Discussion

The journey of prehospital emergency care for a patient, including call receipt, dispatch, travel, triage, diagnosis, treatment, and transport to appropriate definitive care, is efficiently managed by multiple operational and medical components of EMS. This review showed that ML algorithms were implemented in almost all EMS components, including dispatch and communication system, first responders, ambulances and paramedics, medical oversight, and quality improvement. In addition, ML has the potential to augment the decision-making capabilities of EMS personnel in both clinical and operational domains or even supplement their judgment in specific aspects of EMS components. However, the integration of ML in EMS is still evolving, and further research, development, and validation may be needed to fully leverage its capabilities in clinical and operational applications. Considerable heterogeneity was observed in the included studies in terms of population characteristics, the ML algorithms employed, input features, and performance metrics which limited a more comprehensive review. The implications and performance of ML should be in-depth evaluated for each clinical condition, outcome, or operational task.

Several studies showed the potential of ML algorithms in improving emergency medical response to OHCA. More than 15 studies concentrated on implementing ML for classifying cardiac arrest ECG rhythms or detecting the return of spontaneous circulation aiming to improve resuscitation performance and provide timely guidance for delivering electrical defibrillation. Multiple ML algorithms exhibit outstanding performance (AUC > 0.90), demonstrating remarkable accuracy and practical applicability in the real-world context. For example, Jaureguibeitia, et al proposed novel deep learning architectures (AUC = 98.6%) for shock decisions using ECG segments as brief as one second, which likely shorten interruptions in resuscitation for rhythm analysis. ^{Reference Jaureguibeitia, Zubia and Irusta13} Elola, et al also implemented deep neural networks for pulse detection using only the ECG (AUC = 86.2%), ^{Reference Elola, Aramendi and Irusta14} while Alonso, et al (AUC = 92.6%) combined ECG and thoracic impedance signals for superior pulse detection. ^{Reference Alonso, Irusta and Aramendi15} Further, deep neural networks have demonstrated the ability to detect OHCA at the time of calling for help, arguably with greater accuracy and speed than the EMS telecommunicators. ^{Reference Blomberg, Christensen and Lippert6,Reference Blomberg, Folke and Ersbøll16,Reference Byrsell, Claesson and Ringh17}

Predictive ML models were also developed for three clinical outcomes after OHCA, including the return of spontaneous circulation, survival, and neurological recovery. Among many ML models, deep neural networks achieved the highest performance (AUC = 0.84 to 0.90) in predicting the success of defibrillation using ECG features. ^{Reference He, Lu and Zhang18–Reference Shandilya, Kurz and Ward20} Acknowledging the limited number of cases in most of these studies (n <500 cases), larger and more diverse samples of cases are likely essential to limit the risk of bias and to enhance the performance, reliability, and applicability of ML algorithms in predicting defibrillation success. ^{Reference Yang21}

Deep neural networks and ensemble ML algorithms performed exceptionally (AUC = 0.90 to 0.96) among others to predict survival and neurological recovery using prehospital clinical and EMS data. ^{Reference Cheng, Chiu and Zeng22–Reference Kawai, Okuda and Kinoshita24} However, these predictive models exhibit considerable heterogeneity in quality, type of population, input and output features, and algorithm selection, potentially affecting generalizability. Additionally, the complexity and lack of transparency in certain ML algorithms might obstruct their integration into clinical practice. ^{Reference Yang21} Future ML studies should discuss features attribution or explain the ML models output by using several methods such as Shapley Additive Explanation, Local Interpretable Model-agnostic Explanations, and integrated gradients.

Most ML models have also been shown to improve the detection of CVDs in the prehospital setting. A deep learning model achieved an impressive AUC score of 0.997 in detecting ST-elevation myocardial infarction (STEMI). ^{Reference Chen, Wang and Liu25} This model has been deployed in a prehospital 12-lead ECG device and resulted in a shorter median time between first medical contact and hospital arrival (18.5 minutes), compared to the weighted mean estimated by a previous meta-analysis (41 minutes). ^{Reference Alrawashdeh, Nehme and Williams26} Further, Bouzid, et al have developed a fusion model to detect non-STEMI using prehospital 12-lead ECG which significantly outperformed the performance of expert clinicians and commercial software. ^{Reference Bouzid, Faramand and Gregg27} In addition, prehospital stroke diagnostic ML algorithms have been shown to be accurate in identifying and differentiating stroke using prehospital clinical presentation variables. ^{Reference Chen, Zhang and Xu28–Reference Uchida, Kouno and Yoshimura31} Two studies have shown that neural network algorithms out-performed other available prehospital stroke prediction scales, such as the Prehospital Acute Stroke Severity scale and the Cincinnati Prehospital Stroke Severity scale. ^{Reference Chen, Zhang and Xu28,Reference Zhang, Zhou and Zhang29} However, a larger number of parameters (ie, 18 variables) were included in these neural network models, which demand more effort and time to be collected by ambulance clinicians.

For traumatic conditions, ML models have been increasingly applied to primarily improve the triage decision by dispatchers, first responders, and ambulance clinicians in various out-of-hospital contexts (ie, general trauma, road accidents, and military). These models were developed either to classify traumatic patients by their level of severity (ie, Injury Severity Score) ^{Reference Candefjord, Sheikh Muhammad and Bangalore32} or to predict the level of treatment needed (ie, advanced life-saving interventions, trauma team activation, and critical care) ^{Reference Morris, Karam and Zolfaghari33,Reference Kang, Cho and Kwon34} or the clinical outcomes (ie, survival). ^{Reference Kim, You and So35} The input features in most models were relatively limited and mainly consisted of demographics, vital signs, other clinical manifestations, or accident-related variables. Neural networks and ensemble techniques (random forest, XGBoosting) were frequently used and had the best performance compared to other models. Two studies found that binary logistic regression performed better than ML algorithms when classifying the severity of injuries. ^{Reference Candefjord, Sheikh Muhammad and Bangalore32,Reference Larsson, Berg and Gellerfors36} This is likely attributed to the limited number of input features in these studies. Overall, the contribution of ML seems to be promising in this context, but further validation is required to better understand the applicability of ML in supporting triage decisions and optimizing resource allocation.

Several ML models have been implemented to solve operational problems and augment the decision making of EMS personnel in operational tasks. The over-arching aims of the operational studies were to reduce ambulance travel time and deploy the appropriate level of prehospital care to the right patients. A variety of supervised, unsupervised, and reinforcement models have been used to improve EMS demand forecasting in different settings to optimize ambulance allocation strategies, and eventually reduce response times. The characteristics of ML models forecasting EMS demand varied in terms of time horizons (hourly, daily, weekly, and monthly demand) and input features (geospatial, temporal, metrological, and EMS data). Several studies aimed to facilitate ambulance passage through traffic congestion by either audio-based and/or visual-based detection. The audio-based models using random forest, support vector machine, and neural networks achieved AUC scores between 0.97-1.00. ^{Reference Dhakal, Damacharla and Javaid37–Reference Tran and Tsai40} Visual-based models mainly used deep learning with variable performance (AUC = 0.78-1.00) ^{Reference Tran and Tsai40–Reference Roy and Rahman42} while hybrid models incorporated both audio and image inputs and typically relied on deep learning yield strong results (0.86-0.98). ^{Reference Tran and Tsai40,Reference Raman, Kaushik and Rao43} The performance of these models indicates the viability of employing ML to detect ambulances and expedite their passage through intersections and traffic congestion.

Finally, the implementation of the developed ML models in real time was rarely discussed in the included studies. Several challenges such as the generalizability and explainability of these ML models may complicate their implication in clinical and real time. Several studies reported perfect discrimination, especially for operational purposes, which warrant a critical examination of the underlying data and methodologies employed. While perfect discrimination can be attributed to factors like over-fitting, reliance on internal validation, data leakage, or imbalanced datasets, it is imperative to consider the broader context of its applicability, generalizability, and potential limitations in real-world EMS scenarios. For example, the generalization of ML algorithms for ambulance detection using audio or image data may pose significant challenges due to the wide variation in ambulance sirens and morphologies across different EMS systems.

It is also important to acknowledge the complications and limitations associated with the retrospective nature of most included studies. Rich prehospital datasets are still necessary; however, they are likely challenging to obtain, owing to limitations in data quality and availability, patient privacy, and the heterogeneity of EMS system data frameworks. ^{Reference Aggarwal, Sounderajah and Martin44} This necessitates collaborative efforts among researchers, EMS systems, and other health care institutions, as well as policymakers to establish data-sharing frameworks, best practices, and governance models that facilitate the responsible and impactful application of these technologies in the challenging and dynamic landscape of EMS. For better explainability, future ML studies should also discuss features attribution or explain the output of ML models by using available methods such as Shapley additive explanation, local interpretable model-agnostic explanations, and integrated gradients. ^{Reference Petch, Di and Nelson45}