AI in Resuscitation: Smarter Help When Every Second Counts

Marziehsadat Rezaei	Fernando Acosta-Perez
University of Louisville	University of Wisconsin-Madison

A Race Against Time: Imagine a situation where a person suddenly collapses and their heart stops beating. In those first few minutes, everything depends on speed: recognizing what is happening, starting cardiopulmonary resuscitation (CPR), using a defibrillator if needed, and making the right decisions under intense pressure. This is the reality of cardiac arrest, one of the most serious medical emergencies in the world. Despite major advances in emergency medicine, survival rates after cardiac arrest are still low (Perkins et al., 2021). The reason is simple: resuscitation is not just difficult, it is a race against time.

AI Steps Into Resuscitation

When a person suffers a cardiac arrest, survival depends on decisions that need to be made in seconds. Medical personnel or emergency responders should recognize the event quickly to dispatch resources and begin care coordination. Although cardiac arrest constitutes a medical emergency, in practice, it is a system challenge at which detection, communication, and execution must go hand in hand to maximize patient outcomes.

Today, artificial intelligence (AI) is opening a new and exciting chapter in this field. AI is already being used in many areas of medicine to find patterns in large amounts of data and support better decisions (Haug and Drazen, 2023). In resuscitation, the primary form of AI is machine learning (ML), which comprises a set of models that can learn from data to make predictions under uncertainty. Its promise is especially powerful because the challenge is not only to treat cardiac arrest, but also to recognize danger earlier, act faster during the event, and guide care afterward. In other words, ML may help at every stage of the resuscitation process: before, during, and after.

One reason this area is so interesting is that it is not limited to one method or one setting. For instance, a cardiac arrests can occur either in hospital or out of hospital. In hospital cardiac arrests benefit from continuous monitoring and immediate access to care. In contrast, out of hospital cardiac arrest remains a logistical challenge, with slower detection and constrained resources significantly affecting patient outcomes. These challenges highlight that improving outcomes requires not only better prediction, but also better coordination and decision-making across complex systems.

To go from predictions to actions, is where operations research (OR) can make an important contribution. If ML helps answer questions such as “Is this patient deteriorating?” or “Is this emergency call likely to involve cardiac arrest?”, OR helps answer the next question: “What is the best action to take?”

In this article, we examine how can ML and OR support resuscitation across the three different stages of cardiac arrest (i.e., before, during, and after), and discuss the critical role of designing systems capable of leveraging the support from these tools and translating it into better patient outcomes.

From Prediction to Decision Making

Before a cardiac arrest even happens, ML can be used to identify patients who are at high risk. In hospitals, ML models can analyze clinical data such as vital signs, lab results, and monitoring information to detect signs of deterioration (Ueno et al., 2020). This kind of early warning system could alert care teams before a patient reaches the point of collapse. Preventing a cardiac arrest, of course, is even better than responding to one.

Out of hospital, preparing for cardiac arrest events includes location and coverage planning. In this setting, survival often depends on whether an ambulance, AED, or nearby responder can reach the patient fast enough. OR models can help determine where to place these resources in order to maximize coverage, reduce response time, and improve access. This is typically modeled as a location-allocation problem, with common formulations including maximal covering, set covering, and p-median models. For example, optimization studies of public AED placement have reported better OHCA coverage under optimized deployment than under existing placement strategies (Chan et al., 2018). These are not simple placement problems. Cardiac arrests are rare, highly time-sensitive, and geographically uneven, so planners must often balance efficiency with fairness, especially across urban and rural areas. During the arrest itself, ML can support some of the hardest real-time decisions. Researchers have developed tools that can classify heart rhythms, detect whether a patient has a pulse, evaluate CPR quality, and even help determine the best timing for defibrillation (Zace et al., 2025). Some systems use sensors or video-based deep learning models to measure chest compression depth and rate, offering feedback without interrupting CPR (da Costa et al., 2025). In emergency dispatch centers, natural language processing has also shown promise by helping identify cardiac arrest from emergency calls faster than traditional methods, allowing CPR instructions to begin sooner (Byrsell et al., 2021).

During an emergency in the out of hospital setting, emergency systems must quickly decide which resources to send, in what order, and along which routes. Should the closest ambulance be dispatched, or should it be preserved for another nearby high-risk zone? Should a drone carrying an AED be sent alongside EMS? Should fire, police, or trained volunteers also be activated? These are classic OR decision problems involving routing, real-time assignment, and dynamic resource coordination under uncertainty. Related work by Gao et al. (Gao et al., 2024) showed that intelligent real-time UAV–ambulance dispatch policies can outperform simpler benchmark strategies in time-critical EMS settings, especially when capacity is limited.

After the return of spontaneous circulation (ROSC), when the patient’s heart starts beating again, ML models can answer prognosis questions: will the patient survive, and what kind of neurological recovery is likely? Researchers have used clinical records, EEG signals, imaging, and registry data to build models that predict survival and brain outcomes (Pham et al., 2022). These predictions could help clinicians make more informed decisions and improve communication with families during a very critical period.

Similarly, in the out of hospital setting, after an event emergency responders face the problem of redeployment after a call. When one ambulance or response team is sent to a cardiac arrest, the system immediately becomes less prepared for the next emergency. OR can help determine how to reposition the remaining resources to maintain system readiness. For example, Essus et al. (Essus et al., 2024) proposed a real-time optimization framework for EMS relocation and dispatching and found that explicitly accounting for workload balance leads to more effective redeployment decisions. This is especially important in cardiac arrest because even small delays can substantially affect survival. A good emergency system, therefore, must not only respond well to the current patient, but also stay prepared for the next one.

So, what have researchers found so far? Overall, the results are encouraging. Many ML models have shown strong performance in predicting outcomes, detecting deterioration, and improving CPR monitoring. Deep learning models that combine multiple data types sometimes report especially high accuracy, while ML methods such as XGBoost and random forests perform well across many different applications. OR studies have also reported promising results, particularly for improving coverage, dispatch efficiency, and redeployment decisions. At the same time, the field is still developing. Many studies are retrospective, use different outcome definitions, and are difficult to compare directly. In the OR literature as well, results are often highly dependent on local geography, demand patterns, and system design, which can limit generalizability across hospitals or EMS systems. A model or optimization policy that looks impressive in one dataset or system may not work equally well in another hospital, EMS network, or country.

These research findings lead to the most important practical question: how can ML and OR help in real systems? The answer is not that they will replace doctors, nurses, paramedics, or dispatchers. Instead, they can act as smart support tools. ML can warn clinicians earlier, identify high-risk patients, or recognize cardiac arrest during emergency calls. OR can then help determine how to allocate and dispatch resources, improve CPR-related logistics, and support better coordination across the chain of survival.

In a setting where every second matters, even a small improvement in speed or decision quality could translate into saved lives. For example, a more accurate prediction model is valuable, but it becomes far more powerful when linked with an OR model that determines where ambulances should wait, which responder should be activated, or whether an AED drone should be launched. In that sense, the real promise lies not in prediction alone, but in connecting prediction to decision making.

Challenges on the Road Ahead

Although there is substantial evidence highlighting the power of ML in predicting risk across different resuscitation tasks, only a few studies have evaluated the potential of these systems beyond their predictive accuracy (Blomberg et al. (2019, 2021)). For example, in the context of out of hospital cardiac arrest, results from randomized control trials suggest that although ML can help in detecting the cardiac arrest earlier than emergency medical dispatcher by analyzing voice recordings, this did not necessarily translate into improved dispatching decisions (Blomberg et al. (2021)). This highlight the nuanced challenges behind implementing these models in practice.

This limited impact may be explained by several factors: lack of trust from medical dispatchers, limited integration with existing workflows, or the fact that current protocols were not designed to incorporate ML-based decision support. More broadly, this is where many open OR problems remain. It is not enough to develop accurate predictive models if emergency systems do not know how to use them operationally.

To this end, several open challenges are especially important. First, many existing studies stop at prediction and do not evaluate whether the predictions improve real decisions or patient outcomes. Second, uncertainty remains a major issue: demand varies over time and space, travel times can change suddenly, and not all responders have the same availability or skill level. Third, fairness matters. If models are trained mostly on data from certain hospitals or regions, they may not perform equally well everywhere, and optimized systems may unintentionally favor some populations over others.

Another open challenge is human–AI interaction. Even the best decision-support model will have limited impact if clinicians, dispatchers, or EMS leaders do not trust it or cannot easily incorporate it into practice. Future systems must therefore be not only accurate, but also transparent, usable, and aligned with real protocols.

To overcome these challenges, ML and OR systems must be tested in real clinical environments, not only in retrospective studies or simulations. Future progress will depend on stronger validation, more diverse datasets, better reporting standards, and close collaboration between clinicians, engineers, and policymakers (Zace et al., 2025). It will also depend on developing integrated frameworks that combine prediction, optimization, and implementation, rather than treating them as separate tasks.

Conclusions

AI in resuscitation is still evolving, but its potential is hard to ignore. From predicting cardiac arrest before it happens, to guiding CPR during the event, to supporting recovery decisions afterward, AI offers a vision of resuscitation that is faster, smarter, and more personalized. At the same time, the field should not stop at prediction alone. Some of the most important unanswered questions are operational: where to place resources, how to dispatch them, how to manage uncertainty, and how to ensure fairness and readiness across the entire emergency response system.

This is why OR has an important role to play. It helps move the field from knowing more to doing better. In the end, the goal is not to make medicine less human. It is to give healthcare professionals better tools, and better decisions, when human lives are on the line.

References

Blomberg, S.N., Christensen, H.C., Lippert, F., Ersbøll, A.K., Torp-Petersen, C., Sayre, M.R., Kudenchuk, P.J., Folke, F., 2021. Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial. JAMA Network Open 4, e2032320. URL: https://jamanetwork.com/journals/jamanetworkopen/ fullarticle/2774644, doi:10.1001/jamanetworkopen.2020.32320.

Blomberg, S.N., Folke, F., Ersbøll, A.K., Christensen, H.C., Torp-Pedersen, C., Sayre, M.R., Counts, C.R., Lippert, F.K., 2019. Machine learning as a supportive tool to recognize cardiac arrest in emergency calls. Resuscitation 138, 322–329. URL: https://linkinghub. elsevier.com/retrieve/pii/S0300957218309754, doi:10.1016/j.resuscitation.2019.01.015.

Byrsell, F., Claesson, A., Ringh, M., Svensson, L., Jonsson, M., Nordberg, P., Forsberg, S., Hollenberg, J., Nord, A., 2021. Machine learning can support dispatchers to better and faster recognize out-of-hospital cardiac arrest during emergency calls: a retrospective study. Resuscitation 162, 218–226. doi:10.1016/j.resuscitation.2021.02.041.

Chan, T.C., Shen, Z.J.M., Siddiq, A., 2018. Robust defibrillator deployment under cardiac arrest location uncertainty via row-and-column generation. Operations Research 66, 358–379. doi:doi.org/10.1287/opre.2017.1660.

da Costa, L.M.A., Scheidewind, R.O., Peron, R.F., Timmerman, T., de Cerqueira Lario, F., Moreira, F.R., Nunes, R.A.B., Scudeler, T.L., 2025. Effectiveness of a smartwatch-based feedback system in improving cardiopulmonary resuscitation quality: a simulation study. Resuscitation Plus , 101114. doi:10.1016/j.resplu.2025.101114.

Essus, Y., De La Fuente, R., Venkitasubramanian, A., 2024. Real-time optimization for relocation and dispatching of emergency medical services with balanced workload and outsourced ride-hailing services. Computers & Industrial Engineering 187, 109823. doi:10.1016/j.cie.2023.109780.

Gao, X., Kong, N., Griffin, P., 2024. Shortening emergency medical response time with joint operations of uncrewed aerial vehicles with ambulances. Manufacturing & Service Operations Management 26, 447–464. doi:10.1287/msom.2022.0166.

Haug, C.J., Drazen, J.M., 2023. Artificial intelligence and machine learning in clinical medicine, 2023. New England Journal of Medicine 388, 1201–1208. doi:10.1056/NEJMra2302038.

Perkins, G.D., Gräsner, J.T., Semeraro, F., Olasveengen, T., Soar, J., Lott, C., Van de Voorde, P., Madar, J., Zideman, D., Mentzelopoulos, S., et al., 2021. European resuscitation council guidelines 2021: executive summary. Resuscitation 161, 1–60. doi:10.1016/j.resuscitation.2021.02.003.

Pham, S.D., Keijzer, H.M., Ruijter, B.J., Seeber, A.A., Scholten, E., Drost, G., van den Bergh, W.M., Kornips, F.H., Foudraine, N.A., Beishuizen, A., et al., 2022. Outcome prediction of postanoxic coma: a comparison of automated electroencephalography analysis methods. Neurocritical care 37, 248–258. doi:10.1007/s12028-022-01449-8.

Ueno, R., Xu, L., Uegami, W., Matsui, H., Okui, J., Hayashi, H., Miyajima, T., Hayashi, Y., Pilcher, D., Jones, D., 2020. Value of laboratory results in addition to vital signs in a machine learning algorithm to predict in-hospital cardiac arrest: A single-center retrospective cohort study. PloS one 15, e0235835. doi:10.1371/journal.pone.0235835.

Zace, D., Semeraro, F., Schnaubelt, S., Montomoli, J., Ristagno, G., Fijačko, N., Gamberini, L., Bignami, E.G., Greif, R., Monsieurs, K.G., et al., 2025. Artificial intelligence in resuscitation: a scoping review. Resuscitation Plus 24, 100973. doi:10.1016/j. resplu.2025.100973.

Acknowledgements: We would like to thank Ronak Tiwari for taking time to review this article. Photo credit goes to assistance of ChatGPT for the header and footer photos