Ambulance: Emergency Medical Systems

• Ambulance is an emergency vehicle that delivers pre-hospital care, rapid response, and coordinated transport within integrated health systems.
• Research examines operational cycles, stochastic demand modeling, and optimization techniques that reduce response times and enhance service coverage.
• Advanced approaches use simulation, routing innovation, and patient-centered communication to improve dispatch, triage, and overall system preparedness.

An ambulance is an emergency medical vehicle used to provide pre-hospital care, rapid response, and, when necessary, transport to an appropriate hospital or emergency department. In contemporary research, the term covers conventional road ambulances, fast response units, motorcycle ambulances, diversion-capable units, and, in multimodal systems, the ground first- and last-mile component linking hospitals and vertiports. The ambulance is therefore studied not only as a vehicle, but as the central mobile resource in a tightly coupled system of demand forecasting, dispatch, routing, relocation, patient triage, in-transit communication, and fleet-level preparedness under uncertainty (Santis et al., 27 Apr 2026, Guigues et al., 2022, Stratman et al., 29 May 2025, Varnousfaderani et al., 2 Feb 2026).

1. Operational role and system structure

Ambulance systems are organized around a recurrent operational cycle: call receipt, telephone triage, dispatch, preparation, travel to scene, on-scene care, transport if needed, hospital handover, possible cleaning or sanitization, and return to base or redeployment. A realistic regional discrete-event model for ARES 118 in Lazio explicitly represents telephone triage and ambulance preparation times, on-scene service, transport to clinically appropriate emergency departments, ambulance ramping and offload delay, sanitization, and return-to-base or direct reassignment to a nearby new call (Santis et al., 27 Apr 2026).

This operational cycle is constrained by vehicle heterogeneity. Rio de Janeiro studies distinguish BLS and ALS ambulances and model compatibility between ambulance capability and emergency type, including explicit penalties for mismatches such as sending BLS to ALS-preferred emergencies (Guigues et al., 2022). London data distinguish Accident and Emergency Units and Fast Response Units under “blue light” conditions, where ambulances may exceed posted speed limits, treat red traffic lights and zebra crossings as give-way controls, turn against turn restrictions, enter pedestrianized precincts, and use bus lanes during operating hours (Poulton et al., 2018).

The role of the ambulance is not identical across health systems. In low- and middle-income cities such as Dhaka, ambulance travel times may resemble regular traffic because motorists do not yield to emergency vehicles, and many roads are too narrow for vans, which directly alters siting and routing decisions (Boutilier et al., 2018). In multimodal medical transport, ambulances remain the indispensable default mode for patients and the universal first- and last-mile mode for patients, organs, and supplies, even when UAVs or eVTOLs are introduced for middle segments (Varnousfaderani et al., 2 Feb 2026).

2. Demand as a spatio-temporal stochastic process

A central research problem is estimating where ambulance demand will occur at fine spatial and temporal scales. One influential formulation models ambulance calls in period $t$ as a nonhomogeneous Poisson process on a spatial domain $D$ with intensity $\lambda_t(\mathbf{s})$, decomposed as

$\lambda_t(\mathbf{s})=\alpha_t f_t(\mathbf{s}),$

where $\alpha_t$ is expected call volume and $f_t$ is the spatial density of calls. For Toronto, $f_t$ is represented by a Gaussian mixture with component distributions fixed across time and time-varying mixture weights, allowing stable spatial “hotspots” while capturing changing time-of-day and day-of-week patterns (Zhou et al., 2014).

The Toronto model addresses a setting with only about 45 priority calls in a typical 2-hour period across the entire city. To overcome that sparsity, it shares component structure across time, constrains weights to repeat over the 84 two-hour intervals of a week, and uses a conditionally autoregressive prior on transformed weights to encode location-specific short-term serial dependence and daily seasonality. In out-of-sample tests, the mixture model outperformed MEDIC and MEDIC-KDE baselines and reduced the error in predicting operational coverage from the 44 Toronto bases by as much as two-thirds (Zhou et al., 2014).

Demand regularity appears at larger scales as well. A study of 5.6 million ambulance calls in the North West Ambulance Service region identified stable diurnal and weekly cycles by call nature, with morning peaks for Falls, Sick Person, and Chest Pain, and evening or weekend concentration for Overdose/Poisoning and Psychiatric/Suicide. The same work linked ambulance incidence to population activity through Foursquare-derived venue use, defining Spatial Attractiveness, Temporal Synchronicity, the composite $ST\_Risk_{ij}$, and an area-level Urban Activity Risk

$UAR_a := \sum_{j \in Z_a} \sum_{i \in N} ST\_Risk_{ij},$

which ranked above IMD and residential population in feature importance for hour-level call prediction (Todorova et al., 2019).

Regional simulation work uses a different but complementary representation. In Lazio, calls are generated by a two-stage mechanism: arrival time from a generation-zone interarrival distribution, then spatial placement within smaller call squares according to historical density. This preserves urban-rural heterogeneity and supports validation against empirical response-time coverage curves (Santis et al., 27 Apr 2026).

3. Dispatch, stationing, and preparedness optimization

Ambulance operations require repeated decisions at two event types: when a call arrives and when an ambulance becomes free. In Rio de Janeiro, these are formalized as the ambulance selection problem and the ambulance reassignment problem. The corresponding rolling-horizon framework combines immediate decisions with second-stage models of future call uncertainty, using large-scale deterministic integer programs and column generation for the continuous relaxation. With 20 ambulances, the itinerary-based policy achieved penalized response time of approximately 10,742 versus approximately 15,637 for Closest Available; with 12 ambulances, it achieved approximately 14,327 versus approximately 26,789 (Guigues et al., 2022).

A related line of work replaces scenario-based lookahead with a preparedness metric derived from a continuous-time Markov chain at each station. Preparedness is the steady-state expected cost rate of future emergencies given the current mix of available ambulances by type, local arrival rates, service-time characteristics, and dispatch preferences. This metric is embedded in 0–1 linear programs for real-time dispatch and reassignment, and in Rio de Janeiro simulations it outperformed nine literature methods while remaining computationally practical, with mean solve times in the millisecond range for the tested fleet sizes (Guigues et al., 6 May 2026).

Heuristic and rollout methods provide another operational compromise. New heuristics such as Best Myopic, NonMyopic, GHP1, and GHP2 explicitly trade off immediate response against future opportunity cost. Embedded in a two-stage stochastic rollout framework, they produced better response times than using the heuristics alone, while still computing each decision in a few seconds (Guigues et al., 2024).

Strategic stationing problems are also studied at longer horizons. OpenEMS formulates ambulance stationing in Austin as a two-stage stochastic and robust optimization problem over discretized space, with first-stage station decisions and second-stage dispatch decisions. In peak hours, the reported historical baseline was $7.677 \pm 0.443$ minutes, while optimized stochastic stationing achieved $D$0 minutes, an improvement of 88.02 seconds (Ong et al., 2022). In LMIC settings, robust outpost location and routing models further incorporate edge-level travel-time uncertainty and bootstrapped demand scenarios; in Dhaka, a greenfield redesign with 20 optimized outposts and 140 ambulances could replicate current performance with roughly 30–50% of current resources (Boutilier et al., 2018).

These results challenge the common operational simplification that the closest currently visible ambulance is necessarily the best choice. Optimization-augmented machine learning for San Francisco learned dispatching and redeployment policies from full-information solutions and reported reductions in mean response time of up to 30% relative to online benchmarks (Rautenstrauß et al., 14 Mar 2025). Likewise, a market-based multi-robot task allocation formulation for London selected a different first responder than history in 89–97% of evaluated incidents and reduced predicted response time by 48.2% to 66.4% across sampled settings (Schneider et al., 2020).

4. Routing, mobility, and travel-time modeling

Ambulance routing differs fundamentally from civilian routing because the governing traffic regime is different. In London, a Blue Lights Road Network was built by modifying the civilian road graph to reflect legal emergency exemptions and then estimating segment-level emergency speeds from 1,910,941 map-matched journeys and 177,975,172 edge traversals. Google’s Distance Matrix API overestimated blue-light journey times by factors of about 1.4 for AEUs and 1.5 for FRUs in the LAS data (Poulton et al., 2018).

The London study compared several travel-time metrics and found a trade-off between route similarity and ETA accuracy. Metric II, which mirrored the legacy LAS routing engine, achieved about 80% path coincidence, while Metric V, based on edge-specific hour-of-week and vehicle-type harmonic mean speeds, gave the best raw duration accuracy but poorer route similarity. A hybrid approach—routing with optimized Metric II and timing with Metric V—reached approximately 84% path coincidence and, with bias correction, error within approximately 60 seconds up to approximately 15-minute journeys (Poulton et al., 2018).

Routing research is not limited to minimizing travel time. A vibration-aware driving assistance system for ambulances used a Raspberry Pi 3 sensor node with a triaxial accelerometer and GPS, classifying road segments into low, medium, and high vibration classes. The ANN-based classifier achieved 97% accuracy, with 94.34% correctly classified using 29-second buffer features. In route tests, the route with less vibration was preferred when there were low time differences, less than 6%, between the two possible routes, whereas with the current weighting factors the shortest route was preferred when time differences between routes were higher than 20%, regardless of the higher vibrations in the shortest route (Aldegheishem et al., 17 Apr 2026).

In dense LMIC networks, routing flexibility can dominate vehicle speed. In Dhaka, motorcycle-based ambulances accessed 4,828 additional nodes relative to the ambulance network, captured approximately three times more demand, and reduced the median response time by 42% because of the larger accessible road network and greater resilience to localized edge delays (Boutilier et al., 2018).

A separate conceptual literature treats ambulance routing in smart cities as a scheduling problem. One proposal models an autonomous emergency vehicle mission as a mixed-criticality real-time systems task, where route selection, signal pre-emption, and lane reservation are interpreted as scheduling decisions designed to guarantee a response deadline under dynamic congestion, queues, pedestrian flows, and work zones (Humagain et al., 2021). This suggests a convergence between EMS routing, cyber-physical systems, and networked traffic control, although the cited model remains conceptual and reports no simulations or numeric results (Humagain et al., 2021).

5. Patient-centered pathways and in-transit clinical communication

The traditional ambulance workflow assumes transport to an emergency department, but recent work explicitly challenges that assumption. A patient-centered EMS allocation model introduces multiple pathways—ED transport, alternative destination, and treat-in-place—and formulates a two-stage mixed-integer model with adaptive dispatch and secondary assignment. Simulations across diverse regions suggest that agencies can achieve up to 80% of possible ED diversions by equipping only 15 to 25% of their fleet with diversion-capable units, and that adaptive dispatch strategies improve diversion rates by 3.4 to 8.6 times compared to conventional single unit dispatch (Stratman et al., 29 May 2025).

The patient-centered model protects system reliability through a queueing-based availability constraint using an M/G/d/d loss formulation. Operationally, it allows multiple response on ambiguous calls, on-scene reassessment, and secondary assignment of diversion-capable units when initial information is insufficient. A plausible implication is that PC-EMS implementation is not only a clinical redesign but also a fleet-composition and availability-management problem (Stratman et al., 29 May 2025).

Ambulance care during transport also depends on communications infrastructure. A physiology-aware DASH framework for rural ambulance transport treats clinical multimedia streams as a multiple-choice knapsack problem under variable bandwidth. It uses clinical automata, a Physiology-aware Priority Calculator, and adaptation heuristics named Compromise, Round-Robin, and Aggressive to preserve higher fidelity for the most clinically critical streams when bandwidth drops from 4G to 2G to low-speed satellite links. The measured Hoopeston-to-Urbana route exhibited throughput ranging from a few Kbps to several Mbps, and the preliminary evaluation showed that prioritized strategies preserved critical data better than a no-priority baseline (Hosseini et al., 2017).

Ambulances also remain central in multimodal medical transport systems. In a constructive greedy dispatch framework integrating ambulances, UAVs, and eVTOLs, ambulances handle all payload types and remain the default mode when traffic and cost conditions favor ground operations. They also serve as the first- and last-mile connector for air legs. Under extreme time priority, the fully integrated fleet performed best; under cost-dominant settings, ambulances plus UAVs matched or outperformed eVTOL-inclusive plans because eVTOL operating costs dominated when consolidation was weak (Varnousfaderani et al., 2 Feb 2026).

6. Simulation, visualization, and recurring misconceptions

Because ambulance systems are stochastic, event-driven, and spatially heterogeneous, simulation has become a standard instrument for evaluation. A web-based EMS environment integrates data ingestion, Poisson-based forecasting, discrete-event simulation, five dispatch heuristics, and animated geospatial visualization of calls, ambulance trajectories, and response-time distributions. The public Rio de Janeiro dataset supports comparison among Closest Available, Best Myopic, Nonmyopic, GHP1, and GHP2, and the same interface accepts user-uploaded EMS data for custom analysis (Guigues et al., 2024).

High-fidelity regional simulation can materially alter planning conclusions. The Lazio DES explicitly models ED destination choice by clinical suitability, ambulance ramping, handover, sanitization, and calibrated road-network travel times. In that setting, adding one H24 ambulance at the Rieti fire station improved urgent 20-minute coverage from 45.29% to 53.41% and non-urgent 20-minute coverage from 7.28% to 8.53%, whereas converting one H24 ambulance at Rieti to H12 worsened performance (Santis et al., 27 Apr 2026).

Several persistent misconceptions are therefore not supported by the research record. First, closest-available dispatch is often dominated by policies that preserve future coverage or use better travel-time estimates (Guigues et al., 2022, Rautenstrauß et al., 14 Mar 2025, Schneider et al., 2020). Second, civilian navigation engines systematically misestimate emergency vehicle motion under blue-light conditions (Poulton et al., 2018). Third, ambulance operations cannot be reduced to transport-to-ED logistics once diversion-capable pathways are operationalized (Stratman et al., 29 May 2025). Fourth, coarse demand averages obscure the location-specific daily and weekly persistence that drives operational performance at 2-hour and subcity scales (Zhou et al., 2014).

Current research directions accordingly emphasize end-to-end integration: demand forecasting tied to dispatch, routing adapted to emergency traffic regimes, queueing-aware preparedness, patient-centered pathway selection, multimodal transport, and simulation or visualization environments that can expose the operational consequences of policy changes before field deployment (Guigues et al., 2024, Santis et al., 27 Apr 2026). This suggests that the ambulance, in modern systems research, is best understood not as an isolated vehicle but as the mobile control point of a stochastic, geographically embedded, clinically differentiated service network.