Real-Time Ambulance Detection on Resource-Constrained Edge Devices Using YOLOv5 and ONNX Runtime

Background and Purpose: There are no contemporary national-level data on Emergency Medical Services (EMS) response times for suspected stroke in the United States (US). Because effective stroke treatment is time-dependent, we characterized response times for suspected stroke, and examined whether they met guideline recommendations. Methods: Using the National EMS Information System dataset, we included 911 calls for patients ≥ 18 years with an EMS provider impression of stroke. We examined variation in the total EMS response time by dispatch notification of stroke, age, sex, race, region, time of day, day of the week, as well as the proportion of EMS responses that met guideline recommended response times. Total EMS response time included call center dispatch time (receipt of call by dispatch to EMS being notified), EMS dispatch time (dispatch informing EMS to EMS starts moving), time to scene (EMS starts moving to EMS arrival on scene), time on scene (EMS arrival on scene to EMS leaving scene), and transport time (EMS leaving scene to reaching treatment facility). Results: We identified 184,179 events with primary impressions of stroke (mean age 70.4 ± 16.4 years, 55% male). Median total EMS response time was 36 (IQR 28.7–48.0) minutes. Longer response times were observed for patients aged 65–74 years, of white race, females, and from non-urban areas. Dispatch identification of stroke versus “other” was associated with marginally faster response times (36.0 versus 36.7 minutes, p < 0.01). When compared to recommended guidelines, 78% of EMS responses met dispatch delay of <1 minute, 72% met time to scene of <8 minutes, and 46% met on-scene time of <15 minutes. Conclusions: In the United States, time from receipt of 9-1-1 calls to treatment center arrival takes a median of 36 minutes for stroke patients, an improvement upon previously published times. The fact that 22%–46% of EMS responses did not meet stroke guidelines highlights an opportunity for improvement. Future studies should examine EMS diagnostic accuracy nationally or regionally using outcomes based approaches, as accurate recognition of prehospital strokes is vital in order to improve response times, adhere to guidelines, and ultimately provide timely and effective stroke treatment.

Background: Chest pain is a leading complaint in emergency settings. Timely emergency medical services (EMS) responses can reduce delays to treatment and improve clinical outcomes for acute myocardial infarction patients and other medical emergencies. We investigated national-level EMS response, scene, and transport times for acute chest pain patients in the United States. Methods: A retrospective analysis was performed using 2015–2016 data from the National EMS Information System (NEMSIS). Eligible patients were identified as having a provider impression of chest pain or discomfort and not due to trauma or resulting in cardiac arrest during EMS care. Descriptive analyses of prehospital time intervals and patient-, response-, and system-level covariates were performed. Multivariable logistic regression was used to measure associations between meeting response and scene time benchmarks (8-min and 15-min, respectively) and covariates. Results: Our study identified 1,672,893 eligible EMS encounters of chest pain. Patients had a mean age of 63.1 years (SD = 14.8). The population was evenly distributed by sex (51% male; 49% female). Most encounters occurred in home or residence (58%) and had lights and sirens response to scene (84%). Most encounters were in urban areas (78%). The median (interquartile range, IQR) response time was 7 (5–10) minutes. The median (IQR) scene time was 16 (12–20) minutes. The median (IQR) transport time was 13 (8–20) minutes. Generally, median response and transport times were longer in rural and frontier areas compared to urban and suburban areas. Only 65% and 49% met the 8-min response and 15-min scene time benchmarks. Responding with lights and sirens was associated with greater compliance with EMS response time benchmark. EMS care of older age groups and females was less likely to meet the scene time benchmark. Conclusions: Substantial proportions of EMS encounters for chest pain did not meet response and scene time benchmarks. Regional and urban-rural differences were observed in adherence with the response time benchmark. Our findings also suggest age and gender disparities in on-scene delays by EMS. Our study contributes important evidence on timely EMS responses for cardiac chest pain and provides suggestions for EMS system benchmarking and quality improvement.

To describe the use of the Automatic External Defibrillation (AED) device in an urban, two-tiered Emergency Medical Service (EMS) response setting with regard to its potential effects on cardiac arrest patient survival and neurologic outcome. A retrospective and descriptive design was utilized to study all cardiac arrest patients that had resuscitations attempted in the prehospital environment over a 30-month period. The study took place in a two-tiered EMS system serving an urban population of 368,383 persons. The first tier of EMS response is provided by the City Fire Department, which is equipped with a standard AED device. All first-tier personnel are trained to the level of Emergency Medical Technician-Basic. The second tier of EMS response is provided by personnel from one of two ambulance services. All second-tier personnel are trained to the level of Emergency Medical Technician-Paramedic. 271 cardiac arrest patients were identified for inclusion. One-hundred nine of these patients (40.2%) had an initial rhythm of either ventricular fibrillation or pulseless ventricular tachycardia and were shocked using the AED upon the arrival of first-tier personnel. Forty-two patients (38.5%) in this group had a return of spontaneous circulation in the field and 22 (20.2%) survived to hospital discharge. Of the survivors, 17 (77.3%) had moderate to good neurologic function at discharge based on the Glasgow-Pittsburgh Cerebral Performance Categories. Faster response times by the first-tier personnel appeared to correlate with better neurologic outcomes. First responder-based AED usage on patients in ventricular fibrillation or pulseless ventricular tachycardia can be applied successfully in an urban setting utilizing a two-tiered EMS response. In this study, a 20.2% survival to hospital discharge rate was obtained. Seventy-seven percent of these survivors had a moderate to good neurologic outcome based on the Glasgow-Pittsburgh Cerebral Performance Categories.

BackgroundThe national incidence and characteristics of out‐of‐hospital cardiac arrest in the United States is unclear. We sought to describe the national characteristics of adult out‐of‐hospital cardiac arrest reported in the National Emergency Medical Services Information System (NEMSIS).MethodsWe used 2016 NEMSIS data, consisting of most emergency medical services (EMS) responses from 46 states and territories. We limited the analysis to adult (age ≥18 years) emergency “9‐1‐1” events. We defined out‐of‐hospital cardiac arrest as: (1) patient condition reported as cardiac arrest, (2) EMS reported attempted resuscitation of cardiac arrest, (3) EMS performance of cardiopulmonary resuscitation (CPR), or (4) EMS performance of defibrillation. We determined the incidence of adult out‐of‐hospital cardiac arrest among EMS responses. We also determined patient demographics (age, sex, race, ethnicity, location, US census region, and urbanicity), response characteristics (dispatch complaint and elapsed time) and clinical interventions (medications and procedures) of adult out‐of‐hospital cardiac arrest. We analyzed the data using descriptive techniques, calculating binomial proportions with exact 95% confidence intervals (CI).ResultsAmong 18,679,873 adult 9‐1‐1 responses, there were 224,992 with patient condition cardiac arrest, 344,274 with EMS‐reported attempted cardiac arrest resuscitation, 149,775 with EMS performance of CPR, and 185,388 cases with EMS performance of defibrillation, resulting in a total of 574,824 out‐of‐hospital cardiac arrest (incidence 30.8 per 1000 EMS 9‐1‐1 responses, 95% CI = 30.69–30.85). Among identified out‐of‐hospital cardiac arrest responses, most involved patients who were older (mean = 62.4 ± 20.1 years). Most out‐of‐hospital cardiac arrest occurred at home (58.8%), in the South census region (65.4%), and in urban settings (79.8%). The most commonly reported medications used in out‐of‐hospital cardiac arrest were: epinephrine (22.5%), amiodarone (2.9%), sodium bicarbonate (6.2%), glucose (3.0%), and naloxone (5.1%). Commonly reported procedures included CPR (26.1%), orotracheal intubation (14.2%), bag‐valve‐mask ventilation (10.1%), manual defibrillation (29.3%) and automated external defibrillation (5.6%). Out‐of‐hospital cardiac arrest EMS treatment times were: elapsed response time (median = 7 minutes [interquartile range (IQR) = 5–10]), scene time (median = 17 minutes [IQR = 12–25]), and elapsed transport time (median = 11 minutes [IQR = 6–17]).ConclusionsUsing information available in the 2016 NEMSIS data, we estimate that there were over 570,000 reported adult out‐of‐hospital cardiac arrests in the United States. These results highlight the challenges of characterizing the epidemiology of adult out‐of‐hospital cardiac arrest in the United States.

Objective To investigate the epidemiological features of out-of-hospital cardiac arrest (OHCA) in Zhejiang and to analysis factors associated with outcomes for providing evidence on improving the success rate of cardiopulmonary resuscitation (CPR). Methods Clinical data of 493 patients with OHCA collected from the emergency department (ED) of Zhejiang Provincial People's Hospital, Ningbo Emergency Medical Service Center and Shaoxing Emergency Medical Service Center from January 2012 to August 2016 were analyzed retrospectively. All the data were recorded following the Utstein style included causes of arrest, location of arrest, first arrest rhythm, witnessed by bystanders, bystander CPR, pre-hospital defibrillation, pre-hospital intubation, pre-hospital epinephrine administration, emergency medical services (EMS) response time, return of spontaneous circulation (ROSC) at scene/enroute, ROSC at ED, admission to hospital, survival to hospital discharge and neurological outcomes at discharge. Factors associated with 30 days survival rate of patients with OHCA were analyzed. Results 493 patients were enrolled, of whom 342 were male and 151 were female. The average age was (58.8±21.4) years. The causes of arrests consisted of cardiac etiology (219 cases), trauma (155 cases), respiratory disease (22 cases), drowning (19 cases), electrocution (8 cases) and others (70 cases). Most of the events occurred at home (65.1%) and public places (22.7%). 55.2% patients were witnessed by bystanders while bystander CPR was performed in only 2.6% cases. Asystole was the predominant rhythm (78.7%) observed by the ambulance crew at the arrest site while only 5.5% first arrest rhythms were ventricular fibrillation/ventricular tachycardia (VF/VT). Only 6.9% patients underwent pre-hospital defibrillation. Pre-hospital intubations were attempted in 16.4% patients. Epinephrine was administered in 56.4% patients at scene or in ambulances. EMS response time was (13.6±8.0) minutes. 4.5% patients had ROSC at scene/enroute and 7.7% had ROSC at ED. Only 9.7% patients were admitted to hospital and 1.2% discharged alive. 0.8% patients were still in hospital on 30th day. The 30-day survival rate was 2.0% (10/493) and only 0.8% patients had neurologically favorable survivals [with the cerebral performance category (CPC) score of 1 or 2]. Witnessed by bystanders (3.31% vs. 0.45%), VF/VT as the first arrest rhythm (7.41% vs. 1.72%), bystander CPR (15.38% vs. 1.67%), pre-hospital defibrillation (8.82% vs. 1.53%) and EMS response time < 10 minutes (3.57% vs. 0.74%) could improve 30-day survival rate of OHCA significantly (all P < 0.05). Conclusion Resuscitation survival of OHCA in Zhejiang province was unsatisfactory. Improvements are required in series aspects of OHCA survival chain. Key words: Out-of-hospital cardiac arrest; Cardiopulmonary resuscitation; Epidemiology; Outcome; Emergency medical services response time

Background: Fluctuations in emergency medical services (EMS) responses can have a substantial impact on the ability of agencies to meet resource needs within an EMS system. We aimed to identify weather characteristics as potentially predictable factors associated with EMS responses. Methods: We reviewed hourly counts of scene responses documented by 24 EMS agencies in Western Pennsylvania from January 1, 2014 to December 31, 2017 and compared rates of responses to weather characteristics. Responses to counties nonadjacent to the studied weather reporting station and interfacility/scheduled transports were excluded. We identified the mean temperature, meters visibility, dew point, wind speed, total millimeters of precipitation, and presence of rain or snow in 6-hour windows prior to dispatch, in addition to temporal factors of time of day and weekend vs. weekday. Analysis was performed using multivariable linear regression of a negative binomial distribution, reporting incidence rate ratios (IRR) with 95% confidence intervals (CI). Secondary analyses were performed for transports to the hospital and cases involving transports for traumatic complaints and pediatric patients (age <18 years). Results: We included 529,058 responses (54.8% female, mean age 57.2 ± SD 24.7 years). In our multivariable model, responses were associated with (IRR, 95% CI) rain (1.10, 1.08–1.11) snow (1.07, 1.05–1.09), and both rain and snow (1.15, 1.11–1.19). A lower incidence of responses occurred on weekends (0.84, 0.83–0.85) and at night (0.62, 0.61–0.62). Increasing temperature in 5 °C increments was associated with an increase in responses across seasons with an effect that varied between 1.16 (1.15–1.17) in winter to 1.31 (1.28–1.33) in summer. Windy weather was associated with increased responses from light breeze (1.10, 1.09–1.11) to fresh breeze or greater (1.23, 1.16–1.30). Transports occurred in a similar pattern to responses. Trauma transports (n = 64,235) occurred more during weekends (1.04, 1.02–1.06). Pediatric transports (n = 21,880) were not significantly associated with precipitation or season. Conclusion: EMS responses increased with rising temperature and following rain and snow. These findings may assist in planning by EMS agencies and emergency departments to identify periods of greatest resource utilization.

Objective: While emergency medical service (EMS) response times (ERT) remain a leading measure of system performance in many developed countries, relatively few studies have explored the factors associated with meeting benchmark performance for potentially time critical incidents. The purpose of this study was to identify system-level and patient-level factors associated with ERT, which are readily available at the time of ambulance dispatch. Methods: Between July 2009 and June 2014, we included data from 1,000,458 EMS responses to time critical “lights and sirens” incidents in Melbourne, Australia. The primary outcome measure was ERT, defined as the time from emergency call to the arrival of the first EMS team on scene. Quantile regression models were used to identify system-level and patient-level factors associated with 10-percentile intervals of ERT. Results: The median ERT was 10.6 minutes (IQR: 8.1–14.0), increasing from 9.6 minutes (IQR: 7.6–12.5) in 2009/10 to 11.0 minutes (IQR: 8.4–14.7) in 2013/14 (p < 0.001). System-level factors independently associated with the 90th percentile ERT were distance to scene, activation time, turnout time, case upgrade, hour of day, day of week, workload in the previous hour, ambulance skill set, priority zero case (e.g., suspected cardiac or respiratory arrest), and average hospital delay time in the previous hour. Patient-level factors such as age, gender, chief medical complaint, and severity of complaint were also significantly associated with ERT. Conclusions: System-level and patient-level factors available at the time of ambulance dispatch are useful predictors of ERT performance, which could be used to improve the timeliness of EMS response.

Objective. Despite its long history and current prominence in U.S. communities, only limited data describe the national characteristics of emergency medical services (EMS) care in the United States. We sought to characterize out-of-hospital EMS care in the United States. Methods. We conducted an analysis of the 2010 National Emergency Medical Services Information System (NEMSIS) research data set, encompassing EMS emergency response data from 29 states. From these data, we estimated the national number and incidence of EMS responses. We also characterized EMS responses and the patients receiving care. Results. There were 7,563,843 submitted EMS responses, corresponding to an estimated national incidence of 17.4 million EMS emergency responses per year (56 per 1,000 person-years). The EMS response incidence varied by U.S. Census region (South 137.4 per 1,000 population per year, Northeast 85.2, West 39.7, and Midwest 33.3). The use of lights and sirens varied across Census regions (Northeast 90.3%, South 76.7%, West 68.8%, and Midwest 67.5%). The percentage of responses resulting in patient contact varied across Census regions (range 78.4% to 95.7%). The EMS time intervals were similar between Census regions; response median 5 minutes (interquartile range [IQR] 3–9), scene 14 minutes (10–20), and transport 11 minutes (7–19). Underserved populations (the elderly, minorities, rural residents, and the uninsured) were large users of EMS resources. Conclusion. These data highlight the breadth and diversity of EMS demand and care in the United States.

ObjectiveEmergency medical dispatch is intended to ensure that emergency medical services (EMS) allocate appropriate resources for the treatment of patients with status epilepticus (SE). However, it is unclear whether dispatch algorithms accurately identify those patients having a seizure-related medical emergency and how dispatch algorithms influence what prehospital resources are allocated for the encounter.MethodsWe performed a cross-sectional analysis of prehospital encounters for SE using data from the 2019 ESO Data Collaborative. We included patients who were ≥18 years of age, had an EMS diagnostic impression of SE, and did not have a cardiac arrest. We examined the dispatch-determined complaint designated by the emergency medical dispatch (EMD) code, dispatch-determined level of acuity (A, B, C, D), ambulance response, and training level of the responding prehospital professional.ResultsOf the 18,515 patient encounters for SE with an EMD code, 8,279 (44.9%) were women, and the mean age was 40.0 years (SD 19.7). There were 13,829 (75%) encounters that received a dispatch code for seizures/convulsions and 4,686 (25%) with a dispatch code for a non-seizure-related condition. Among encounters for SE identified by dispatch as seizures/convulsions, 6,412 (46.4%) were designated high acuity, 6,626 (63.6%) were designated low acuity, and the majority received emergent ambulance responses (98.1% among those designated high acuity and 81.8% among those designated low acuity) and an Advanced Life Support-trained responder (93.7% among those designated high acuity and 92.7% among those designated low acuity). Median response times were similar for all acuity levels (9.1, 8.8, 9.1, and 8.3 minutes for A-D, respectively).ConclusionApproximately one-fourth of SE cases were categorized as a non-seizure related condition at dispatch, and fewer than half received the highest acuity determinant code. Despite this, dispatch-assigned acuity did not have a strong relationship with the ambulance response or training level of the EMS responder or response time, suggesting that use of dispatch algorithms might be further optimized and highlighting a potential area to improve quality of EMS care.

The advent of Tiny Machine Learning (TinyML) has unlocked the potential to deploy machine learning models on resource-constrained edge devices, revolutionizing real-time monitoring in Internet of Medical Things (IoMT) applications. This study introduces a novel approach to real-time electrocardiogram (ECG) anomaly detection by integrating TinyML with edge Artificial Intelligence (AI) on low-power embedded systems. We demonstrate the feasibility and effectiveness of deploying optimized models on edge devices, such as the Raspberry Pi and Arduino, to detect ECG anomalies, including arrhythmias. The proposed workflow encompasses data preprocessing, feature extraction, and model inference, all executed directly on the edge device, eliminating the need for cloud resources. To address the constraints of memory and power consumption in wearable devices, we applied advanced optimization techniques, including model pruning and quantization, achieving an optimal balance between accuracy and resource utilization. The optimized model achieved an accuracy of 92.3% while reducing the power consumption to 0.024 mW, enabling continuous, long-term health monitoring with minimal energy requirements. This work highlights the potential of TinyML to advance edge AI for real-time medical applications.

People with mental illness have higher use of emergency services than the general population and may experience problems with navigating in complex healthcare systems. Poor physical health contributes to the excess mortality among the mentally ill. To compare the level of Emergency Medical Services (EMS) response and subsequent contacts emergency between patients with and without a history of mental illness. A nationwide cohort study was conducted in Denmark including medical 1-1-2 calls 2016 2017. The healthcare system is financed through taxation allowing free access to healthcare services including ambulance services. Exposed groups had a history of major, moderate, or minor mental illness. We studied seven national prehospital care Performance Indicators (PI 1-7). The selected PI concerned EMS response levels and subsequent contacts to prehospital and in-hospital services. Exposed groups were compared to nonexposed groups using regression analyses. We included 492 388 medical 1-1-2 calls: 8, 10, and 18% of calls concerned patients with a history of major, moderate, or minor mental illness, respectively.There were no clinically relevant differences regarding response times (PI 1-2) or registration of symptoms (PI 3) between groups.If only telephone advice was offered, patients with a history of major, moderate or minor mental illness were more likely to recall within 24 h (PI 4): adjusted risk ratio (RR) 2.11 (1.88-2.40), 1.96 (1.20-2.21), and 1.38 (1.20-1.60), but less or equally likely to have an unplanned hospital contact within 7 days (PI 6): adjusted RRs 1.05 (0.99-1.12), 1.04 (0.99-1.10), and 0.90 (0.85-0.94), respectively.If released at the scene, the risk of recalling (PI 5) or having an unplanned hospital contact (PI 7) was higher among patients with a history of mental illness: adjusted RRs 2.86 (2.44-3.36), 2.41 (2.05-2.83), and 1.57 (1.35-1.84), and adjusted RRs 2.10 (1.94-2.28), 1.68 (1.55-1.81), and 1.25 (1.17-1.33), respectively.Patients with a history of mental illness were more likely to receive telephone advice only adjusted RRs 1.61 (1.53-1.70), 1.30 (1.24-1.37), and 1.08 (1.04-1.13), and being released at scene adjusted RRs 1.11 (1.08-1.13), 1.03 (1.01-1.04), and 1.05 (1.03-1.07). More than one-third of the study population had a history of mental illness. These patients received a significantly lighter EMS response than patients with no history of mental illness. They were significantly more likely to use the emergency care system again if released at scene. This risk increased with the increasing severity of the mental illness.

Advances in deep learning, especially Convolutional Neural Networks (CNNs) have revolutionized intelligent frame-works such as Human Activity Recognition (HAR) systems by effectively and efficiently inferring human activity from various modalities of data. However, the training and inference of CNNs are often resource-intensive. Recent research developments are focused on bringing the effectiveness of CNNs in resource con-strained edge devices through Tiny Machine Learning (TinyML). However, this is extremely hard to achieve due to the limitations in memory, compute power, and energy of resource constrained edge devices. This paper provides a benchmark to understand these trade-offs among variations of CNN network architectures, different training methodologies, and different modalities of data in the context of HAR, TinyML, and edge devices. We tested and reported the performance of CNN and Depthwise Separable CNN (DSCNN) models as well as two training methodologies: Quantization Aware Training (QAT) and Post-training Quantization (PTQ) on five commonly used benchmark datasets containing image and time-series data: UP-Fall, Fall Detection Dataset (FDD), PAMAP2, UCI-HAR, and WISDM. We also deployed and tested the performance of the model-based standalone applications on multiple commonly available resource constrained edge devices in terms of inference time and power consumption. The experimental results demonstrate the effectiveness and feasibility of Tiny ML for HAR in edge devices.

Earthquake P-wave detection is the first and most critical step in Earthquake Early Warning (EEW), as detection speed determines how much time is available before strong shaking arrives, while detection accuracy governs the reliability of the warning. Conventional P-Wave detection methods such as Short-Term Average over Long-Term Average (STA/LTA) often struggle in complex noise environments and exhibit higher detection latency. Deep learning offers a more robust alternative by learning complex waveform characteristics directly from data, enabling faster and more reliable detections. However, the high computational cost of deep models limits their use on low-cost, resource-constrained edge devices in decentralized EEW networks. Here we present a lightweight Convolutional Neural Network optimized for real-time P- and S-wave detection on edge hardware using only two seconds of seismic input waveforms. Trained on approximately 89K waveform segments recorded by strong-motion sensors across New Zealand, the model achieves an overall accuracy of 97.12% and correctly identifies 98% of P-wave segments. With ~ 38K trainable parameters and sub-7-millisecond inference time on Raspberry Pi 5 hardware, the model demonstrates efficient operation under edge-computing constraints. Its ability to generalize to high-magnitude events highlights its strong potential for integration into distributed EEW systems.

Convolutional neural networks (CNNs) have gained a huge attention for real-world artificial intelligence (AI) applications such as image classification and object detection. On the other hand, for better accuracy, the size of the CNNs’ parameters (weights) has been increasing, which in turn makes it difficult to enable on-device CNN inferences in resource-constrained edge devices. Though weight pruning and 5-bit quantization methods have shown promising results, it is still challenging to deploy large CNN models in edge devices. In this paper, we propose an encoding and hardware-based decoding technique which can be applied to 5-bit quantized weight data for on-device CNN inferences in resource-constrained edge devices. Given 5-bit quantized weight data, we employ arithmetic coding with range scaling for lossless weight compression, which is performed offline. When executing on-device inferences with underlying CNN accelerators, our hardware decoder enables a fast in-situ weight decompression with small latency overhead. According to our evaluation results with five widely used CNN models, our arithmetic coding-based encoding method applied to 5-bit quantized weights shows a better compression ratio by 9.6× while also reducing the memory data transfer energy consumption by 89.2%, on average, as compared to the case of uncompressed 32-bit floating-point weights. When applying our technique to pruned weights, we obtain better compression ratios by 57.5×–112.2× while reducing energy consumption by 98.3%–99.1% as compared to the case of 32-bit floating-point weights. In addition, by pipelining the weight decoding and transfer with the CNN execution, the latency overhead of our weight decoding with 16 decoding unit (DU) hardware is only 0.16%–5.48% and 0.16%–0.91% for non-pruned and pruned weights, respectively. Moreover, our proposed technique with 4-DU decoder hardware reduces system-level energy consumption by 1.1%–9.3%.

Objectives Emergency medical services (EMS) systems increasingly grapple with rising call volumes and workforce shortages, forcing systems to decide which responses may be delayed. Limited research has linked dispatch codes, on-scene findings, and emergency department (ED) outcomes. This study evaluated the association between dispatch categorizations and time-critical EMS responses defined by prehospital interventions and ED outcomes. Secondarily, we proposed a framework for identifying dispatch categorizations that are safe or unsafe to hold in queue. Methods This retrospective, multi-center analysis encompassed all 9-1-1 responses from 8 accredited EMS systems between 1/1/2021 and 06/30/2023, utilizing the Medical Priority Dispatch System (MPDS). Independent variables included MPDS Protocol numbers and Determinant levels. EMS treatments and ED diagnoses/dispositions were categorized as time-critical using a multi-round consensus survey. The primary outcome was the proportion of EMS responses categorized as time-critical. A non-parametric test for trend was used to assess the proportion of time-critical responses Determinant levels. Based on group consensus, Protocol/Determinant level combinations with at least 120 responses (∼1 per week) were further categorized as safe to hold in queue (<1% time-critical intervention by EMS and <5% time-critical ED outcome) or unsafe to hold in queue (>10% time-critical intervention by EMS or >10% time-critical ED outcome). Results Of 1,715,612 EMS incidents, 6% (109,250) involved a time-critical EMS intervention. Among EMS transports with linked outcome data (543,883), 12% had time-critical ED outcomes. The proportion of time-critical EMS interventions increased with Determinant level (OMEGA: 1%, ECHO: 38%, p-trend < 0.01) as did time-critical ED outcomes (OMEGA: 3%, ECHO: 31%, p-trend < 0.01). Of 162 unique Protocols/Determinants with at least 120 uses, 30 met criteria for safe to hold in queue, accounting for 8% (142,067) of incidents. Meanwhile, 72 Protocols/Determinants met criteria for unsafe to hold, accounting for 52% (883,683) of incidents. Seven of 32 ALPHA level Protocols and 3/17 OMEGA level Protocols met the proposed criteria for unsafe to hold in queue. Conclusions In general, Determinant levels aligned with time-critical responses; however, a notable minority of lower acuity Determinant level Protocols met criteria for unsafe to hold. This suggests a more nuanced approach to dispatch prioritization, considering both Protocol and Determinant level factors.