Network Foundations

Module 4: Network Flow and System Connectivity Depth: Foundation | Target: ~2,500 words

Thesis: Healthcare delivery is a network of flows — patients, referrals, information, money — and network analysis reveals structural vulnerabilities invisible at the node level.

The Operational Problem

A health system director reviews her facilities: a regional hospital running smoothly, three critical access hospitals with adequate staffing, two FQHCs meeting their visit targets. Every node looks healthy. Then the region’s only cardiologist retires, and within six weeks, cardiac patients at four of those five facilities have nowhere to go. Transfers spike. Patients drive 90 miles to the next referral option. Two of the critical access hospitals lose a revenue stream they depended on. The system did not lose a facility. It lost an arc — a connection — and the consequences propagated across the entire network in ways that no facility-level dashboard predicted.

This is the core argument for network analysis in healthcare operations. Individual facilities and providers are nodes. The relationships between them — referrals, transfers, information exchange, funding flows — are arcs. The performance properties that matter most — reachability, fragility, throughput capacity, redundancy — are properties of the network, not of the nodes. You cannot see them by examining nodes one at a time, no matter how carefully.

What a Network Is: The Formal Object

In operations research, a network is not a metaphor. It is a mathematical object — a directed graph G = (N, A) — with computable properties.

Nodes (N) are the entities in the system. In a healthcare delivery network, nodes are facilities (hospitals, clinics, FQHCs, skilled nursing facilities), providers (specialists, primary care), administrative functions (grant offices, billing departments), or even patients when modeling individual care trajectories.

Arcs (A) are the directed connections between nodes. Each arc has a direction: a referral flows from a primary care provider to a specialist, not the reverse. Each arc has a capacity — the maximum flow it can carry per unit time. A cardiologist who can see 18 patients per week from referring facilities has a capacity of 18 on each inbound referral arc. An information reporting arc between a subgrantee and a grant administrator may have a capacity defined by the administrator’s processing bandwidth — say, 12 reports per month before review quality degrades.

Flow is the actual movement along arcs. Patient referrals, data transmissions, funding disbursements, lab results — all are flows. A flow must satisfy two constraints: it cannot exceed the arc’s capacity, and at every intermediate node (neither source nor sink), flow in must equal flow out. This is conservation of flow — the network equivalent of mass balance. Patients referred into a facility must either be seen there or referred onward. Funding received must be either spent or reported as unspent. Violations of conservation indicate leakage: patients lost to follow-up, funds unaccounted for, information that enters a node and never exits.

Why the formalism matters: Once a healthcare system is represented as a directed graph with capacities and flows, the entire toolkit of network optimization becomes available. Maximum throughput, weakest links, shortest paths, single points of failure, redundancy levels — all become computable rather than debatable. The argument shifts from “I think we have enough cardiology access” to “the maximum referral flow to cardiology services is 18 patients per week, and demand from the network is 24.”

Max-Flow / Min-Cut: The Throughput Is Set by the Weakest Link

The max-flow / min-cut theorem is the foundational result in network flow theory. Proved independently by L.R. Ford Jr. and D.R. Fulkerson (1956) and by P. Elias, A. Feinstein, and C.E. Shannon (1956), it states:

The maximum flow from a source to a sink in a network equals the minimum capacity of any cut separating the source from the sink.

A cut is a partition of the network’s nodes into two sets — one containing the source, one containing the sink — such that removing the arcs crossing the partition disconnects them. The capacity of a cut is the sum of the capacities of the arcs crossing it (in the source-to-sink direction). The minimum cut is the cut with the smallest total capacity.

The theorem’s operational meaning is direct: the maximum throughput of a care network is determined not by its total capacity but by its narrowest bottleneck. Adding capacity anywhere except at the min-cut does not increase system throughput.

Karl Menger’s theorem (1927), the graph-theoretic precursor, establishes the dual relationship between connectivity and cuts: the maximum number of node-disjoint paths between two vertices equals the minimum number of nodes whose removal disconnects them. In healthcare terms: the number of independent referral pathways a patient population has to reach specialty care equals the minimum number of providers or facilities whose loss would cut off access entirely.

The Ford-Fulkerson algorithm finds the max-flow (and implicitly the min-cut) by iteratively finding augmenting paths — routes with remaining capacity — and pushing flow along them until no more augmenting paths exist. The computational procedure maps to a practical diagnostic: start with the current referral flows, ask “can any additional patients be routed to this specialty service by any path?”, and when the answer is no, the arcs at capacity constitute the bottleneck.

What This Reveals in Practice

Consider a behavioral health referral network where three FQHCs refer patients to two outpatient providers and one crisis stabilization unit, which in turn can escalate to a regional inpatient psychiatric facility. If the crisis unit can process 8 patients per week and the inpatient facility admits 4, then no matter how many outpatient appointments are available, the acute care pathway is capped at 8 crisis evaluations and 4 admissions per week. The min-cut is at the crisis stabilization unit. Expanding outpatient capacity without addressing the crisis bottleneck produces a system that can identify more patients in need but cannot serve them when they decompensate — arguably a worse outcome than the status quo, because it increases the volume of visible, unmet acute need.

Shortest Path: Routing Patients Through the System

Not all paths through a care network are equal. A patient needing orthopedic surgery might be referred from a rural clinic to a regional hospital directly (one hop, 45 miles), or through an intermediate facility for imaging first (two hops, 30 + 60 miles, plus a 5-day wait for the imaging appointment). The shortest path problem asks: what is the minimum-cost route from origin to destination, where “cost” can be distance, time, number of handoffs, or any weighted combination?

Dijkstra’s algorithm (Edsger Dijkstra, 1956, published 1959) solves this for networks with non-negative arc weights. It works by iteratively selecting the unvisited node with the smallest known distance from the source, updating its neighbors’ distances, and marking it visited. The result is the shortest path from the source to every other node in the network.

In healthcare, the “weight” on each arc is rarely pure distance. A useful composite weight for patient routing might be:

Travel time (30 minutes by car)
Wait time to next available appointment (12 days for the specialist)
Handoff risk (each transfer introduces a probability of information loss or care discontinuity)

A shortest-path analysis using these composite weights often reveals that the geographically closest referral is not the fastest path to treatment. A specialist 90 miles away with a 3-day wait may produce faster time-to-treatment than one 20 miles away with a 6-week wait. Patient navigation programs that route by geography alone systematically underperform those that route by weighted path cost.

Critical path analysis — an extension of shortest path methods applied to project networks — identifies which tasks in a transformation program have zero slack. If a 988 crisis line implementation requires facility licensing (8 weeks), staff hiring (12 weeks), and IT integration (6 weeks), and hiring cannot begin until licensing is complete, the critical path is licensing + hiring = 20 weeks. IT integration has 14 weeks of slack. Delays to hiring propagate directly to the go-live date; delays to IT integration (up to 14 weeks) do not. This is the CPM (Critical Path Method), developed by Morgan Walker and James Kelley at DuPont in 1957.

Network Reliability: What Happens When a Node Fails

Network reliability analysis asks: what is the probability that the network remains connected — that patients can still reach needed services — given that nodes and arcs can fail?

The key concept is vertex connectivity — the minimum number of nodes whose removal disconnects the network. A network with vertex connectivity of 1 has a single point of failure: one node whose loss breaks the system. Menger’s theorem tells us that vertex connectivity equals the maximum number of node-disjoint paths between the most vulnerable pair of nodes.

In healthcare, single points of failure are disturbingly common:

The only cardiologist in a 5-county region. More than 86% of rural U.S. counties have zero cardiologists (American College of Cardiology, JACC, 2023). In those that have one, that provider is the network’s single point of failure for cardiac care. Retirement, illness, or relocation does not reduce capacity — it eliminates it.
The only lab capable of a specific test. If toxicology screening for a medication-assisted treatment program depends on a single reference lab, a processing delay at that lab halts the entire clinical workflow.
The sole grant administrator who understands the federal reporting system. When this person is unavailable, the entire reporting chain from subgrantees to the federal agency is severed. This is an information flow network with vertex connectivity of 1.

Redundancy is the network design response to reliability concerns. A network with two independent paths from every spoke to specialty care can survive the loss of any single intermediate node. But redundancy has costs: maintaining two cardiologists where demand supports one means one is underutilized, which creates its own financial and retention problems.

The engineering tradeoff is precise: each additional disjoint path between a population and a critical service increases vertex connectivity by 1, increasing the number of simultaneous failures the network can absorb. The cost of that redundancy is the cost of maintaining capacity on the redundant path. Network reliability analysis quantifies this tradeoff rather than leaving it to intuition.

Hub-and-Spoke vs. Distributed: Architecture Tradeoffs

Healthcare networks are not random graphs. They are designed — often implicitly — around one of two basic topologies.

Hub-and-spoke concentrates advanced services at a central hub (typically a regional hospital or academic medical center) with peripheral spokes (critical access hospitals, FQHCs, rural clinics) routing patients inward for complex care. Washington State’s Flex Program explicitly supports this model for rural health systems, encouraging CAH networks that regionalize specialty services at hub facilities.

The advantages are real: the hub accumulates volume, which supports specialization, quality, and financial viability. A surgeon who performs 200 procedures per year maintains sharper skills than one performing 40. Equipment investments are not duplicated. Training, credentialing, and compliance are centralized.

The fragility is equally real. Hub-and-spoke networks have low vertex connectivity by design — often 1 for the hub itself. If the hub is disrupted (a ransomware attack, a major staffing crisis, a natural disaster), every spoke loses access simultaneously. The network’s topology guarantees correlated failure across all peripheral sites. Research published in JACC Advances (2023) documented that hub inefficiencies, including prolonged lengths of stay, affect downstream capacity across the entire spoke network.

Distributed networks spread capabilities across multiple nodes, with each node capable of handling a broader range of services and multiple peer-to-peer referral paths. Vertex connectivity is higher. No single node failure cascades across the system. But distributed networks sacrifice specialization — each node handles lower volume — and coordination costs increase because there is no single authority setting standards.

The practical design question is not which topology to choose in the abstract. It is which services concentrate well (surgery, complex imaging, neonatal ICU) and which distribute well (primary care, behavioral health screening, chronic disease management). Most real healthcare networks are hybrids: hub-and-spoke for high-acuity specialty care, distributed for primary and preventive services. Network analysis makes the tradeoff explicit by computing connectivity, redundancy, and throughput under each design.

Healthcare Example: A 5-County Eastern Washington Network

Consider a stylized but realistic rural health network in eastern Washington:

Nodes: 1 regional hospital (hub, in the county seat), 3 critical access hospitals (CAHs in adjacent counties), 2 FQHCs (serving Medicaid and uninsured populations). The regional hospital has cardiology (1 cardiologist, capacity 18 referrals/week), orthopedics (2 surgeons, capacity 24/week), and behavioral health (3 outpatient providers, capacity 45 visits/week). The CAHs provide primary care and stabilization. The FQHCs provide primary care and behavioral health screening.

Arcs (referral flows):

Each CAH and FQHC refers cardiology patients to the regional hospital: 5 arcs, each carrying 3-6 referrals/week
Orthopedic referrals follow the same pattern: 5 arcs, 2-5 referrals/week
Behavioral health referrals flow both to the regional hospital and between FQHCs and CAHs: a denser sub-network with 8 arcs

Network analysis reveals:

Min-cut for cardiology: The single cardiologist is the min-cut. Total inbound referral demand is approximately 22 patients/week. Capacity is 18. The network is already over the max-flow threshold — 4 patients per week are being deflected, delayed, or lost to follow-up. Adding a sixth referring site (say, a new FQHC) without adding cardiology capacity would worsen the bottleneck without improving access.
Vertex connectivity for cardiology = 1. The cardiologist is a single point of failure. If this provider takes a 3-week vacation, cardiology access for 5 counties drops to zero. If this provider leaves permanently, the region’s cardiac patients face a 90-mile drive to the nearest alternative — assuming that cardiologist has capacity, which network analysis of the adjacent region could verify.
Behavioral health sub-network is more resilient. With 3 outpatient providers at the hub plus behavioral health screening at both FQHCs, the vertex connectivity for behavioral health is 2-3 depending on service level. The network can absorb the loss of one provider without disconnecting any population from care. This resilience is a function of the distributed topology for behavioral health, in contrast to the hub-and-spoke concentration for cardiology.
Shortest path analysis for orthopedics shows that patients from the most remote CAH (65 miles from the hub) face a 14-day wait for an orthopedic consult. If a telehealth arc were added — a virtual orthopedic consultation capability at the CAH — the effective path cost drops from (65 miles + 14 days) to (0 miles + 5 days) for initial evaluation, with only surgical cases requiring the physical transfer. The shortest path changes, and the min-cut shifts from the surgeon’s consultation capacity to the surgical schedule.

What the network analysis does that node-level analysis cannot: It reveals that the system’s most urgent vulnerability is not any facility’s performance — all facilities are meeting their individual metrics — but the structural dependence on a single cardiology provider, a property visible only at the network level.

Limitations: What Networks Do Not Model

Network analysis reveals topology and capacity, but it has specific blind spots that must be acknowledged:

No queueing dynamics. A network model shows that an arc has capacity 18 and flow 22, identifying the bottleneck. It does not show how long patients wait in the resulting queue, how wait times vary by day, or how the queue interacts with patient abandonment. For that, you need the queueing models from Module 2 applied at each congested node.

No human behavior at the nodes. A network model treats each node as a processing function with a capacity. It does not model the grant administrator who triages reports informally, the specialist who prioritizes certain referring physicians’ patients, or the clinic scheduler who routes around a provider’s known unavailability. Human factors — covered in the Human Factors discipline — determine how capacity is actually used at each node.

Static topology. Standard network models assume fixed nodes and arcs. Real healthcare networks are dynamic: providers join and leave, referral relationships strengthen or atrophy, telehealth arcs are added, facilities close. Temporal network analysis exists but adds substantial complexity.

No cost modeling. Max-flow finds the maximum throughput, not the minimum-cost flow. A minimum-cost flow formulation adds cost coefficients to each arc and optimizes for least-cost routing — relevant when payer mix, travel burden, or operational cost varies by path.

Network analysis is the structural skeleton. It must be combined with queueing theory for dynamic behavior at congested nodes, human factors for realistic node performance, and cost analysis for financial sustainability.

Integration Points

Human Factors Module 5: Human Error and Failure Modes. Every arc in a healthcare network is a handoff — a transfer of responsibility for a patient, a piece of information, or a task. Handoffs are where human error concentrates: information is lost, context is dropped, accountability is ambiguous. The network’s topology determines how errors propagate. In a hub-and-spoke network, an error at the hub (a lost referral, a misfiled report) affects every spoke. In a distributed network, the same error is contained locally. Network analysis identifies which handoff arcs carry the highest consequence of failure — those on the min-cut or the sole path between a population and a critical service — and those arcs should receive the highest investment in handoff reliability protocols.

Workforce Module 5: Organizational Design and Team Coordination. An organization chart is an information network. Reporting relationships are directed arcs. The span of control at each node is a capacity constraint. The topology determines coordination cost: a flat organization with many peer-to-peer arcs has high coordination overhead but low vertex distance (any node can reach any other in few hops). A deep hierarchy has low coordination overhead but high path length, creating delays in information flow and decision-making. Network analysis of organizational structure reveals the same properties — bottlenecks, single points of failure, min-cuts — that it reveals in patient flow networks.

Product Owner Lens

What is the operational problem? Healthcare networks have structural vulnerabilities — single points of failure, bottlenecks, disconnection risks — that are invisible when each node is assessed independently. Failures cascade in ways that node-level metrics do not predict.

What mechanism explains the system behavior? Network topology determines throughput (max-flow/min-cut), accessibility (shortest path), and resilience (vertex connectivity). These are computable properties of the directed graph, not subjective assessments.

What intervention levers exist?

Add capacity at the min-cut (recruit a second cardiologist, cross-train a nurse practitioner for cardiology triage)
Add arcs to increase connectivity (telehealth links, reciprocal referral agreements with adjacent regions, locum tenens contracts)
Redistribute services from hub-and-spoke to distributed topology where clinically appropriate (behavioral health screening, chronic disease management)
Formalize handoff protocols on high-consequence arcs (those on the min-cut or sole referral paths)

What should software surface?

Network topology map with real-time flow volumes on each arc, colored by utilization (flow/capacity ratio)
Min-cut identification: which providers, facilities, or connections, if lost, would reduce system throughput most?
Vertex connectivity score for each specialty service: how many simultaneous failures can the network absorb?
Shortest-path routing for patient referrals, weighted by wait time + travel time + handoff count, not geography alone
Alerts when a single provider or facility exceeds a threshold fraction of total network flow for any service line (e.g., “Dr. Torres handles 87% of cardiology referrals for the network”)

What metric reveals degradation earliest? The concentration ratio — the fraction of total network flow for a service that passes through a single node or arc. When one provider handles more than 70% of a specialty’s referrals, the network’s effective vertex connectivity for that specialty is approaching 1, regardless of how many other nodes nominally exist. This metric degrades gradually as referral patterns consolidate, providing months or years of warning before a retirement or departure triggers the access crisis that the concentration predicted.

Summary

A healthcare delivery system is a network in the precise, mathematical sense: nodes connected by directed arcs with capacities and flows. The max-flow/min-cut theorem (Ford-Fulkerson, 1956) establishes that the system’s throughput is set by its narrowest bottleneck, not its total capacity. Dijkstra’s algorithm finds the fastest path through the system when multiple routes exist. Menger’s theorem quantifies resilience as the number of independent paths between populations and services. Hub-and-spoke architectures concentrate expertise at the cost of fragility; distributed architectures increase resilience at the cost of specialization.

None of these properties are visible at the node level. A facility can meet every internal metric while sitting on a network with vertex connectivity of 1 — one retirement away from an access catastrophe. Network analysis makes the invisible visible, converting structural intuitions (“we’re too dependent on that one cardiologist”) into computable facts (“the cardiology min-cut capacity is 18/week against demand of 22/week, and vertex connectivity is 1”). That conversion — from intuition to computation — is what makes the difference between a system that reacts to failures and one that anticipates them.