occurs.14 Further,
these predictions are also valuable to estimate demand for bed
capacity and to appropriately allocate medical resources.15 Some
Prediction models of these ML models are implemented in electronic medical record
systems or as applications on tablets or smartphones, which auto-
The most common use of ML is predictive modeling.1 Prediction matically input the data into the model and output the calculated
models (also known as supervised learning) are commonly used to results, improving user availability and accessibility.14,15
predict a patient’s diagnosis or outcomes based on clinical data.
For example, ML models can be helpful to diagnose, estimate the Prediction models using unstructured data
severity in triage, and understand the risk of complications in Images and bio-signals (EEG, and ECG)
decision-making for surgery, which can allow us to develop more ML has been increasingly utilized in resuscitation research to
appropriate treatment plans and potentially improve patient progno- enhance diagnostic and prognostic accuracy in unstructured data
sis in a more objective manner.2–4 This type of prediction model such as various imaging modalities, including CT scans, EEG, and
may also be applied to adjust for severity when considering the qual- ECG. For example, there are some researchers developing ML mod-
ity of care and assuming the counterfactual scenario (such as if a els to predict neurological outcomes using head CT images16–18 and
certain treatment was not performed with the resulting outcome) EEG,19–21 potentially leading to more accurate and timely diagnoses.
when discussing causal inference.5,6 Similarly, ML models have been employed to analyze ECG data,
Prediction models may incorporate a wide array of data including enabling the prediction of critical events such as in-hospital cardiac
structured data such as demographic information, clinical variables, arrest, ventricular arrhythmia, sudden cardiac death, and the suc-
biomarkers, and blood test results, and also unstructured data such cess of defibrillation during resuscitation.22–26 These applications
as images and bio-signals like electrocardiograms (EEG) and elec- of ML models using medical imaging and bio-signals are expected
troencephalograms (EEG), to predict outcomes (Fig. 1). We intro- to contribute to facilitating early detection, improving predictive accu-
duce some examples of research on prediction models based on racy, and ultimately enhancing more appropriate resuscitation, emer-
the type of data. gency, or intensive care.

Prediction models using structured data

Structured data is one of the most common sources of data for ML Exploring sub-phenotypes and treatment
models in resuscitation research.7,8 This type of data is typically pre- heterogeneity
sented in a tabular format with clear rows and columns, representing
patients and their respective features or attributes. These may ML is also used to explore sub-phenotypes, an emerging concept in
include demographic information, medical history, vital signs, labora- precision medicine. (Fig. 2) Sub-phenotypes are distinct subgroups
tory test results, and more. For out-of-hospital cardiac arrest (OHCA) within a disease or condition characterized by different clinical fea-
research, the Utstein format is established worldwide as a standard- tures such as disease progression, outcomes, and underlying biolog-
ized data format. This enables us to easily develop ML models ical mechanisms.27,28 Whereas phenotypes represent categories of
applied to the data.9,10 One of the primary uses of ML with tabular patients with common features such as a specific syndrome, e.g.,
data in resuscitation research are predictive models to estimate the sepsis or acute respiratory distress syndrome,27,28 sub-phenotype
likelihood of outcomes such as return of spontaneous circulation is particularly relevant when discussing subgroups with heterogene-
(ROSC), survival, or neurological recovery after cardiac arrest.7,8,11 ity on treatment effect.27 Heterogeneity on treatment effect refers to
In another example, tabular data was also used to develop early the variation in how different individuals or groups respond to the
warning systems (EWS) that predict the risk of cardiac arrest or other same treatment.5 It means that not all patients respond to treatments
serious adverse events among patients admitted to hospital.12–14 in the same way due to various factors such as genetic differences,
These systems use ML models to analyze various data such as heart lifestyle factors, pre-existing health conditions, and more.5 Under-
rate, blood pressure, respiratory rate, oxygen saturation, tempera- standing the concept of sub-phenotypes and the complexities of
treatment effect heterogeneity are anticipated to advance the devel-
opment of personalized medicine, moving beyond the conventional
’one-size-fits-all’ treatment approach. For example, some research
in the resuscitation context suggests the hypothesis that sub-
phenotypes exhibit heterogeneity of effect of targeted temperature
management, such as some subgroups may have the potential ben-
efit of hypothermia (e.g., at 33 ), while others may not.29,30 For
exploring such treatment heterogeneity, ML such as “clustering” or
“causal machine learning” are utilized in some research.31,32

Fig. 1 – The concept of prediction models applied to Clustering

predict mortality A prediction model is one type of ML Clustering is a type of unsupervised machine learning that can be
developed to predict the outcome. Various patterns of used to identify subgroups who share similar clinical characteristics
clinical information can be utilized to develop and explore treatment heterogeneity or novel association between
prediction models. the subgroups and events, using data such as patients’ characteris-
 R E S U S C I T A T I O N P L U S 15 (2023) 100435 3

Fig. 2 – The concept of clustering and sub-phenotypes Phenotypes (e.g., sepsis, acute respiratory distress
syndrome) are categorized by clustering to sub-phenotypes with different clinical features and the heterogeneous
response to the treatment.

tics, biomarker values, and genomic data (Fig. 2).31,33 One of the targeted to specific genetic features,47,48 there will be an increasing
strengths of clustering is its ability to manage data complexity and number of studies on treatment heterogeneity and pharmacoge-
discover hidden patterns, making the data easier to understand nomics in the resuscitation field.
and visualize. Previously, this clustering analysis was used in
research exploring novel sub-phenotypes among patients with vari-
ous patterns in emergency medicine and critical care such as sepsis, Reinforcement learning to optimize treatment
ARDS, trauma, and cardiac arrest.27,28,34–39
For instance, various clinical patterns in coagulopathy among Reinforcement learning is a type of machine learning that autono-
patients with severe head trauma are associated with different out- mously chooses actions to maximize rewards obtained from the
comes.38 There are also subgroups among OHCA patients with dif- given environment. The system learns through trial and error to
ferent clinical outcomes when treated with ECPR.39 Some research select actions that lead to the highest possible reward. Reinforce-
suggests the effect of early goal-directed treatment or the effect of ment learning has broad applications and is particularly useful for
drugs on coagulopathy are different among subgroups in sepsis.36,37 complex tasks, such as games, autonomous driving, robotics, and
This technique is also utilized to summarize the risk factors as a sub- logistics.49 For example, in 2015, AlphaGo, an AI developed using
group. One example is the subgroups with environmental features reinforcement learning, famously defeated a world champion Go
characterized by environmental parameters such as temperature, player.50
wind speed, and air pollution are suggested to be associated with In the field of medicine and healthcare, reinforcement learning
the occurrence of acute myocardial infarction or acute ischemic has potential applications in optimizing treatment strategies.
stroke.40,41 (Fig. 4) For example, one notable example of using reinforcement
learning, in the context of intensive care, is the development of an
Causal machine learning “AI Clinician“ for sepsis treatment in managing fluids and vasopres-
Causal machine learning is an ML approach to investigate causal sors.51 This AI system analyzed two ICU databases and learned opti-
inference, which is particularly valuable in assessing heterogeneity mal treatment strategies by examining numerous treatment
in treatment effects (Fig. 3).5,42,43 Causal forest, one approach within decisions to maximize the expected survival outcome. As a result,
causal machine learning based on the random forest, works by split- this AI model could select the optimal treatment strategy which
ting the data into different subgroups and assessing the treatment showed the lowest mortality rates. Another model using reinforce-
effect within each subgroup by handling the no-linear and/or hi- ment learning suggested personalized optimization of mechanical
dimensional data.5,42,43 For example in critical care fields, the causal ventilation in patients staying at cardiovascular ICUs.52 In other
forest was used on data from an RCT about the effect of using a bou- examples, some reinforcement learning programs were suggested
gie during intubation.44 This RCT found that using a bougie did not to investigate the optimal dose of sedative agents in general anes-
increase the incidence of successful intubation on the first attempt thesia.53 Although there are few published research using reinforce-
in all critically ill adults; however, the causal forest analysis sug- ment learning in the resuscitation field, it has potential for future
gested some individuals who had the potential benefit of using a studies.
bougie.
The application of machine learning using genetic and molecular
data (omics data) to treatment heterogeneity and precision medicine Natural language processing
is also expected to result in a more personalized approach to health-
care such as investigating the heterogeneity of the treatment Natural language processing (NLP) is a subset of ML technology that
response or adverse events of drugs among patients with certain enables computers to analyze the language that humans usually use
genetic features.45,46 Although this type of research is mainly in daily life. This technology is prevalent in our modern lives with
focused on the oncology field because the drugs are commonly applications using voice recognition such as voice assistant
4 R E S U S C I T A T I O N P L U S 15 (2023) 100435

Fig. 3 – The concept of treatment heterogeneity (Left) Assuming that the difference between outcomes when
treatment is performed and when it is not, is the same in each patient: treatment effect is homogenous between
individual patients. (Right) Assuming that the difference between outcomes when treatment is performed and when
it is not, is different in each patient: treatment effect is heterogenous between individual patients.

Fig. 4 – The concept of reinforcement learning in medical research. Patient status is changed to a different status by
the action, and consequently, the reward is obtained based on the status. Reinforcement learning can find the best
strategy to maximize the rewards based on many trials.

programs like Apple’s Siri or Google Assistant and using text like to enable faster and more accurate deployment of emergency med-
chatbots or language translation tools. ical services, which can improve patient outcomes.
In the field of research in resuscitation, NLP models are being uti- NLP technology can also be utilized to analyze clinical data from
lized in innovative ways. One notable example of using voice data is the free text in medical records such as medical history or physical
ML programs to help recognize cardiac arrest and support initiating findings.59 Algorithms can be developed to predict emergency condi-
bystander-CPR during emergency calls to the dispatch center tions such as in-hospital cardiac arrest or give decision support on
(Fig. 5).54,55 These programs can analyze the caller’s words during the appropriate disposition of patients at the emergency depart-
an emergency call and estimate the probability of the patient being ment.59–63 This technique can also be used to accurately predict
in cardiac arrest. This kind of program has also been applied in neurological outcomes such as a modified Rankin scale by analyzing
research to detect other emergencies such as severe trauma after free text data in clinical notes.64 Additionally, chatbot tools using NLP
road trauma and stroke.56,57 Additionally, NLP voice recognition have also been developed in the resuscitation research fields. One
technology offers practical benefits for paramedics in the field. Para- example is a preliminary chatbot to guide users on how to perform
medics can use voice commands to create prehospital records bystander CPR.62 In summary, NLP-applied research using voice
thereby reducing the need for manual data entry and enabling them or text is increasing and they can analyze communication or medical
to focus more on patient care.58 These programs have the potential records to predict events and be a guide to action in resuscitation.
 R E S U S C I T A T I O N P L U S 15 (2023) 100435 5

Fig. 5 – Example of Natural Language Processing for Activating Bystander CPR NLP: Natural Language Processing,
CPR: Cardiopulmonary resuscitation In the emergency call dispatch center, the application utilizes natural
language processing (NLP) to analyze the caller’s words, aiding the dispatcher in identifying potential cases of
cardiac arrest.54

Large language model (LLM) is one domain of research in NLP harmful. A prediction model may simply be biased because of the
fields that can understand and generate natural language used by original data it is trained on, reflecting the existing bias as is. For
humans. Typically, by learning patterns from large amounts of textual example, an AI model may reflect historical disparities in healthcare
data, these models can generate answers to new questions, or pro- access and outcomes, and inadvertently perpetuate these biases by
duce text to accomplish specific tasks such as translation or revising recommending differential treatment based on factors such as race,
the text. Recently, the GPT-3 and GPT-4 developed by OpenAI have gender, age, or socioeconomic status.75,76 It is therefore essential
attracted a lot of attention for their wide adaptability and flexibility.65 If that the training data is diverse and representative of the patient pop-
you enter the prompt “What should we do if we encounter a patient ulation. However, in the actual scene of resuscitation, obtaining com-
who has suddenly collapsed?” into the application, the application prehensive and diverse datasets can be challenging. Clinical
can provide plausible answers as if they are provided by a healthcare situations can change drastically in a short time, making it difficult
professional. (However, it should be noted that these answers may to comprehensively collect data in a timely manner, such as in a
be incorrect.) One representative example of using LLM is that the resource-limited environment like the prehospital setting or a
LLM can pass the medical licensing examination without any addi- crowded emergency department.77,78 Furthermore, in many settings
tional training data.66,67 Further, some research indicated that LLM of resuscitation, clinical data is still being recorded using paper and
can provide quality and empathetic responses to patient ques- pen, and some backend data entry process is needed to integrate
tions.68,69 Further, the LLM is also expected to summarize the clinical the data into electronic medical records for it to be utilized for ML
information from medical records like a professional or perform the application.79 Yet, ensuring the availability of comprehensive and
systematic review instead of humans.70,71 Although research in the representative data is crucial to develop accurate and generalizable
resuscitation field has not yet been published, it is expected to models.
develop in the future. In contrast, this LLM has also caused various
controversial issues, such as the accuracy, validity, and responsibil-
ity of the generated sentences and ethical issues that may arise Validation process to verify the reproducibility
(more detail is discussed in the next paragraph).65 Although several Once ML models have been developed, they should be reproduca-
concerns, LLM has great potential to improve the burden on health- ble.80 Previously, it has been reported that many prediction models
care providers, especially in terms of decision-making, documenta- have a high risk of bias, especially due to the lack of the validation
tion, and summarizing medical information. process to confirm the reproducibility of the models using different
datasets.80–83 One of the problems to validate the ML and AI models
using different datasets is the difficulty in obtaining different data
Challenges for AI and ML in resuscitation from the original data with consistent format and definition of the vari-
research and implementation ables. In the resuscitation fields, the Utstein format is broadly
accepted as a universal data-collecting standard mainly in pre-
Despite the extensive research conducted, actual implementation of hospital settings; however, some of the in-hospital data have not
AI and ML in the clinical setting remains limited, though some prac- yet been standardized (e.g., some variables in the emergency
tices have implemented AI and ML-based algorithms in resuscitation department or intensive care unit have still not been strictly
and intensive care.14,15,54,72,73 Widespread adoption may be slow defined).10
due to several concerns and limitations.74 Here we give an overview Another concern is inappropriate reporting of the originally devel-
of the most important challenges and barriers that prevent proper oped models.83 Reproducibility can be difficult to ascertain as details
implementation. of the models are not reported.83 Furthermore, validation study risks
selective reporting bias, meaning that validation studies reporting
Data quality and availability models with poor performance are less likely to get published.81
AI and ML algorithms heavily depend on the quality of data they are Yet, ensuring robustness in AI and ML models, including their relia-
trained on. If the data is unreliable, missing, incomplete, or biased, bility and reproducibility, is essential to prevent or minimize unin-
the model’s predictions or performance can be inaccurate or even tended harm.
6 R E S U S C I T A T I O N P L U S 15 (2023) 100435

Generalizability and clinical integration essence, past mistakes lead to new self-fulfilling prophecies, rein-
Verifying the Generalizability is also essential to validate the AI and forcing predictions that generate inappropriate clinical judgments,
ML models prior to clinical application. Again, ML models depend on creating a vicious cycle; an automated feedback loop of self-
data, and if the model too strongly fits certain features of the data fulfilling poor outcomes for future cardiac arrest patients.88 Further-
(“overfitting”), the results may not be generalizable to the different more, the lack of error signals due to confirmative outcomes com-
population without those features. Resuscitation practices vary bined with the lack of interpretability of ML models greatly hinders
across different healthcare settings, geopolitical contexts, and clinicians from recognizing such biased predictive feedback loops.
patient populations.84–87 AI models developed in one context may Catching false positives retrospectively is near to impossible, since
not generalize well when traveling to other settings. Ensuring the this would require counterfactual data. Clinical guidelines suggest
generalizability and applicability the models to diverse populations, the need for a multi-modal approach to predict the outcome of car-
different clinical protocols and resource-constrained environments diac arrest patients to minimize the potential harm of false-positive
is essential for their widespread application.87 of predictions.92 When advanced AI models are developed, clinicians
Additionally, other practical barriers exist to implementing AI and must remain aware of the risk for amplified bias through self-fulfilling
ML in clinical settings. It includes not only regulatory approval but prophecies and feedback loops.
also integration into clinical workflows. Moreover, the adoption of
ML models necessitates clear benefits in routine clinical practice, Transparency, Interpretability, and trust
such as improving patients’ outcomes and reducing workload or A key challenge when applying AI and ML to the actual resuscitation
costs. However, few randomized controlled trials (RCTs) have scene is the interpretability of and trust in ML models.80–82 ML mod-
shown the actual benefit of ML models in clinical settings.42,80 If inte- els are often described as a ’black box’ due to the complexity of the
gration of ML models into general clinical workflows does not yield models that generate the results. This lack of transparency can hin-
clear benefits for clinicians, patients, or other stakeholders, no one der clinicians’ or patients’ trust towards ML models. One example is,
would use these models. The actual benefit of ML tools in clinical set- as mentioned above, an ML model was developed to detect potential
tings compared to existing clinical workflows need to be demon- cardiac arrest cases using the voice data of emergency calls at the
strated in research before widespread adoption will follow. dispatch center.54 The retrospective observational study using the
voice recordings indicated that the ML model outperformed human
Self-fulfilling prophecies and feedback loops dispatchers.77 However, the RCT comparing the dispatcher assisted
Another important issue to be focused on in the resuscitation field is by the ML model to those without such assistance, did not demon-
the risk of hidden false positive bias by self-fulfilling prophecy and strate any improvement in the performance to recognize the cardiac
feedback loop when predicting the prognosis of cardiac arrest arrest cases.54 One of the potential mechanisms of this result sug-
patients.88,89 A self-fulfilling prophecy is a prediction that influences gested by the research team was that the dispatcher could not
people’s beliefs and behavior through which the prediction is then understand the ML model’s decision-making process and the dis-
realized.90 In resuscitation, if clinicians expect that a particular patcher possibly did not trust the alert from the ML program.93 Had
patient may not survive despite the best treatment, the expectation the advice come from human experts instead of the ML model, the
could influence their decision to forego further treatment, allowing dispatchers might have asked the rationale why and how they con-
the patient’s death, thereby fulfilling the initial prediction (self- cluded, considered accepting (or rejecting) their suggestion, and
fulfilling prophecy). This becomes especially problematic if the initial thereby improved their performance to recognize the cardiac arrest
prediction was incorrect (a false-positive), which could result in the case. As such, achieving interpretability and trust in ML models
patient not receiving the potentially beneficial care. While these may be essential to successfully implement AI and ML into real-
issues have existed even before AI and ML are developed (because world clinical practice.
predictions of clinicians are sometimes inaccurate),91 there is grow-
ing concern that AI and ML might amplify the bias due to self-fulfilling Regulatory and legal challenges
feedback loops (Fig. 6). If a model trained on biased data is applied While proper data collection and management is an essential prereq-
to guide clinical decision-making, and the new data influenced by the uisite for developing and applying ML models to clinical settings,
model’s results are then used as input data again to “improve” the such data collection and management must of course respect pri-
model, there is a risk that the initial biases will be reinforced and vacy and comply with the law.94 Furthermore, liability and responsi-
amplified. To illustrate, if a prediction model is developed using data bility frameworks need to be developed and implemented for AI-
from a hospital where resuscitation efforts were consistently termi- driven and ML-based resuscitation interventions, in order to ensure
nated early for OHCA patients aged over 70 years old during a speci- accountability and patient safety. As seen in this article, AI and ML
fic period due to temporary limitation of resources (such as limitation can raise several ethical concerns when it is applied to the actual
of intensive care during the COVID-19 pandemic), the model may medical system and care, although the ethical concerns far exceed
inevitably predict the lower probability of survival for similar patients the ones we mention here. Generally speaking, the Ethics Guideline
than is accurate. This prediction merely reflects the flawed input data for Trustworthy AI suggested seven key requirements including
itself rather than the truth under ideal circumstances. Yet, if clinicians human agency and oversight, technical robustness and safety, pri-
perceive this prediction as “accurate” and terminate resuscitation vacy and data governance, transparency, diversity, non-
efforts based on such false positives, no one will notice the missed discrimination and fairness, environmental and societal well-being,
opportunities for successful resuscitation of OHCA patients over and, accountability.94 While we have selected several significant
70 years, since the outcome confirms the prediction.88 If new models issues particular to resuscitation, these ethical principles should be
are then trained based on the confirmed biased data, it can further addressed across all AI applications in medicine, regardless of the
amplify the biased prediction and inappropriate withdrawal rates. In specialty. Indeed, many non-profit institutions, regulatory, and gov-
 R E S U S C I T A T I O N P L U S 15 (2023) 100435 7

Fig. 6 – Concept of self-fulfilling prophecy and its feedback loop. A patient who could be saved is mistakenly
assessed, due to a false positive, as having a “Very low possibility to survive”. Such a prognosis can inform the
decision to withdraw treatment. As a result, the initial prediction “Very low possibility to survive” is self-realized,
thereby showing as a true positive. If this faulty and biased data is utilized to develop or improve the ML models, it
reproduces and amplifies the false positive predictions. This leads to further harm in that more viable patients lose
the opportunity to be treated. If this new data gets used in its turn to further develop the model, it leads to a vicious
cycle of harm.

ernmental bodies across the world are currently collaborating to Consent for publication
ensure (inter)national laws that better protect citizens from the
rapidly increasing impacts of AI and ML-driven models. Not applicable.

Conclusion Availability of data and materials

In this article, we introduce and illustrate important concepts within AI Not applicable.
and ML research in the resuscitation field. The application of AI and
ML in resuscitation research holds significant potential to revolution-
ize the field by improving prediction, supporting decision-making, and Funding
developing personalized treatment strategies. However, various lim-
itations and ethical concerns must be addressed to ensure the This study was supported by a scientific research grant from the
responsible and effective implementation of these technologies in JSPS KAKENHI of Japan (JP22K21143) and the Zoll foundation.
actual clinical practice. As more high-quality data becomes available, YO has received an overseas scholarship from the Japan Society
it is expected that AI-driven models and ML-based algorithms will for the Promotion of Science, the FUKUDA Foundation for medical
play an increasingly important role in resuscitation research and technology, and the International medical research foundation. MM
practice. Moving forward, it will be essential for researchers, com- is funded by the European Union, through the HORIZON-MSCA-
puter scientists, clinicians, ethicists, policymakers, and other stake- 2022-PF-01-01 Marie Curie Postdoctoral Fellowship, Project
holders to work together to overcome the challenges and harness 101107292 ‘PredicGenX’.
the full potential of AI and ML in resuscitation, ultimately leading to
better patient outcomes and more efficient healthcare systems.
CRediT authorship contribution statement

Ethical approval Yohei Okada: Conceptualization, Writing – original draft. Mayli Mer-
tens: Conceptualization, Writing – original draft. Nan Liu: Writing –
review & editing. Sean Shao Wei Lam: Writing – review & editing.
Not applicable.
Marcus Eng Hock Ong: Writing – review & editing.
8 R E S U S C I T A T I O N P L U S 15 (2023) 100435

Declaration of Competing Interest learning: A retrospective study. EClinicalMedicine 2022;48:101422.

https://doi.org/10.1016/j.eclinm.2022.101422.
9. Jacobs I, Nadkarni V, Bahr J, et al. Cardiac arrest and
YO has received a research grant from the ZOLL Foundation and
cardiopulmonary resuscitation outcome reports: update and
overseas scholarships from the Japan Society for Promotion of
simplification of the Utstein templates for resuscitation registries: a
Science, the FUKUDA Foundation for medical technology, and the statement for healthcare professionals from a task force of the
International medical research foundation. These organizations have International Liaison Committee on Resuscitation (American Heart
no role in conducting this research. MEHO reports grants from the Association, European Resuscitation Council, Australian
Laerdal Foundation, Laerdal Medical, and Ramsey Social Justice Resuscitation Council, New Zealand Resuscitation Council, Heart
Foundation for funding of the Pan-Asian Resuscitation Outcomes and Stroke Foundation of Canada, InterAmerican Heart Foundation,
Resuscitation Councils of Southern Africa). Circulation 2004.
Study an advisory relationship with Global Healthcare SG, a com-
10. Perkins GD, Jacobs IG, Nadkarni VM, et al. Cardiac Arrest and
mercial entity that manufactures cooling devices; and funding from Cardiopulmonary Resuscitation Outcome Reports: Update of the
Laerdal Medical on an observation program to their Community Utstein Resuscitation Registry Templates for Out-of-Hospital Cardiac
CPR Training Centre Research Program in Norway. MEHO is a Sci- Arrest. Circulation 2015;132:1286–300. https://doi.org/10.1161/
entific Advisor to TIIM Healthcare SG and Global Healthcare SG. CIR.0000000000000144.
11. Nanayakkara S, Fogarty S, Tremeer M, et al. Characterising risk of
in-hospital mortality following cardiac arrest using machine learning:
Appendix A. Supplementary data A retrospective international registry study. PLoS Med 2018;15:
e1002709. https://doi.org/10.1371/journal.pmed.1002709.
12. Pimentel MAF, Redfern OC, Malycha J, et al. Detecting Deteriorating
Supplementary data to this article can be found online at https://doi. Patients in the Hospital: Development and Validation of a Novel
org/10.1016/j.resplu.2023.100435. Scoring System. Am J Respir Crit Care Med 2021;204:44–52. https://
doi.org/10.1164/rccm.202007-2700OC.
13. Bartkowiak B, Snyder AM, Benjamin A, et al. Validating the
Author details Electronic Cardiac Arrest Risk Triage (eCART) Score for Risk
Stratification of Surgical Inpatients in the Postoperative Setting:
a
Duke-NUS Medical School, National University of Singapore, Retrospective Cohort Study. Ann Surg 2019;269:1059–63.
14. Winslow CJ, Edelson DP, Churpek MM, et al. The Impact of a
Singapore bPreventive Services, Graduate School of Medicine, Kyoto
Machine Learning Early Warning Score on Hospital Mortality: A
University, Kyoto, Japan cAntwerp Center for Responsible AI, Antwerp
d
Multicenter Clinical Intervention Trial. Crit Care Med
University, Belgium Centre for Ethics, Department of Philosophy, 2022;50:1339–47. https://doi.org/10.1097/ccm.0000000000005492.
Antwerp University, Belgium eDepartment of Emergency Medicine, 15. Goldstein BA, Cerullo M, Krishnamoorthy V, et al. Development
Singapore General Hospital and Performance of a Clinical Decision Support Tool to Inform
Resource Utilization for Elective Operations. JAMA Netw Open
2020;3:e2023547. https://doi.org/10.1001/jamanetworkopen.2020.
23547.
R E F E R E N C E S
16. Kawai Y, Kogeichi Y, Yamamoto K, Miyazaki K, Asai H, Fukushima
H. Explainable artificial intelligence-based prediction of poor
neurological outcome from head computed tomography in the
1. Kühl N, Schemmer M, Goutier M, Satzger G. Artificial intelligence immediate post-resuscitation phase. Sci Rep 2023;13:5759. https://
and machine learning. Electronic Markets 2022;32:2235–44. https:// doi.org/10.1038/s41598-023-32899-5.
doi.org/10.1007/s12525-022-00598-0. 17. Mansour A, Fuhrman JD, Ammar FE, et al. Machine Learning for
2. Goto T, Camargo Jr CA, Faridi MK, Freishtat RJ, Hasegawa K. Early Detection of Hypoxic-Ischemic Brain Injury After Cardiac
Machine learning-based prediction of clinical outcomes for children Arrest. Neurocrit Care 2022;36:974–82.
during emergency department triage. JAMA Netw Open 2019;2: 18. Elmer J, Liu C, Pease M, et al. Deep learning of early brain imaging
e186937. https://doi.org/10.1001/jamanetworkopen.2018.6937. to predict post-arrest electroencephalography. Resuscitation
3. Okada Y, Matsuyama T, Morita S, et al. Machine learning-based 2022;172:17–23. https://doi.org/10.1016/j.resuscitation.2022.01.004.
prediction models for accidental hypothermia patients. J Intensive 19. Zheng WL, Amorim E, Jing J, et al. Predicting neurological outcome
Care 2021;9:6. in comatose patients after cardiac arrest with multiscale deep neural
4. Bihorac A, Ozrazgat-Baslanti T, Ebadi A, et al. MySurgeryRisk: networks. Resuscitation 2021;169:86–94.
development and validation of a machine-learning risk algorithm for 20. Zheng WL, Amorim E, Jing J, et al. Predicting Neurological Outcome
major complications and death after surgery. Ann Surg From Electroencephalogram Dynamics in Comatose Patients After
2019;269:652. Cardiac Arrest With Deep Learning. IEEE Trans Biomed Eng
5. Gong X, Hu M, Basu M, Zhao L. Heterogeneous treatment effect 2022;69:1813–25.
analysis based on machine-learning methodology. CPT 21. Jonas S, Rossetti AO, Oddo M, Jenni S, Favaro P, Zubler F. EEG-
Pharmacometrics Syst Pharmacol 2021;10:1433–43. https://doi.org/ based outcome prediction after cardiac arrest with convolutional
10.1002/psp4.12715. neural networks: Performance and visualization of discriminative
6. Riascos A, Romero M, Serna N. Risk adjustment revisited using features. Hum Brain Mapp 2019;40:4606–17. https://doi.org/
machine learning techniques. Documento Cede 2017. 10.1002/hbm.24724.
7. Nishioka N, Kobayashi D, Kiguchi T, et al. Development and 22. Hajeb-M S, Cascella A, Valentine M, Chon KH. Deep Neural Network
validation of early prediction for neurological outcome at 90 days Approach for Continuous ECG-Based Automated External
after return of spontaneous circulation in out-of-hospital cardiac Defibrillator Shock Advisory System During Cardiopulmonary
arrest. Resuscitation 2021;168:142–50. https://doi.org/10.1016/j. Resuscitation. J Am Heart Assoc 2021;10:e019065.
resuscitation.2021.09.027. 23. Kwon J-M, Kim K-H, Jeon K-H, Lee SY, Park J, Oh B-H. Artificial
8. Liu N, Liu M, Chen X, et al. Development and validation of an intelligence algorithm for predicting cardiac arrest using
interpretable prehospital return of spontaneous circulation (P-ROSC) electrocardiography. Scand J Trauma, Resus Emergency Med
score for patients with out-of-hospital cardiac arrest using machine 2020;28:98. https://doi.org/10.1186/s13049-020-00791-0.
 R E S U S C I T A T I O N P L U S 15 (2023) 100435 9

24. Kolk MZH, Deb B, Ruipérez-Campillo S, et al. Machine learning of 42. Athey S, Wager S. Estimating treatment effects with causal forests:
electrophysiological signals for the prediction of ventricular an application. Observational Studies 2019;5:37–51.
arrhythmias: systematic review and examination of heterogeneity 43. Wager S, Athey S. Estimation and inference of heterogeneous
between studies. eBioMedicine 2023:89. https://doi.org/10.1016/j. treatment effects using random forests. J Am Stat Assoc
ebiom.2023.104462. 2018;113:1228–42.
25. Sem M, Mastrangelo E, Lightfoot D, Aves T, Lin S, Mohindra R. The 44. Seitz KP, Spicer AB, Casey JD, et al. Individualized Treatment
ability of machine learning algorithms to predict defibrillation success Effects of Bougie vs Stylet for Tracheal Intubation in Critical Illness.
during cardiac arrest: a systematic review. Resuscitation 2023;185. Am J Respir Crit Care Med 2023. https://doi.org/10.1164/
https://doi.org/10.1016/j.resuscitation.2023.109755. rccm.202209-1799OC.
26. Kenet AL, Pemmaraju R, Ghate S, et al. A pilot study to predict 45. Syrowatka A, Song W, Amato MG, et al. Key use cases for artificial
cardiac arrest in the pediatric intensive care unit. Resuscitation intelligence to reduce the frequency of adverse drug events: a
2023;185:109740. https://doi.org/10.1016/j. scoping review. Lancet Digital Health 2022;4:e137–48. https://doi.
resuscitation.2023.109740. org/10.1016/S2589-7500(21)00229-6.
27. Reddy K, Sinha P, O’Kane CM, Gordon AC, Calfee CS, McAuley DF. 46. Kline A, Wang H, Li Y, et al. Multimodal machine learning in precision
Subphenotypes in critical care: translation into clinical practice. health: a scoping review. npj Digital Med 2022;5:171. https://doi.org/
Lancet Respir Med 2020;8:631–43. https://doi.org/10.1016/s2213- 10.1038/s41746-022-00712-8.
2600(20)30124-7. 47. Azuaje F. Artificial intelligence for precision oncology: beyond patient
28. Wildi K, Livingstone S, Palmieri C, LiBassi G, Suen J, Fraser J. The stratification. npj Prec Oncol 2019;3:6. https://doi.org/10.1038/
discovery of biological subphenotypes in ARDS: a novel approach to s41698-019-0078-1.
targeted medicine? J Intensive Care 2021;9:14. https://doi.org/ 48. Lauschke VM, Ingelman-Sundberg M. Emerging strategies to bridge
10.1186/s40560-021-00528-w. the gap between pharmacogenomic research and its clinical
29. Callaway CW, Coppler PJ, Faro J, et al. Association of Initial Illness implementation. npj Genom Med 2020;5:9. https://doi.org/10.1038/
Severity and Outcomes After Cardiac Arrest With Targeted s41525-020-0119-2.
Temperature Management at 36 °C or 33 °C. JAMA Netw Open 49. Liu S, See KC, Ngiam KY, Celi LA, Sun X, Feng M. Reinforcement
2020;3:e208215–e. https://doi.org/ learning for clinical decision support in critical care: comprehensive
10.1001/jamanetworkopen.2020.8215. review. J Med Internet Res 2020;22:e18477.
30. Nishikimi M, Ogura T, Nishida K, et al. Outcome Related to Level of 50. Silver D, Huang A, Maddison CJ, et al. Mastering the game of Go
Targeted Temperature Management in Postcardiac Arrest Syndrome with deep neural networks and tree search. Nature 2016;529:484–9.
of Low, Moderate, and High Severities: A Nationwide Multicenter https://doi.org/10.1038/nature16961.
Prospective Registry. Crit Care Med 2021;8:e741–50. 51. Komorowski M, Celi LA, Badawi O, Gordon AC, Faisal AA. The
31. Loftus TJ, Shickel B, Balch JA, et al. Phenotype clustering in health Artificial Intelligence Clinician learns optimal treatment strategies for
care: A narrative review for clinicians. Front Artif Intell sepsis in intensive care. Nat Med 2018;24:1716–20. https://doi.org/
2022;5:842306. https://doi.org/10.3389/frai.2022.842306. 10.1038/s41591-018-0213-5.
32. Jawadekar N, Kezios K, Odden MC, et al. Practical Guide to Honest 52. Peine A, Hallawa A, Bickenbach J, et al. Development and validation
Causal Forests for Identifying Heterogeneous Treatment Effects. Am of a reinforcement learning algorithm to dynamically optimize
J Epidemiol 2023. Published by Oxford University Press on behalf of mechanical ventilation in critical care. NPJ Digit Med 2021;4:32.
the Johns Hopkins Bloomberg School of Public Health For https://doi.org/10.1038/s41746-021-00388-6.
permissions, please e-mail: [email protected].; 2023. 53. Yun WJ, Shin M, Jung S, Ko J, Lee HC, Kim J. Deep reinforcement
33. Sinha P, Calfee CS, Delucchi KL. Practitioner’s Guide to Latent learning-based propofol infusion control for anesthesia: A feasibility
Class Analysis: Methodological Considerations and Common Pitfalls. study with a 3000-subject dataset. Comput Biol Med
Crit Care Med 2021;49:e63–79. https://doi.org/10.1097/ 2023;156:106739. https://doi.org/10.1016/
CCM.0000000000004710. j.compbiomed.2023.106739.
34. Okada Y, Komukai S, Kitamura T, et al. Clustering out-of-hospital 54. Blomberg SN, Christensen HC, Lippert F, et al. Effect of Machine
cardiac arrest patients with non-shockable rhythm by machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac
learning latent class analysis. Acute Med Surg 2022;9:e760. Arrest During Calls to Emergency Medical Services: A Randomized
35. Wilson JG, Subphenotypes CCSA. Understanding a Heterogeneous Clinical Trial. JAMA Netw Open 2021;4:e2032320–e. https://doi.org/
Syndrome. Crit Care 2020;24:102. https://doi.org/10.1186/s13054- 10.1001/jamanetworkopen.2020.32320.
020-2778-x. 55. Byrsell F, Claesson A, Ringh M, et al. Machine learning can support
36. Seymour CW, Kennedy JN, Wang S, et al. Derivation, validation, and dispatchers to better and faster recognize out-of-hospital cardiac
potential treatment implications of novel clinical phenotypes for arrest during emergency calls: A retrospective study. Resuscitation
sepsis. JAMA 2019;321:2003–17. 2021;162:218–26. https://doi.org/10.1016/j.
37. Kudo D, Goto T, Uchimido R, et al. Coagulation phenotypes in sepsis resuscitation.2021.02.041.
and effects of recombinant human thrombomodulin: an analysis of 56. Chin KC, Cheng YC, Sun JT, et al. Machine Learning-Based Text
three multicentre observational studies. Crit Care 2021;25:114. Analysis to Predict Severely Injured Patients in Emergency Medical
38. Fujiwara G, Okada Y, Shiomi N, Sakakibara T, Yamaki T, Hashimoto Dispatch: Model Development and Validation. J Med Internet Res
N. Derivation of Coagulation Phenotypes and the Association with 2022;24. https://doi.org/10.2196/30210.
Prognosis in Traumatic Brain Injury: A Cluster Analysis of Nationwide 57. Scholz ML, Collatz-Christensen H, Blomberg SNF, Boebel S,
Multicenter Study. Neurocrit Care 2023. Verhoeven J, Krafft T. Artificial intelligence in Emergency
39. Okada Y, Komukai S, Kitamura T, et al. Clinical Phenotyping of Out- Medical Services dispatching: assessing the potential impact of an
of-Hospital Cardiac Arrest Patients With Shockable Rhythm - automatic speech recognition software on stroke detection taking the
Machine Learning-Based Unsupervised Cluster Analysis. Circ J Capital Region of Denmark as case in point. Scand J Trauma
2022;86:668–76. https://doi.org/10.1253/circj.CJ-21-0675. Resusc Emerg Med 2022;30:36. https://doi.org/10.1186/s13049-
40. Koo GPY, Zheng H, Pek PP, et al. Clustering of Environmental 022-01020-6.
Parameters and the Risk of Acute Myocardial Infarction. Int J Environ 58. Fukaguchi K, Goto T, Yamamoto T, Yamagami H. Experimental
Res Public Health 2022;19. https://doi.org/10.3390/ijerph19148476. Implementation of NSER Mobile App for Efficient Real-Time Sharing
41. Koo GPY, Zheng H, Aik JCL, et al. Clustering of Environmental of Prehospital Patient Information With Emergency Departments:
Parameters and the Risk of Acute Ischaemic Stroke. Int J Environ Interrupted Time-Series Analysis. JMIR Formative Research 2022;6:
Res Public Health 2023:20. https://doi.org/10.3390/ijerph20064979. e37301.
10 R E S U S C I T A T I O N P L U S 15 (2023) 100435

59. Goto T, Hara K, Hashimoto K, et al. Validation of chief complaints, Access in Algorithms, Mechanisms, and Optimization. Association
medical history, medications, and physician diagnoses structured for Computing Machinery 2021. Article 17.
with an integrated emergency department information system in 77. Frisch A, Reynolds JC, Condle J, Gruen D, Callaway CW.
Japan: the Next Stage ER system. Acute Med Surg Jan-Dec 2020;7: Documentation discrepancies of time-dependent critical events in out
e554. of hospital cardiac arrest. Resuscitation 2014;85:1111–4. https://doi.
60. Sterling NW, Patzer RE, Di M, Schrager JD. Prediction of emergency org/10.1016/j.resuscitation.2014.05.002.
department patient disposition based on natural language processing 78. Sundermann ML, Salcido DD, Koller AC, Menegazzi JJ. Inaccuracy
of triage notes. Int J Med Inf 2019,;129:184–8. https://doi.org/ of patient care reports for identification of critical resuscitation events
10.1016/j.ijmedinf.2019.06.008. during out-of-hospital cardiac arrest. Am J Emerg Med
61. Brown JR, Ricket IM, Reeves RM, et al. Information Extraction From 2015;33:95–9.
Electronic Health Records to Predict Readmission Following Acute 79. Hani M, Christine N, Gerard O, et al. Emergency care surveillance
Myocardial Infarction: Does Natural Language Processing Using and emergency care registries in low-income and middle-income
Clinical Notes Improve Prediction of Readmission? Journal of the countries: conceptual challenges and future directions for research.
American Heart Association 2022;;11:e024198. BMJ Glob Health 2019;4:e001442.
62. Ivanov O, Wolf L, Brecher D, et al. Improving ED Emergency 80. Volovici V, Syn NL, Ercole A, Zhao JJ, Liu N. Steps to avoid overuse
Severity Index Acuity Assignment Using Machine Learning and and misuse of machine learning in clinical research. Nat Med
Clinical Natural Language Processing. J Emerg Nurs 2022,;28:1996–9. https://doi.org/10.1038/s41591-022-01961-6.
2021,;47:265–278.e7. https://doi.org/10.1016/j.jen.2020.11.001. 81. Andaur Navarro CL, Damen JAA, Takada T, et al. Risk of bias in
63. Okada Y, Okada A, Ito H, Sonoo T, Goto T. External validation of the studies on prediction models developed using supervised machine
POP score for predicting obstetric and gynecological diseases in the learning techniques: systematic review. BMJ 2021;375:n2281.
emergency department. The. Am J Emerg Med 2022,;51:348–53. https://doi.org/10.1136/bmj.n2281.
https://doi.org/10.1016/j.ajem.2021.11.022. 82. Ramspek CL, Jager KJ, Dekker FW, Zoccali C, van Diepen M.
64. Fernandes MB, Valizadeh N, Alabsi HS, et al. Classification of External validation of prognostic models: what, why, how, when and
neurologic outcomes from medical notes using natural language where? Clin Kidney J 2021;14:49–58. https://doi.org/10.1093/ckj/
processing. Expert Syst Appl 2023:214. https://doi.org/10.1016/j. sfaa188.
eswa.2022.119171. 83. Yang C, Kors JA, Ioannou S, et al. Trends in the conduct and
65. Haupt CE, Marks M. AI-Generated Medical Advice—GPT and reporting of clinical prediction model development and validation: a
Beyond. JAMA 2023;329:1349–50. https://doi.org/ systematic review. J Am Med Inform Assoc 2022;29:983–9. https://
10.1001/jama.2023.5321. doi.org/10.1093/jamia/ocac002.
66. Kung TH, Cheatham M, Medenilla A, et al. Performance of ChatGPT 84. Ong MEH, Do Shin S, De Souza NNA, et al. Outcomes for out-of-
on USMLE: Potential for AI-assisted medical education using large hospital cardiac arrests across 7 countries in Asia: The Pan Asian
language models. PLoS digital health 2023;2:e0000198. Resuscitation Outcomes Study (PAROS). Resuscitation
67. Kasai J, Kasai Y, Sakaguchi K, Yamada Y, Radev D. Evaluating gpt- 2015;96:100–8.
4 and chatgpt on japanese medical licensing examinations. arXiv 85. Nichol G, Thomas E, Callaway CW, et al. Regional variation in out-
2023. preprint arXiv:230318027. of-hospital cardiac arrest incidence and outcome. JAMA
68. Ayers JW, Poliak A, Dredze M, et al. Comparing Physician and 2008;300:1423–31. https://doi.org/10.1001/jama.300.12.1423.
Artificial Intelligence Chatbot Responses to Patient Questions 86. Tagami T, Tanaka H, Shin SD, et al. Impact of population aging on
Posted to a Public Social Media Forum. JAMA Intern Med 2023. the presentation of out-of-hospital cardiac arrest in the Pan Asian
https://doi.org/10.1001/jamainternmed.2023.1838. Resuscitation Outcomes Study. Acute Med Surg 2020;7:e430.
69. Sarraju A, Bruemmer D, Van Iterson E, Cho L, Rodriguez F, Laffin L. https://doi.org/10.1002/ams2.430.
Appropriateness of Cardiovascular Disease Prevention 87. Van Calster B, Steyerberg EW, Wynants L, van Smeden M. There is
Recommendations Obtained From a Popular Online Chat-Based no such thing as a validated prediction model. BMC Med 2023;21:70.
Artificial Intelligence Model. JAMA 2023;329:842–4. https://doi.org/ https://doi.org/10.1186/s12916-023-02779-w.
10.1001/jama.2023.1044. 88. Mertens M, King OC, van Putten M, Boenink M. Can we learn from
70. Patel SB, Lam K. ChatGPT: the future of discharge summaries? hidden mistakes? Self-fulfilling prophecy and responsible
Lancet Digital Health 2023;5:e107–8. https://doi.org/10.1016/S2589- neuroprognostic innovation. J Med Ethics 2022;48:922–8. https://doi.
7500(23)00021-3. org/10.1136/medethics-2020-106636.
71. Qureshi R, Shaughnessy D, Gill KAR, Robinson KA, Li T, Agai E. Are 89. De-Arteaga M, Elmer J. Self-fulfilling prophecies and machine
ChatGPT and large language models “the answer” to bringing us learning in resuscitation science. Resuscitation 2023;183:109622.
closer to systematic review automation? Syst Rev 2023;12:72. https://doi.org/10.1016/j.resuscitation.2022.10.014.
https://doi.org/10.1186/s13643-023-02243-z. 90. King OC, Mertens M. Self-fulfilling Prophecy in Practical and
72. Pham SD, Keijzer HM, Ruijter BJ, et al. Outcome Prediction of Automated Prediction. Ethical Theory Moral Pract 2023;26:127–52.
Postanoxic Coma: A Comparison of Automated https://doi.org/10.1007/s10677-022-10359-9.
Electroencephalography Analysis Methods. Neurocrit Care 91. Detsky ME, Harhay MO, Bayard DF, et al. Discriminative Accuracy of
2022;37:248–58. Physician and Nurse Predictions for Survival and Functional
73. Aellen FM, Alnes SL, Loosli F, et al. Auditory stimulation and deep Outcomes 6 Months After an ICU Admission. JAMA
learning predict awakening from coma after cardiac arrest. Brain 2017;317:2187–95. https://doi.org/10.1001/jama.2017.4078.
2023. 92. Nolan JP, Sandroni C, Böttiger BW, et al. European Resuscitation
74. Chan SL, Lee JW, Ong MEH, et al. Implementation of Prediction Council and European Society of Intensive Care Medicine guidelines
Models in the Emergency Department from an Implementation 2021: post-resuscitation care. Intensive Care Med 2021;47:369–421.
Science Perspective-Determinants, Outcomes, and Real-World https://doi.org/10.1007/s00134-021-06368-4.
Impact: A Scoping Review. Ann Emerg Med 2023;82:22–36. https:// 93. Zicari RV, Brusseau J, Blomberg SN, et al. On assessing trustworthy
doi.org/10.1016/j.annemergmed.2023.02.001. Epub 2023 Mar 14. AI in healthcare. Machine learning as a supportive tool to recognize
75. Di Nucci E, Lee J-Y, Wagner IA. The Rowman & Littlefield Handbook cardiac arrest in emergency calls. Front Hum Dynam 2021;3:673104.
of Bioethics. Rowman & Littlefield; 2022. 94. Floridi L. Establishing the rules for building trustworthy AI. Nat Mach
76. Suresh H, Guttag J. A Framework for Understanding Sources of Intell 2019;1:261–2. https://doi.org/10.1038/s42256-019-0055-y.
Harm throughout the Machine Learning Life Cycle. Equity and