Data-driven network diagnostics for optimizing POCUS management: an actionable hospital analytical approach Authors: Chengchen Xu, Chenchen Liu Published: 2026 ABSTRACT BackgroundPoint-of-care ultrasound (POCUS) is vital in critical care, yet hospitals struggle with inefficient resource allocation, leading to delays, staff burnout, and system fragility during demand surges. Static scheduling often inadequately addresses dynamic patient needs and hidden workflow bottlenecks.MethodsWe developed a three-tier, data-driven analytical approach to diagnose and optimize POCUS management. The study was conducted at Yixing People’s Hospital, a 1,700‑bed tertiary teaching hospital in Jiangsu, China. Using data from 2,281 POCUS examinations performed during the one‑year study period (2024), we modeled the POCUS system as a tripartite network (patients, departments, sonographers). Key clinical units involved included the intensive care unit (ICU, 34 beds), the emergency department (ED, 12 resuscitation beds plus a 15‑bed EICU), and the Department of Respiratory and Critical Care Medicine (81 beds). We integrated structural, temporal, and functional analyses to identify bottlenecks, forecast demand surges, and detect latent collaborative communities.ResultsThe network exhibited efficient but vulnerable “small‑world” properties, with over‑reliance on critical hubs (e.g., ICU, sonographer U2165). Time‑series analysis revealed periodic extreme surges coinciding with network contraction. Community detection exposed workload imbalances, with one community handling 57.8% of examinations.ConclusionsWe propose an actionable management analytical approach: (1) a demand‑responsive staffing strategy activated by forecast‑based alerts; (2) priority green channels for critical diagnostic pathways; and (3) functionally differentiated allocation based on community workload. This approach may enable hospital administrators to systematically balance efficiency with resilience, moving POCUS management from intuition toward evidence‑based optimization.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12913-026-14926-z. Introduction Point-of-care ultrasound (POCUS) has become a critical tool in high-acuity settings such as the intensive care unit and emergency department, providing critical real-time imaging for rapid clinical decision-making [1–3]. However, its growing use, combined with fluctuating patient demands, poses substantial operational challenges for hospital administrators [4–7]. Effectively matching limited sonographer and equipment resources to dynamic examination requests is essential to reduce delays, alleviate resource congestion, and ensure equitable staff workload distribution [8–10]. Current POCUS management faces three main structural shortcomings. First, dynamic demand often misaligns with static resource allocation. Fixed scheduling models cannot accommodate daily or seasonal variations, leading to backlogs during peak periods and underused capacity during slower times [11, 12]. This inefficiency is compounded by fragmented data systems that hinder responsive scheduling. Second, interdepartmental collaboration remains fragmented, creating systemic inefficiencies. Without effective coordination, departments may submit redundant examination requests, prolonging workflows and wasting valuable sonographer time and equipment [13, 14]. Third, resource allocation must balance equity and efficiency [15–17]. Over-serving high-demand departments may delay care elsewhere, whereas rigid parity can undermine throughput in critical areas, ultimately reducing system-wide performance. However, in dynamic clinical service environments such as POCUS, defining and quantifying fairness still requires careful contextual adaptation. While existing research has highlighted the need to optimize POCUS resources, most studies operate at a macro level, lacking quantitative analysis based on granular data. Furthermore, they have not adequately employed systematic modeling methods from operational management, such as network theory, to analyze the dynamic interdependencies among patients, departments, and sonographers. To bridge this gap, this study proposes a tripartite network framework that models the POCUS service delivery as a complex adaptive system. Integrating structural, temporal, and functional analyses, it addresses three core management questions: (1) Where are the key structural bottlenecks and critical dependency paths? (2) How does demand evolve temporally, and how can peak loads be forecast for proactive allocation? (3) How can allocation be optimized based on latent functional communities, beyond rigid departmental boundaries? By answering these, we develop a data-driven decision-support analytical approach, providing administrators with a roadmap to advance from empirical intuition to evidence-based, operational optimization. It is important to note that this study focuses on health workforce and operations management; the network analysis is used to guide personnel deployment and schedule design, not as an information technology implementation. Materials and methods Study setting and data collection The study was conducted at Yixing People’s Hospital, a 1,700-bed tertiary teaching hospital in Jiangsu, China. In 2024, the hospital had approximately 80,000 inpatient discharges and over 200,000 emergency visits. The intensive care unit (ICU) has 34 beds (including 3 negative-pressure isolation beds), the emergency department (ED) has 12 resuscitation beds and a 15-bed emergency intensive care unit (EICU), and the Department of Respiratory and Critical Care Medicine has two general wards (68 beds) and a 13-bed respiratory intensive care unit (RICU), totaling approximately 81 beds. Other clinical departments include general internal medicine, surgery, oncology, geriatrics, obstetrics and gynecology, pediatrics, and 49 clinical and technical departments in total. Patients treated at this hospital represent a general mixed adult population, covering medical, surgical, emergency, critical, geriatric, oncological, and obstetric cases. Point‑of‑care ultrasound (POCUS) examinations at our institution are performed exclusively by dedicated sonographers (ultrasound technicians) from the Department of Ultrasound Medicine. Physicians do not perform independent POCUS. The main types of POCUS exams included: FAST/eFAST (focused assessment with sonography in trauma), lung ultrasound (for pneumothorax, pleural effusion, and interstitial syndrome), deep vein thrombosis (DVT) screening, focused cardiac ultrasound (for pericardial effusion, global cardiac function), vascular access assessment, and abdominal fluid evaluation. These six categories accounted for over 95% of all examinations. Conceptual framework and model assumptions This study uses network analysis to represent the POCUS service delivery system as a dynamic, three‑part network (patients, departments, sonographers). Our analytical framework is based on three assumptions that link real‑world operations to network theory: Relational Abstraction Assumption: The interactions among patients, departments, and sonographers can be represented as a network with three types of nodes (patients, departments, sonographers). In this network, each connection (edge) links two nodes and is weighted by how often they collaborate. This abstraction allows us to use standard network measures to assess how the system works as a whole.Temporal Stationarity Assumption for Structural Analysis: While daily demand fluctuates, the core collaborative patterns between clinical departments and sonographer teams remain structurally stable over the mid-term. This permits a meaningful analysis of the static network topology to identify persistent dependencies.To assess the validity of this assumption, we performed a sensitivity analysis by constructing monthly subnetworks and comparing their core structure. The overlap of the top 10% of nodes ranked by betweenness centrality (a measure of how often a node serves as a bridge) across consecutive months was consistently > 85%, supporting the temporal stationarity assumption for the purpose of identifying persistent structural bottlenecks and dependencies.Functional Community Hypothesis: Clinical collaboration naturally organizes into functional modules (communities) that may transcend formal departmental boundaries. Detecting these empirically derived communities offers a more valid unit for resource planning than administrative silos. Relational Abstraction Assumption: The interactions among patients, departments, and sonographers can be represented as a network with three types of nodes (patients, departments, sonographers). In this network, each connection (edge) links two nodes and is weighted by how often they collaborate. This abstraction allows us to use standard network measures to assess how the system works as a whole. Relational Abstraction Assumption: The interactions among patients, departments, and sonographers can be represented as a network with three types of nodes (patients, departments, sonographers). In this network, each connection (edge) links two nodes and is weighted by how often they collaborate. This abstraction allows us to use standard network measures to assess how the system works as a whole. Temporal Stationarity Assumption for Structural Analysis: While daily demand fluctuates, the core collaborative patterns between clinical departments and sonographer teams remain structurally stable over the mid-term. This permits a meaningful analysis of the static network topology to identify persistent dependencies.To assess the validity of this assumption, we performed a sensitivity analysis by constructing monthly subnetworks and comparing their core structure. The overlap of the top 10% of nodes ranked by betweenness centrality (a measure of how often a node serves as a bridge) across consecutive months was consistently > 85%, supporting the temporal stationarity assumption for the purpose of identifying persistent structural bottlenecks and dependencies. Temporal Stationarity Assumption for Structural Analysis: While daily demand fluctuates, the core collaborative patterns between clinical departments and sonographer teams remain structurally stable over the mid-term. This permits a meaningful analysis of the static network topology to identify persistent dependencies.To assess the validity of this assumption, we performed a sensitivity analysis by constructing monthly subnetworks and comparing their core structure. The overlap of the top 10% of nodes ranked by betweenness centrality (a measure of how often a node serves as a bridge) across consecutive months was consistently > 85%, supporting the temporal stationarity assumption for the purpose of identifying persistent structural bottlenecks and dependencies. Functional Community Hypothesis: Clinical collaboration naturally organizes into functional modules (communities) that may transcend formal departmental boundaries. Detecting these empirically derived communities offers a more valid unit for resource planning than administrative silos. Functional Community Hypothesis: Clinical collaboration naturally organizes into functional modules (communities) that may transcend formal departmental boundaries. Detecting these empirically derived communities offers a more valid unit for resource planning than administrative silos. Guided by these assumptions, our framework examines the system along three dimensions—structural, temporal, and functional—to convert complex network data into stratified managerial insights (Fig. 1). Fig. 1Analytical framework for point-of-care ultrasound (POCUS) resource optimization. Note: The framework examines the POCUS system through three complementary dimensions: structural, temporal, and functional. Each dimension addresses specific management questions using quantitative analysis, and the results are combined to generate integrated optimization strategies Analytical framework for point-of-care ultrasound (POCUS) resource optimization. Note: The framework examines the POCUS system through three complementary dimensions: structural, temporal, and functional. Each dimension addresses specific management questions using quantitative analysis, and the results are combined to generate integrated optimization strategies Data sources and preprocessing The analysis was conducted using retrospective POCUS examination records extracted from the Hospital Information System (HIS) and the Picture Archiving and Communication System (PACS) at Yixing People’s Hospital for the period from January 1, 2024 to December 31, 2024. Following a standardized data quality control protocol to remove entries with missing critical fields, inconsistent timestamps, or duplicate identifiers, a final cohort of 2,281 valid records was retained for network modeling. Each record comprised the following structured fields, which were directly mapped to network components: (1) a unique anonymized patient identifier (n = 1,500), representing patient nodes; (2) the requesting clinical department (n = 49), representing department nodes; (3) a unique sonographer identifier (n = 24), representing sonographer nodes; (4) a binary examination result (positive/negative), used subsequently to construct the positive-diagnosis subnetwork; and (5) the examination timestamp, enabling the construction of time-resolved collaboration networks. Definition of collaboration frequency: In this study, “collaboration frequency” refers to the number of POCUS examinations performed by a sonographer for a specific requesting department. In our network models, each edge between a department node and a sonographer node is weighted by the total count of examinations that involved that department–sonographer pair. One examination equals one unit of weight. This definition applies to all network analyses, including the static tripartite network, the time‑resolved networks, and the positive‑diagnosis subnetwork. All personally identifiable information was irreversibly anonymized prior to analysis. This study was conducted in accordance with the Declaration of Helsinki. It was approved by the Ethics Committee of Yixing People’s Hospital (Approval Number: 2025-KY-190-01). Integrated analytical methodology Static network analysis: topological metrics and their managerial interpretation Why we used this method: To evaluate the inherent trade-off between efficiency and robustness in the POCUS system’s architecture, we first constructed and analyzed a static tripartite network (“patient–department–sonographer”). We calculated several network measures to see if the system is set up for rapid service or for withstanding disruptions. The network was visualized using Gephi (v0.10.1; Supplementary Fig. 1)[see Additional file]. Four measures were chosen for their practical relevance: Average path length: how many steps, on average, a request travels through the network—shorter means faster service. Network density: how well‑connected the system is—higher density means more backup routes. Clustering coefficient: how tightly groups of nodes work together; a high value may indicate “silos” (teams that work mostly within themselves). Betweenness centrality: how often a node serves as a bridge that many requests pass through—high values indicate bottlenecks. Detailed formulas are provided in Additional file (Eqs. 1–3). This initial analysis focused on the first three measures to get a basic understanding of the system. We then used betweenness centrality in Sect. “Positive-Diagnosis Subnetwork Analysis” to pinpoint critical pathways for severe cases. Small‑world verification: To confirm whether the observed high clustering coefficient and short average path length indicate a small‑world structure, we compared the real network with 100 Erdős–Rényi random graphs having the same number of nodes (1,573) and edges (3,869), using Python (v3.9.7) and NetworkX (v2.6.3). For each random graph, we computed the clustering coefficient and the average shortest path length (taking the largest connected component if disconnected). The mean values (C_rand, L_rand) and standard deviations were then used to calculate the small‑world index σ = (C_real / C_rand) / (L_real / L_rand). A σ > 1 indicates a small‑world network. Time series and anomaly detection We analyzed daily changes in exam volume using Python (v3.9.7) and NetworkX (v2.6.3). Daily examination demand was modeled as a time series covering all 366 days of 2024 (including days with zero examinations). To find days when demand was unusually high—so that we could prepare in advance—we used a standard statistical method (Z‑score, Eq. 4, Additional file) to detect anomalies. The mean (µ) and standard deviation (σ) of daily examination counts were calculated from the complete time series. Days with a daily count exceeding µ + 2σ were flagged as potential stress periods. Why this threshold: This cutoff is common in quality control; it catches true surges without too many false alarms. Checking reliability: Sensitivity analyses using thresholds of µ + 1.5σ and µ + 2.5σ confirmed that the core set of extreme-peak days was consistently identified. This confirms that the “network contraction” we observed under peak demand is a real pattern, not a random fluctuation. Positive-diagnosis subnetwork analysis Why we focused on positive findings: To improve care for the sickest patients, we need to know which departments and sonographers are most involved in severe cases. We therefore constructed a dedicated subnetwork comprising only positive-diagnosis cases. A “positive finding” was defined as an examination result indicating a condition requiring immediate intervention (e.g., significant pleural effusion, deep vein thrombosis, pneumothorax). From these positive cases, we built a two‑part network connecting departments and sonographers, where the number of shared positive exams determines the strength of each connection. We then calculated betweenness centrality for every node in this subnetwork. This tells us which departments and sonographers are most critical for the flow of severe cases, helping us target improvements where they matter most. Assessing stability of centrality rankings: To evaluate the robustness of the betweenness centrality rankings, we performed bootstrap resampling (1,000 iterations, sampling with replacement 80% of nodes from the positive‑diagnosis subnetwork). In each iteration, we recalculated betweenness centrality for the sampled nodes and recorded their ranks. For each node, we then computed the median rank and the 95% confidence interval (2.5th–97.5th percentiles) across all bootstrap samples. Community detection Finding natural work groups: To see how departments and sonographers naturally cluster together (beyond formal organizational charts), we built a combined network of all three node types (patients, departments, sonographers) and used a standard algorithm (Louvain) to detect communities. This algorithm groups nodes that have many connections within the group and few connections to outside groups—a property called modularity (Q). Checking stability: To assess the robustness of the detected communities, we ran the Louvain algorithm 50 times with different random seeds, using a resolution parameter γ = 1.0 (the default value that controls the granularity of the partition). The modularity Q across the 50 runs was 0.6608 ± 0.0001 (mean ± SD), indicating very high stability with respect to algorithmic randomness. Resolution sensitivity analysis: We further examined the effect of the resolution parameter by testing γ = 0.8 (coarser partitions) and γ = 1.2 (finer partitions). These settings produced 4 and 6 communities, respectively, but the core community (Community 3, comprising the Intensive Care Unit and its closely collaborating sonographers) remained consistent across all values of γ. This confirms that the identification of clinically relevant functional modules is robust to the choice of resolution parameter. Measuring group cohesion: For each detected community *c*, we computed an internal cohesion Index (Ic) (Eq. 5, Additional file). This index tells us how tightly the community works together—higher values mean more self‑contained groups. This helps us design different resource strategies for different types of communities. From descriptive analysis to prescriptive modeling: a bridge to future research The current study establishes a diagnostic, evidence-based framework. We outline how this foundation naturally extends toward prescriptive decision-support: Agent-Based Simulation (ABS): The validated community structure and collaboration rules can seed an ABS model to virtually test scheduling policies (e.g., elastic staffing, priority pathways) against stochastic demand forecasts. Integer Linear Programming (ILP): The quantified workload and demand patterns can be formulated as an ILP problem to optimize shift-by-shift resource deployment, subject to constraints derived from our findings (e.g., coverage for high-CB nodes, community-based demand). Results Static network analysis reveals an efficient yet vulnerable “small-world” structure The static tripartite network exhibited “small-world” (i.e., high efficiency but over-reliance on a few core nodes) characteristics (Table 1): a short average path length (L = 1.91) and a high clustering coefficient (C = 0.444), indicating efficient coordination but low redundancy (density D = 0.002). Comparison with 100 Erdős–Rényi random graphs of equal size (see Methods) confirmed a strong small‑world structure: the real network’s clustering coefficient was 142 times higher (0.444 vs. 0.00313 ± 0.00005) and its average path length was only 0.41 times as long (1.91 vs. 4.62 ± 0.09), yielding a small‑world index σ = 344 ( > > 1). Such a sparse, hub‑dependent architecture may constrain system flexibility and resilience, suggesting vulnerability to overload or failure of core nodes. Table 1Topological metrics of the static tripartite POCUS networkMetricNumber of NodesTotal EdgesNetwork DiameterNetwork Density(D)Clustering Coefficient(C)Average Path Length(L)Value1573386920.0020.4441.91Note. The key topological metrics collectively characterize a “small-world” network (high C, short L) that is efficient yet fragile (very low D). This structural profile underscores the inherent trade-off between operational efficiency and systemic robustness in the current POCUS resource allocation system. Small‑world index σ = 344 (compared to 100 equivalent random networks) Table 1Topological metrics of the static tripartite POCUS networkMetricNumber of NodesTotal EdgesNetwork DiameterNetwork Density(D)Clustering Coefficient(C)Average Path Length(L)Value1573386920.0020.4441.91Note. The key topological metrics collectively characterize a “small-world” network (high C, short L) that is efficient yet fragile (very low D). This structural profile underscores the inherent trade-off between operational efficiency and systemic robustness in the current POCUS resource allocation system. Small‑world index σ = 344 (compared to 100 equivalent random networks) Topological metrics of the static tripartite POCUS network Note. The key topological metrics collectively characterize a “small-world” network (high C, short L) that is efficient yet fragile (very low D). This structural profile underscores the inherent trade-off between operational efficiency and systemic robustness in the current POCUS resource allocation system. Small‑world index σ = 344 (compared to 100 equivalent random networks) Implications for practice (hypothesis‑generating): The observed “small‑world” architecture implies that while the current system is efficient for daily operations, it could be vulnerable to overload or failure of a few key hubs (e.g., the ICU and specific sonographers). This finding suggests a potential need to build strategic redundancy, though further prospective studies would be required to confirm causality. Temporal analysis identifies demand anomalies and systemic overload Time-series analysis reveals extreme demand surges—sporadic peaks that far exceed baseline levels (Fig. 2). Applying the Z-score anomaly detection method (threshold = µ + 2σ, corresponding to 15.0 daily examinations), we identified six days of anomalously high demand (Table 2). Collaboration frequency peaked at 33 examinations on November 17, 2024 (Z-score = 6.10). A detailed analysis of the 3-day window surrounding this peak event (Table 3) showed a 44.9% contraction in network scale (both nodes and edges) concurrent with a 211.1% surge in the maximum collaboration frequency recorded for a single node. This pattern is consistent with a “resource overload” response: under peak demand pressure, the network appeared to shed peripheral, non‑core collaborations and concentrate workload onto a limited set of established core pathways. This observation suggests that static scheduling models may have important limitations during major demand surges. Fig. 2Temporal variation in daily POCUS collaboration frequency (2024). Note: The blue line shows the daily number of department–sonographer collaborations (exams/day); the orange dashed line is a 7‑day moving average. Red dots indicate weekends Temporal variation in daily POCUS collaboration frequency (2024). Note: The blue line shows the daily number of department–sonographer collaborations (exams/day); the orange dashed line is a 7‑day moving average. Red dots indicate weekends Table 2Days with anomalously high POCUS demand identified in 2024Date(2024)Daily Examinations (Xt)Z-score(Zt)Deviation from Mean2024-03-08162.23+ 2.23σ2024-05-16172.45+ 2.45σ2024-07-12182.68+ 2.68σ2024-11-17336.10+ 6.10σ2024-11-18203.14+ 3.14σ2024-11-20254.28+ 4.28σNote. Anomalous days were identified using the Z-score method with a threshold of µ + 2σ (15.0 examinations/day). The mean (µ) and standard deviation (σ) of daily examinations are calculated from all 366 days of 2024 (including days with zero examinations), yielding µ = 6.23, σ = 4.39. The peak on 2024-11-17 (Z = 6.10) represents an extreme outlier, providing a critical case for analyzing system adaptation under stress Table 2Days with anomalously high POCUS demand identified in 2024Date(2024)Daily Examinations (Xt)Z-score(Zt)Deviation from Mean2024-03-08162.23+ 2.23σ2024-05-16172.45+ 2.45σ2024-07-12182.68+ 2.68σ2024-11-17336.10+ 6.10σ2024-11-18203.14+ 3.14σ2024-11-20254.28+ 4.28σNote. Anomalous days were identified using the Z-score method with a threshold of µ + 2σ (15.0 examinations/day). The mean (µ) and standard deviation (σ) of daily examinations are calculated from all 366 days of 2024 (including days with zero examinations), yielding µ = 6.23, σ = 4.39. The peak on 2024-11-17 (Z = 6.10) represents an extreme outlier, providing a critical case for analyzing system adaptation under stress Days with anomalously high POCUS demand identified in 2024 Note. Anomalous days were identified using the Z-score method with a threshold of µ + 2σ (15.0 examinations/day). The mean (µ) and standard deviation (σ) of daily examinations are calculated from all 366 days of 2024 (including days with zero examinations), yielding µ = 6.23, σ = 4.39. The peak on 2024-11-17 (Z = 6.10) represents an extreme outlier, providing a critical case for analyzing system adaptation under stress Table 3Network adaptation during the peak-load anomaly (2024-11-17)Metric3-Day Period Before Anomaly (Daily Average)3-Day Period After Anomaly (Daily Average)Change (After vs. Before)Number of Nodes18.710.3↓44.9%Number of Edges18.710.3↓44.9%Average Collaboration Frequency5.15.0↓2.0%Maximum Collaboration Frequency9.028.0↑211.1%Note: The 3-day windows represent daily averages centered on the peak day (2024-11-17). The data show a “shrink-to-core” adaptation pattern: a sharp contraction in overall network scale alongside an extreme concentration of workload on a limited set of core pathways. This pattern highlights limitations of static scheduling when facing demand surges Table 3Network adaptation during the peak-load anomaly (2024-11-17)Metric3-Day Period Before Anomaly (Daily Average)3-Day Period After Anomaly (Daily Average)Change (After vs. Before)Number of Nodes18.710.3↓44.9%Number of Edges18.710.3↓44.9%Average Collaboration Frequency5.15.0↓2.0%Maximum Collaboration Frequency9.028.0↑211.1%Note: The 3-day windows represent daily averages centered on the peak day (2024-11-17). The data show a “shrink-to-core” adaptation pattern: a sharp contraction in overall network scale alongside an extreme concentration of workload on a limited set of core pathways. This pattern highlights limitations of static scheduling when facing demand surges Network adaptation during the peak-load anomaly (2024-11-17) Note: The 3-day windows represent daily averages centered on the peak day (2024-11-17). The data show a “shrink-to-core” adaptation pattern: a sharp contraction in overall network scale alongside an extreme concentration of workload on a limited set of core pathways. This pattern highlights limitations of static scheduling when facing demand surges Implications for practice (hypothesis‑generating): The network contraction observed during peak demand—rather than expansion—may indicate that static scheduling is ill‑suited to handle extreme loads. This finding supports the consideration of forecast‑driven, flexible scheduling strategies, but these remain to be tested in prospective or simulation‑based settings. Positive-diagnosis subnetwork identifies critical clinical pathways and structural bottlenecks Analysis of betweenness centrality within the positive‑diagnosis subnetwork identified the Intensive Care Unit (ICU) and sonographer U2165 as the highest‑ranked nodes (Table 4). To assess the stability of these rankings, we performed bootstrap resampling (1,000 iterations, sampling 80% of nodes each time). The ICU had a median bootstrap rank of 5.0 (95% CI: 1–15), and sonographer U2165 had a median rank of 3.0 (95% CI: 1–13), confirming that both nodes are consistently among the top ranks across bootstrap samples. In contrast, lower‑ranked nodes (e.g., Emergency Medicine Department, Respiratory & Critical Care Medicine I) exhibited wide rank variability (see Supplementary Table S1 for full results), suggesting that only the top two nodes represent stable structural bottlenecks. The collaborative pathway between ICU and U2165 therefore appears to be the most critical and reliable conduit for severe‑case workflows. Table 4Key bridge nodes in the positive-diagnosis subnetwork ranked by betweenness centralityCategoryRankNodeBetweenness Centrality(CB)Departments1Intensive Care Unit (ICU)0.40712Neurosurgery ICU0.15143Emergency Medicine Department0.07894Respiratory & Critical Care Medicine I0.07255Respiratory & Critical Care Medicine ICU0.0585Sonographers1U21650.47812U23530.47063U18570.47034U20800.47035U19080.4684Note: Betweenness centrality identifies nodes critical for high-value diagnostic pathways. The highest-ranked nodes (U2165 and ICU) are primary structural bottlenecks. Bootstrap resampling (1,000 iterations, 80% node resampling) confirmed their stability: median rank ICU 5.0 (95% CI 1–15), U2165 3.0 (1–13). Ranks 2–5 showed greater variability (see Supplementary Table S1), indicating that only the top two nodes are robust bottlenecks Table 4Key bridge nodes in the positive-diagnosis subnetwork ranked by betweenness centralityCategoryRankNodeBetweenness Centrality(CB)Departments1Intensive Care Unit (ICU)0.40712Neurosurgery ICU0.15143Emergency Medicine Department0.07894Respiratory & Critical Care Medicine I0.07255Respiratory & Critical Care Medicine ICU0.0585Sonographers1U21650.47812U23530.47063U18570.47034U20800.47035U19080.4684Note: Betweenness centrality identifies nodes critical for high-value diagnostic pathways. The highest-ranked nodes (U2165 and ICU) are primary structural bottlenecks. Bootstrap resampling (1,000 iterations, 80% node resampling) confirmed their stability: median rank ICU 5.0 (95% CI 1–15), U2165 3.0 (1–13). Ranks 2–5 showed greater variability (see Supplementary Table S1), indicating that only the top two nodes are robust bottlenecks Key bridge nodes in the positive-diagnosis subnetwork ranked by betweenness centrality Note: Betweenness centrality identifies nodes critical for high-value diagnostic pathways. The highest-ranked nodes (U2165 and ICU) are primary structural bottlenecks. Bootstrap resampling (1,000 iterations, 80% node resampling) confirmed their stability: median rank ICU 5.0 (95% CI 1–15), U2165 3.0 (1–13). Ranks 2–5 showed greater variability (see Supplementary Table S1), indicating that only the top two nodes are robust bottlenecks Implications for practice (hypothesis‑generating): The ICU‑U2165 pathway may represent an important bottleneck for severe cases. Ensuring its robustness and efficiency could be a priority, but targeted interventions (e.g., dedicated resources, priority scheduling, backup planning) should first be evaluated in implementation research. Community detection reveals structural imbalances in resource allocation Community detection applied to the collaborative hypernetwork yielded a well-defined partition with modularity (Q = 0.662), segmenting the system into five distinct functional communities (Fig. 3). The modularity was highly stable across 50 algorithm runs (mean Q = 0.6608, SD = 0.0001). Resolution sensitivity analysis (γ = 0.8 and 1.2) produced 4 and 6 communities, respectively, but the core community (Community 3, comprising the Intensive Care Unit and its closely collaborating sonographers) remained consistent, confirming the robustness of the detected structure. Fig. 3Community detection results showing the partition of the collaborative network into five functional communities (modularity Q = 0.662). Note: Node colors indicate community membership, and node sizes are proportional to service volume (total number of examinations handled by that node) Community detection results showing the partition of the collaborative network into five functional communities (modularity Q = 0.662). Note: Node colors indicate community membership, and node sizes are proportional to service volume (total number of examinations handled by that node) The distribution of service load across these communities was profoundly uneven (Table 5). Community 3 emerged as the dominant core, accounting for 1,318 service instances—3.7 times the volume of the next largest community (Community 4). This concentration of workload, supported by strong internal connectivity (22.47), points to a cluster of high efficiency but also a potential risk of overload. The system’s centralization was further reflected by the role of a few core sonographers; for instance, U2165 served as a key contributor across four of the five communities, spanning diverse clinical areas from intensive care to obstetrics. This pattern suggests that over‑reliance on specific individuals could pose a threat to workflow stability. In contrast, Community 1 involved the greatest number of departments (n = 26) but exhibited the lowest functional cohesion (0.56) and weakest internal connectivity (0.80), indicating more fragmented and potentially less efficient collaboration. Significant variation in service frequency among sonographers across different communities (Fig. 4) further suggests that the current allocation system may be associated with workload imbalances. Table 5Characteristics and workload distribution of the five functional communities detected in the POCUS collaboration networkCommunity IDFunctional ProfileNumber of DepartmentsCore Sonographers (Top 3 by Service Count)Internal Connection StrengthExternalConnection StrengthCohesion Index(Ic)Number of SonographersServices Volume0Obstetrics & Perinatal Management3U2165: 18, U7301: 11, U2353: 93.000.400.88218771Surgical & Emergency Care26U2165: 38, U2080: 28, U1857: 280.800.620.560212962Geriatric & Chronic Disease Management3U2353: 40, U1908: 27, U1907: 2711.001.100.909142363Intensive & Critical Care5U2165: 184, U1908: 165, U7301: 15822.470.600.9742113184Internal Medicine & Oncology12U1857: 59, U2165: 48, U2353: 472.490.530.82317354Note: Functional communities were identified via the Louvain algorithm (50 runs, γ = 1.0; modularity Q = 0.6608 ± 0.0001). Sensitivity analyses (γ = 0.8,1.2) produced 4 and 6 communities, but the core community (Community 3) remained consistent. The Cohesion Index (Ic) = Internal / (Internal+External). Community sorted by service volume: Community 3 handles 57.8% of exams (Ic=0.974) indicating high efficiency but potential overload risk. Sonographer U2165 appears across multiple communities, underscoring a critical dependence on individual expertise Table 5Characteristics and workload distribution of the five functional communities detected in the POCUS collaboration networkCommunity IDFunctional ProfileNumber of DepartmentsCore Sonographers (Top 3 by Service Count)Internal Connection StrengthExternalConnection StrengthCohesion Index(Ic)Number of SonographersServices Volume0Obstetrics & Perinatal Management3U2165: 18, U7301: 11, U2353: 93.000.400.88218771Surgical & Emergency Care26U2165: 38, U2080: 28, U1857: 280.800.620.560212962Geriatric & Chronic Disease Management3U2353: 40, U1908: 27, U1907: 2711.001.100.909142363Intensive & Critical Care5U2165: 184, U1908: 165, U7301: 15822.470.600.9742113184Internal Medicine & Oncology12U1857: 59, U2165: 48, U2353: 472.490.530.82317354Note: Functional communities were identified via the Louvain algorithm (50 runs, γ = 1.0; modularity Q = 0.6608 ± 0.0001). Sensitivity analyses (γ = 0.8,1.2) produced 4 and 6 communities, but the core community (Community 3) remained consistent. The Cohesion Index (Ic) = Internal / (Internal+External). Community sorted by service volume: Community 3 handles 57.8% of exams (Ic=0.974) indicating high efficiency but potential overload risk. Sonographer U2165 appears across multiple communities, underscoring a critical dependence on individual expertise Characteristics and workload distribution of the five functional communities detected in the POCUS collaboration network Note: Functional communities were identified via the Louvain algorithm (50 runs, γ = 1.0; modularity Q = 0.6608 ± 0.0001). Sensitivity analyses (γ = 0.8,1.2) produced 4 and 6 communities, but the core community (Community 3) remained consistent. The Cohesion Index (Ic) = Internal / (Internal+External). Community sorted by service volume: Community 3 handles 57.8% of exams (Ic=0.974) indicating high efficiency but potential overload risk. Sonographer U2165 appears across multiple communities, underscoring a critical dependence on individual expertise Fig. 4Workload distribution across the five functional communities. Note: Bars represent total service frequency (examinations handled per community); the line shows per‑capita frequency (examinations per sonographer). Community legend: 0 = Obstetrics & Perinatal; 1 = Surgery & Emergency; 2 = Geriatrics & Chronic Disease; 3 = Intensive & Critical Care (ICU/CC); 4 = Internal Medicine & Oncology Workload distribution across the five functional communities. Note: Bars represent total service frequency (examinations handled per community); the line shows per‑capita frequency (examinations per sonographer). Community legend: 0 = Obstetrics & Perinatal; 1 = Surgery & Emergency; 2 = Geriatrics & Chronic Disease; 3 = Intensive & Critical Care (ICU/CC); 4 = Internal Medicine & Oncology Implications for practice (hypothesis‑generating): The observed workload imbalance across functional communities and the cross‑community reliance on key sonographers (e.g., U2165) suggest that uniform resource distribution based on administrative departments may be inefficient. A shift toward differentiated resource allocation—tailored to community function and load—could be considered, along with support or backup plans for overloaded core personnel. These strategies are hypothesis‑generating and require validation through simulation or controlled operational experiments. Discussion This study used network analysis to explore possible drivers of resource imbalance in POCUS services by examining structural, temporal, and functional dimensions. Rather than establishing causality, our approach provides a diagnostic and hypothesis‑generating analytical framework that quantifies inefficiencies and suggests potential leverage points for intervention. This aligns with recent calls to apply complexity science to health care management [18, 19] and offers a data‑based foundation for moving from intuition toward evidence‑informed planning. Building structural resilience: identifying core dependencies and strategic redundancy Our static network analysis showed that the POCUS system has a “small‑world” structure [20–24]: it appears efficient but relies heavily on a few key hubs (the ICU and sonographer U2165), with very few backup connections (density = 0.002). This pattern may make the system vulnerable—if one hub becomes overloaded or fails, the whole system could be affected [25, 26]. Based on these observations, we suggest two hypothesis‑generating strategies that could be explored in future research or simulation studies: A/B team backup for key sonographers: For a core sonographer like U2165, training a backup person (“B”) with equivalent skills in the same critical tasks (e.g., ICU exams) might reduce the risk of disruption. Formalizing handover protocols and shared schedules could help ensure continuity when the primary sonographer is overloaded or absent. Cross‑community skill training: Building a reserve of versatile sonographers by training them to handle high‑demand exams across several work groups (e.g., both intensive care and internal medicine) could increase redundancy and allow flexible deployment during surges or staff shortages Enabling dynamic scheduling: a forecast-based elastic capacity adjustment model The time‑series analysis found occasional but extreme spikes in demand that coincided with a contraction of the network—a pattern suggesting that static scheduling may not be well suited to these surges [27–30]. However, other factors could also contribute to the observed contraction, such as changes in staff shift patterns, departmental prioritization policies (e.g., delaying non‑urgent exams), data recording artifacts (e.g., delayed entry during busy periods), or clinical triage decisions that route severe cases to core sonographers. Our data cannot distinguish among these mechanisms, and future studies with real‑time workflow logs would be needed to clarify the drivers. Regardless, the pattern highlights potential limitations of static scheduling under extreme demand. To respond proactively, we propose a two‑stage model that is hypothesis‑generating and would need validation: Prediction and tiered alerts: Using the Z‑score method described earlier, the system could automatically generate alerts when forecasted demand exceeds the µ + 2σ threshold (15.0 exams/day). Level‑1 alert (forecast > 15.0): Consider activating a mobile “surge” POCUS team to increase capacity by 10‑20%. Level‑2 alert (Z‑score > 4.0): Consider contingency plans, such as rescheduling non‑urgent outpatient exams to other times or using alternative imaging methods [31, 32]. Dynamic response: Once an alert is triggered, a decision rule could determine how to reallocate staff and equipment based on current availability, equipment status, and patient priority. A potential enabler is a centralized scheduling pool [33–35]—a hospital‑wide system that would replace fixed departmental assignments with flexible, demand‑driven deployment of sonographers and equipment. Optimizing critical pathways: enhancing flow in high-value diagnostic channels The analysis of positive‑finding cases identified the ICU–U2165 connection as a potential bottleneck for severe cases. Improving this single pathway might have a substantial impact on the whole system [36, 37]. We suggest three specific actions that could be tested in implementation research: Green channels for urgent requests: Priority rules in the scheduling system could automatically route requests from the ICU and other high‑acuity units to designated, experienced sonographers, potentially reducing waiting time. Delegate non‑diagnostic tasks: Assigning support staff (nurses or aides) to core sonographers to handle paperwork, equipment setup, and patient positioning might free the sonographers to focus on image acquisition and interpretation. Standardize high‑value exams: Developing standard operating procedures (SOPs) for key exams (e.g., “ICU Rapid Ultrasound for Shock”) could reduce variability, improve reliability, and provide a baseline for ongoing quality improvement using methods like Plan‑Do‑Study‑Act [38]. Implementing differentiated allocation: resource deployment based on functional communities The community analysis showed a substantial workload imbalance (Community 3 handled 3.7 times as many exams as the next largest community) and wide variation in how tightly each community worked together. This suggests that assigning resources simply by department name may be inefficient. Instead, we propose different hypothesis‑generating strategies for different community types: For high‑load, tightly‑knit communities (e.g., Community 3 – ICU/CC): A “dedicated team” model could be considered, where specific sonographer‑equipment pairs are assigned to this community, protected scheduling blocks are set aside, and priority is given for hospital‑wide resources. This might ensure stable, high‑volume throughput. For loosely‑connected, fragmented communities (e.g., Community 1 – Surgery/Emergency): A “flexible pool” model might be more appropriate. Sonographers from these communities could be combined into a hospital‑wide pool that can be sent to high‑demand areas—including core communities during surges—using optimization algorithms. This could maximize overall equipment use and system flexibility. Synthesis: from diagnostic insights to a prescriptive management analytical Collectively, our analyses converge into an integrated, three‑tier decision‑support framework that is diagnostic and hypothesis‑generating. The Diagnostic Tier uses network metrics to map structural bottlenecks, temporal volatility, and functional communities. The Predictive & Strategic Tier translates diagnostics into an Elastic Capacity Model and a Community‑Informed Allocation Framework. The Prescriptive & Operational Tier would require prospective testing of specific interventions such as green channels and strategic redundancy. This framework provides a systematic roadmap for moving from reactive intuition to proactive, evidence‑informed optimization, but all proposed strategies should be validated in simulation or implementation studies before real‑world adoption. Limitations and future research This study has several limitations. First, this is a single‑center study at a tertiary teaching hospital in China. The POCUS delivery model at our site is centralized (sonographer‑performed), which differs from the “democratic” clinician‑performed model common elsewhere. Regarding staffing and training, our hospital operates a rotating shift system for sonographers, with daily on‑duty numbers typically ranging from 6 to 10 depending on workload; all sonographers completed an internal POCUS competency certification (including theoretical and practical assessments) as required by our hospital. The findings may not directly generalize to community hospitals or health systems in other countries with different governance models, staffing levels, or resource allocation policies. Future multi‑center studies with larger and more diverse populations are needed to validate our findings. Second, the number of POCUS examinations (2,281) is modest, which limits the statistical power for certain subgroup analyses. However, because this dataset represents the complete annual workload of the hospital (capturing all patients, departments, and sonographers), the network constructed for analysis still contains 1,573 nodes and 3,869 edges—sufficient for stable topological metrics (e.g., average path length, clustering coefficient) and meaningful community detection. Nevertheless, the modest examination count implies that our findings may not capture rare or seasonal patterns that could emerge in larger datasets from busier centers. Third, bootstrap resampling (1,000 iterations) showed that only the top‑two ranking nodes (ICU and U2165) are robust bottlenecks; lower‑ranked nodes exhibited considerable variability (see Supplementary Table S1). Therefore, our managerial recommendations focus primarily on the ICU–U2165 pathway, and other lower‑ranked nodes should be interpreted with caution. Fourth, our model is diagnostic and predictive but does not generate real‑time optimal schedules or simulate policy impacts under uncertainty. Future research should build on these findings in several directions: (1) combine network diagnostics with cost‑effectiveness analysis (e.g., comparing the incremental cost of A/B backup teams vs. reduced delay penalties); (2) develop an agent‑based simulation to test the proposed elastic scheduling and differentiated allocation strategies under various surge scenarios; and (3) formulate an integer programming model to compute mathematically optimal resource assignment schedules using the quantified workload disparities and critical pathway constraints. Integrating real‑time data on equipment status and clinician skill levels would further enhance fidelity, paving the way for a closed‑loop decision‑support system. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2