Emergency Department Flow

Module 7: Healthcare Operations Applications Depth: Application | Target: ~2,500 words

Thesis: Emergency department crowding is a system-level flow problem with known OR solutions that are underdeployed because the problem is misattributed to demand volume alone.

The Operational Problem

Emergency department crowding kills people. Not metaphorically. The evidence is direct: Chalfin et al. (2007) showed that patients who boarded in the ED for more than six hours before ICU transfer had significantly higher mortality than those transferred promptly. The IOM’s landmark 2007 report “Hospital-Based Emergency Care: At the Breaking Point” called ED crowding a national crisis with measurable harm to patients, providers, and health system capacity. Nearly two decades later, the problem has worsened.

The standard explanation is wrong. The default narrative — “too many patients showing up at the ED” — attributes crowding to input volume and prescribes demand management: urgent care centers, nurse triage lines, public education campaigns. These interventions address a real but secondary factor. The dominant cause of ED crowding is not that too many patients arrive. It is that patients who need to leave the ED cannot, because the rest of the hospital will not absorb them. This is an output failure, not an input problem. And the distinction between the two is the difference between solving ED crowding and perpetually managing it.

Operations research has known this for decades. The tools — queueing models, simulation, scheduling optimization, network flow analysis — are mature and validated. What is missing is not the math. It is the diagnostic framework that correctly locates the problem before applying the solution.

The Input-Throughput-Output Model

Asplin et al. (2003) formalized the canonical framework for understanding ED flow by decomposing crowding into three distinct components: input, throughput, and output. This is not a metaphor. It is a systems decomposition that maps directly to queueing and flow models.

Input is the arrival process. How many patients present to the ED, at what rate, with what acuity distribution, through which channels (walk-in, ambulance, transfer). Input is governed by community demand, seasonal patterns, time-of-day effects, and the availability of alternative care sites. In queueing terms, input defines the arrival rate lambda and its variability structure.

Throughput is the service process. Once a patient arrives, how quickly and effectively the ED moves them through triage, evaluation, diagnostic workup, treatment, and disposition. Throughput is a function of provider staffing, bed availability within the ED, ancillary service turnaround (lab, imaging), consultation response times, and the physical layout of the department. In queueing terms, throughput determines the service rate mu and the number of effective servers c.

Output is the departure process. Once a patient has a disposition — admit, discharge, or transfer — how quickly they actually leave the ED. For discharged patients, output is relatively fast (paperwork, prescriptions, discharge instructions). For admitted patients, output depends entirely on the availability of an inpatient bed. When no bed is available, the admitted patient remains in the ED, occupying an ED bed, consuming ED nursing resources, and contributing nothing to ED throughput. This is boarding.

The Asplin framework matters because it makes the diagnostic question precise: Is the crowding problem driven by input, throughput, or output? Each has different mechanisms, different interventions, and different responsible parties. Treating an output problem with input solutions wastes time and money. Treating an input problem with throughput solutions produces the same result.

Input Factors: Arrivals Are Variable, Not Random

ED arrival patterns are stochastic but not unstructured. They follow well-characterized time-of-day and day-of-week patterns: arrivals peak in late morning and early evening, fall overnight, and are lower on weekends than weekdays for most community EDs. Hoot and Aronsky (2008), in their systematic review of ED crowding, confirmed that while overall volume contributes to crowding, the variability and timing of arrivals matter more than the mean.

The input side includes three specific factors that operators can act on:

Arrival patterns. Poisson arrival assumptions hold reasonably well for short intervals (the classic M in Kendall notation), but arrival rates are non-stationary — they vary by hour. Models that use a single daily average arrival rate will underestimate peak-hour crowding and overestimate overnight crowding. Time-varying queueing models or simulation with non-stationary arrival schedules (Module 6) are required for realistic analysis.

Ambulance diversion. When an ED goes on diversion — signaling that it cannot accept additional ambulance patients — it temporarily reduces its own input by redirecting arrivals to other facilities. But diversion is a system-level game: when one ED diverts, neighboring EDs absorb the load, pushing them closer to their own diversion thresholds. Multi-hospital systems in metropolitan areas can experience cascading diversion, where the diversion of one facility triggers a chain reaction. This is a network flow problem (Module 4), not a single-queue problem.

Triage acuity mix. Not all arrivals impose equal load. An ESI Level 1 (resuscitation) patient consumes far more provider time, nursing attention, and ancillary resources than an ESI Level 5 (non-urgent) patient. The acuity mix determines the effective service demand per arrival. A shift in acuity mix — more high-acuity patients due to aging population, fewer low-acuity patients diverted to urgent care — can increase throughput demand without changing arrival volume.

Throughput Factors: The Internal Service Process

Throughput failures are real and addressable, but they are not where most of the crowding lives. Still, they deserve precise treatment because they are the factors most within the ED’s direct control.

Provider staffing and the service rate. The number of physicians, advanced practice providers, and nurses on shift at any given hour determines the ED’s instantaneous service rate. This is the c in the M/M/c queueing model (Module 2). Staffing that is calibrated to average demand rather than peak demand will produce the predictable result: adequate throughput at 10 AM, catastrophic queuing at 6 PM. Erlang-C calculations can determine the minimum number of providers needed to maintain a target probability of waiting, given arrival rates and average service times per acuity level.

Bed assignment. Within the ED, patients need physical treatment spaces. When all beds are occupied — by patients being treated, by patients awaiting results, and critically, by admitted patients boarding — new arrivals queue in the waiting room. Bed assignment is a resource allocation problem: which arriving patient gets the next available bed, given acuity, time waiting, and expected treatment duration? Most EDs use informal triage-based assignment. Formal assignment algorithms that account for expected length of stay can improve bed turnover.

Lab and imaging turnaround. A patient awaiting a troponin result, a CT read, or a basic metabolic panel is occupying an ED bed but cannot progress toward disposition. Ancillary turnaround time directly extends the effective service time for patients requiring diagnostics. A 45-minute lab turnaround adds 45 minutes to the patient’s ED stay regardless of provider efficiency. This is a serial process constraint: the patient’s total ED time is the sum of sequential stages, and the longest stage dominates.

Consultation delays. When an ED patient requires an inpatient specialty consultation before disposition, the ED is dependent on an external service rate. A hospitalist who responds in 20 minutes versus one who responds in 90 minutes changes the patient’s ED length of stay by over an hour. In queueing terms, the consultation process is a separate queue that the ED patient must transit, and its service rate is outside the ED’s control.

Output Factors: Where the Crowding Actually Lives

Here is where the standard narrative fails. The dominant driver of ED crowding in most hospitals is not input volume and not throughput inefficiency. It is output failure: the inability to move admitted patients out of ED beds and into inpatient beds.

Boarding is the practice of holding admitted patients in the ED because no inpatient bed is available. The patient has been seen, evaluated, diagnosed, and dispositioned — the ED’s clinical work is done — but they remain in the ED, occupying a bed, requiring ongoing nursing care, and blocking that bed from being used for a new arrival.

The American College of Emergency Physicians (ACEP) has documented boarding as a patient safety crisis. Their 2022 data showed that 73% of emergency physicians reported boarding patients in hallways, and the average boarding time in many institutions exceeds four hours. ACEP’s position is unambiguous: boarding patients in the ED for prolonged periods constitutes an unsafe practice that increases mortality, morbidity, and patient suffering.

The mechanism is straightforward in queueing terms. The ED is a multi-server queue. Its effective number of servers is the number of beds staffed and available for new patients. When boarding patients occupy ED beds, the effective server count drops. If a 30-bed ED has 8 beds occupied by boarding patients, it is functioning as a 22-bed ED. The utilization-delay curve (Module 2) does the rest: a reduction from 30 to 22 effective beds, at the same arrival rate, can push utilization from a manageable 75% into the steep part of the curve above 90%, where waits explode.

Discharge delays are the upstream cause of boarding. Inpatient beds are not available because the patients in them have not been discharged, even when they are medically ready. Late-in-the-day discharge practices — attending physicians rounding in the morning, writing discharge orders in the afternoon, with actual departures in the evening — mean that inpatient beds turn over late, long after the morning and midday ED admission surge has begun. The mismatch between inpatient discharge timing and ED admission demand is the structural cause of boarding.

Bed management failures compound the problem. Even when beds are physically available, administrative delays in bed assignment, transport, cleaning, and handoff can add 60-90 minutes between a bed being vacated and the next patient occupying it. This is a bed turnaround time problem — the setup time in scheduling terms (Module 5) — and it is directly addressable with process engineering.

The Diagnostic Case: A Community Hospital ED

Consider a 30-bed community hospital ED seeing 45,000 visits per year — roughly 123 patients per day, or an average of 5.1 patients per hour. Average door-to-disposition time is 4.5 hours. The boarding rate is 25%: one in four patients ultimately admitted waits for an inpatient bed in the ED after disposition.

The administration’s instinct is to add ED capacity — more beds, more staff. Before investing, an OR analysis decomposes the problem using the Asplin framework.

Input analysis. Arrival rate peaks at 7.2 patients/hour between 11 AM and 7 PM, and troughs at 2.1 patients/hour overnight. Overall volume has grown 3% annually, but the crowding problem worsened faster than volume growth would predict. Input is not the binding constraint.

Throughput analysis. Median provider evaluation time (door to medical decision-making) is 2.8 hours, which is within national benchmarks. Lab turnaround averages 52 minutes. Imaging turnaround averages 38 minutes. Consultation response averages 65 minutes. These are improvable — shaving 15 minutes off lab turnaround helps — but they are not the structural problem.

Output analysis. Average boarding time for admitted patients is 3.2 hours. At any given time during peak hours, 7-8 of the 30 ED beds are occupied by boarding patients. This effectively converts the 30-bed ED into a 22-23 bed ED. With peak-hour arrivals of 7.2/hour and average service time of 4.5 hours, the effective utilization of available beds is:

rho = (7.2 x 4.5) / 22.5 = 1.44

This is a utilization above 1.0 — the system is in overload during peak hours. Not “stressed.” Not “busy.” Mathematically unstable. The queue grows without bound until arrivals slow in the evening.

The root cause of the boarding is inpatient discharge timing. Sixty-two percent of inpatient discharges occur after 2 PM. The morning ED admission surge begins at 10 AM. For four to six hours every day, the ED is generating admissions faster than the hospital is discharging inpatients.

Intervention. Two changes target the output bottleneck: (1) smoothing elective surgical admissions across the week, eliminating the Monday-Tuesday surgical surge that consumes inpatient beds mid-week, and (2) implementing a discharge-before-noon protocol that moves target discharges earlier in the day. Neither intervention adds ED capacity. Both address the inpatient flow constraint.

Result: boarding drops 40%. Average boarding time falls from 3.2 hours to 1.9 hours. Peak-hour effective bed count rises from 22-23 to 26-27. Effective utilization drops from the overload zone into the steep-but-manageable range around 82%. Door-to-disposition improves from 4.5 to 3.4 hours.

The ED was never the problem. The ED was where the hospital’s flow failure became visible.

Split-Flow: Queue Discipline Redesign

The split-flow model, developed and validated by Welch et al., addresses throughput by redesigning the ED’s queue discipline rather than adding capacity. In queueing terms, it replaces a single FIFO queue with a set of priority queues, each with its own service process optimized for the acuity class.

Fast-track creates a dedicated service line for low-acuity patients (ESI 4-5). These patients have predictable, short service times and require minimal diagnostics. Routing them to a separate queue with dedicated providers prevents them from competing for main ED beds and prevents high-acuity cases from blocking their rapid treatment. This is the classic priority queue separation: when service time distributions differ dramatically between customer classes, separating queues outperforms pooling. The math is in Module 2 — pooling helps when service times are homogeneous, but hurts when it forces short-service customers to wait behind long-service customers.

Vertical flow (or intake and rapid medical evaluation) targets medium-acuity patients (ESI 3). Instead of placing these patients in a bed for the entire visit, the provider evaluates them while standing — “vertically” — initiates diagnostic workups, and returns the patient to a results-waiting area that does not consume an ED bed. The patient occupies a treatment bed only for procedures that require one. This decouples the physician evaluation from the bed resource, increasing effective bed capacity without adding physical beds.

Split-flow works because it matches service process to demand class. It is the same principle as express checkout lanes in retail or priority boarding in airlines — not a novel insight, but a rigorous application of queue discipline theory to a setting where the stakes are materially higher.

LWBS: The Real-Time Failure Signal

Left Without Being Seen rate is the canary in the ED mine. As established in the abandonment page (Module 2), LWBS is a queueing phenomenon: patients abandon when the perceived wait exceeds their patience threshold. In the ED, LWBS is the real-time signal that the system is failing.

The Emergency Department Benchmarking Alliance flags LWBS above 2% as a performance concern. Sustained rates above 5% indicate structural flow failure, not episodic surges. LWBS responds to the tail of the wait distribution, not the mean — patients leave when waits are extreme, which happens before average wait metrics register alarm. This makes LWBS a leading indicator: it degrades before door-to-provider time, before boarding hours, and before diversion is triggered.

For operators, LWBS is the earliest actionable metric. A rising LWBS rate in real time signals that interventions — pulling boarding patients to hallway inpatient beds, activating fast-track, calling in additional providers — are needed now, not after retrospective analysis confirms the crowding was real.

Product Implications

Software supporting ED operations should surface the flow decomposition, not just aggregate metrics.

Decompose delay by phase. Dashboard design should separately display door-to-triage, triage-to-bed, bed-to-provider, provider-to-decision, and decision-to-departure times. When total door-to-disposition climbs, the phase breakdown shows whether the failure is throughput (long bed-to-decision) or output (long decision-to-departure, i.e., boarding). Most ED dashboards aggregate these phases into a single metric that obscures the root cause.

Track effective bed count in real time. Display the number of ED beds available for new patients, not the total physical bed count. Subtract boarding patients, beds out of service, and beds in terminal cleaning. This effective bed count, combined with current arrival rate, yields real-time utilization — the operating point on the utilization-delay curve. An alert when effective utilization exceeds 85% gives operators a 30-60 minute window to act before waits become critical.

Surface LWBS as a live indicator. Do not relegate LWBS to monthly retrospective reports. Display the current-shift LWBS rate alongside current wait times. When LWBS begins rising, the system is shedding the patients it was built to serve. This is the earliest degradation signal available.

Model the output bottleneck. Integrate inpatient discharge projections with ED boarding forecasts. If the discharge-before-noon target is being missed, the system should project the boarding impact on ED flow three to four hours forward. This converts a reactive boarding crisis into a predictable and manageable inpatient discharge problem.

Warning Signs

• “We need more ED beds” as the default response to crowding. If the boarding rate exceeds 15%, the problem is not ED capacity. It is hospital flow. Adding ED beds creates more places to board patients, not more throughput.
• Boarding treated as an ED problem rather than a hospital problem. When the CNO or CMO frames boarding as “the ED can’t manage its flow,” the diagnostic is backwards. Boarding is an inpatient discharge failure that manifests in the ED.
• Staffing models that ignore boarding load. An ED staffing model calibrated to arrival volume and average service time will understaff during boarding surges because it does not account for the nursing hours consumed by patients who have already been dispositioned but have not left.
• LWBS above 2% without root cause analysis. Sustained LWBS elevation is not a PR problem. It is an access failure that means a measurable fraction of patients who needed emergency care did not receive it.
• Elective surgical scheduling that ignores downstream effects. Surgical scheduling that front-loads the week creates predictable mid-week inpatient census peaks that consume beds the ED needs for admissions. If surgery scheduling and ED flow are managed by different departments with no coordination, the output bottleneck is structurally guaranteed.

Integration Hooks

Human Factors M2 (Fatigue and Decision Degradation). Boarding crises impose sustained cognitive and physical load on ED clinicians. An ED nurse assigned four patients — two actively being treated, two boarding and awaiting inpatient beds — is managing two fundamentally different care models simultaneously: acute emergency care and inpatient holding care. The context-switching degrades performance on both. Boarding does not just reduce ED bed capacity; it degrades the effective service rate of the remaining servers by fatiguing and fragmenting the attention of the humans who operate them. Staffing models that account for bed count but not cognitive load will systematically underestimate the throughput impact of boarding.

Workforce M1 (Workforce as Capacity Infrastructure). ED staffing models typically set nurse and provider counts based on expected patient volume and acuity. But these models treat the service rate as fixed — a given number of providers produces a given throughput. In reality, the service rate degrades as boarding increases, because providers spend time on boarding patients (medication administration, reassessments, family communication) that does not advance any ED patient toward disposition. A staffing model that does not account for boarding effectively overstates the ED’s service capacity by the fraction of provider time consumed by non-ED work. This is the same mechanism as the utilization-delay curve, but operating on the service-rate side: boarding does not just reduce bed supply; it reduces service supply.