Emergency Medical Service (EMS) responders operate in highly variable situations with unfamiliar environments, diverse patient conditions, and varying protocols. This problem is exacerbated for infrequent high-acuity pediatric cases which exhibit a high frequency of dosing errors of +/- 20% the intended dose. This study utilizes time-study methodology to analyze the effects of using a newly developed AR cognitive aid for administering medication in 24 high-fidelity simulations of newborn cardiac arrest scenarios. The goal was to identify and assess differences and variations in time to dose, medication administration procedures, and errors between EMS crews using current cognitive aids versus crews using the AR cognitive aid. The introduction of the AR cognitive aid resulted in a drastic reduction in the dosing error rate from 83.3% to 8.3% but increased the time to medication administration by 55%. The mean time to medication administration using the AR cognitive aid remained below the accepted 10-minute threshold for positive cardiac arrest patient outcomes. Variabilities in the task sequencing and omission were reduced with the AR cognitive aid. The authors believe that this reduction in procedural variability played a significant role in reducing dosing errors. Further work is necessary to fully harness the benefits of such an AR cognitive aid for EMS. These efforts should focus on 1) Evaluating the balance between dosing accuracy and time to dose for optimal patient outcomes, 2) Further developing EMS best practices and standard operating procedures for high-acuity pediatric cases, and 3) Expanding the capabilities of the AR cognitive aid beyond pediatrics.
We present a framework for constructing individual-level metamodels of complex simulations to support rapid, data-driven decision-making. By training models at the individual level and aggregating predictions, our framework enables flexible representation of any target population and facilitates comparison of alternative strategies. We explore regression-based and machine learning-based approaches and demonstrate the framework's utility in a colorectal cancer screening case study spanning diverse scenarios and interventions. Outputs such as cancer cases averted and life years lost are accurately predicted, with metamodels closely reproducing simulation results while substantially reducing computational burden. We further show how metamodel outputs can support cost-effectiveness comparisons and guide intervention selection without additional simulation runs. Although demonstrated in a colorectal cancer screening context, the proposed framework can be used in many domains and can be applied to simulations that produce individual-level outputs. Overall, our framework provides a scalable approach for accelerating policy evaluation and strategy comparison across diverse healthcare and operations environments.
School-based asthma programs are a proven vehicle for reaching underserved children, yet program impact hinges on where to send scarce mobile-clinic days and how to allocate appointments. We formulate a two-level, finite-horizon capacity allocation problem that jointly selects schools and schedules patients within schools, integrating a pediatric asthma state-transition model with operational constraints. We present a mixed-integer programming (MIP) formulation, a sub-second index policy with myopic look-ahead, and a rolling-horizon MIP relaxation, and evaluate them using data from a large urban district (89 schools; 5 975 children with asthma). Across horizons and capacities, the optimal policy balances treatment for the sickest with prevention for moderate illness, yielding marked improvements in health: as monthly service-day capacity rises, the share of children in mild illness increases (e.g., 30% -> 74%) while the worst-illness share declines (e.g., 38% -> 7%). The index policy achieves 0.2-1.5% gaps with orders-of-magnitude faster run times, enabling practical deployment. We discuss design guidelines for mobile asthma programs-including capacity sizing, planning horizon, and school prioritization-and show how the framework extends to other community-based chronic care models.
The growing prevalence of critical illness presents an urgent public health challenge, exacerbated by the aging population and associated comorbidities. This has led to an increased demand for intensive care, alongside rising complexity in managing critically ill patients and ongoing strain on the healthcare system with limited monetary and personnel resources. This research aims to develop novel multi-method simulation, hybridizing discrete event and agent-based models with AI-informed agents for supporting operational decision making in the intensive care unit (ICU) setting. We develop an AI pipeline to cluster, model, and predict clinical trajectory phenotypes that a critically ill patient should follow. We then integrate these models in the simulation environment, allowing for runtime evaluation of AI models to drive simulated agent behavior. The data-driven patient trajectory modeling and prediction enable us to capture the interplay between patients and clinicians, thereby providing a refined assessment of clinician workload and ICU resource utilization. A case study is developed based on data from one large American academic hospital system. Using the validated model, resource management under patient surge and dynamic nurse staffing was investigated through what-if analyses, demonstrating the effectiveness of this model to support ICU operational decisions.
In this study, a resilient inventory management model with minimized cost for the Chimeric Antigen Receptor (CAR) T-cell therapy production process is presented. While current autologous and allogeneic CAR T-cell therapy manufacturing follows a strict pull-based, patient-specific flow that leads to high cost, long vein-to-vein time, and limited efficiency, this research shows that integrating periodic inventory planning with lean manufacturing principles can improve both efficiency and operational resilience through reduced shortages, improved supply continuity, and greater responsiveness. Using demand forecasts generated through ridge regression on global health indicators, two tailored planning models were formulated: an allogeneic model incorporating holding, expiry, and shortage costs under shelf-life and capacity constraints, and an autologous model reflecting patient-specific manufacturing through a simplified planning structure. The models are optimized over 52 weekly periods and evaluated through sensitivity analysis. The results show that the shift from the autologous configuration to the allogeneic buffered configuration is the main driver of performance improvement, reducing total cost to approximately one-third of the autologous baseline and shortening vein-to-vein time from 35 days to about 9 days. These improvements enhance process flexibility and mitigate bottlenecks without compromising biological and regulatory constraints. The proposed framework offers a scalable decision-support tool for biotherapeutic production systems and provides a practical case study for applying inventory management and optimization in highly constrained, perishable, and high-value healthcare supply chains. This work highlights the potential of bridging industrial engineering and medicine to improve accessibility, efficiency, and operational resilience in CAR T-cell therapy.
We study the vaccine distribution problem over a three-echelon network during a sudden pandemic outbreak. We consider uncertainties in vaccine supply due to delays in production, disruption in global supply chain, or lockdown policies in the early phases of the pandemic outbreak. We also consider limited resources including cold trucks for vaccine transportation and healthcare workers for vaccine administration. The problem is to dynamically allocate these limited resources over the distribution network in response to periodic and uncertain vaccine supply, such that vaccine can be transported and administered in a timely and cost-effective manner. We formulate a dynamic programming model for this problem and propose solution algorithms based on approximate dynamic programming. We conduct a case study on a vaccine distribution network in Iran and report computational results that demonstrate the superiority of our solution approach over three benchmark approaches. We also perform various sensitivity analyses and provide useful insights into the benefits of more transportation and human resources, and the impacts of vaccine shelf life and vaccine supply amount.
As the center of hospital operations, operating room (OR) scheduling directly shapes patient safety, operational efficiency, and system resilience. In practice, OR scheduling is challenged by uncertain demand fluctuations and coordination difficulties both within and across departments. To address these challenges, and motivated by collaboration with a large tertiary hospital, we propose the Surgical Pool-Augmented, Capacity-Embedded (SPACE) framework to achieve demand smoothing and enhance intra- and inter-departmental coordination. Within the SPACE framework, we worked closely with our partner hospital to redesign the OR scheduling process by introducing a surgical pool mechanism, which enables proactive management of surgical postponements and helps reduce daily demand variability. The redesigned process combines several methodological tools. Specifically, machine learning is applied to improve the accuracy of surgery duration prediction, while an OR scheduling optimization model introduces structured flexibility that facilitates collaboration among multiple stakeholders within the OR department. In addition, a queueing-based approximation model is employed to derive theoretical lower bounds on regular OR capacity subject to surgical pool backlog constraints, informing coordination between ORs and inpatient beds. We further analyze the mechanism's robustness against strategic behavior, confirming its resilience to potential manipulation by stakeholders. Extensive numerical experiments using real hospital data yield practical insights for applying decision-support methods in the redesigned scheduling process of the SPACE framework. The framework highlights how process redesign, artificial intelligence, and decision-support approaches can be jointly leveraged to systematically address the challenges of OR scheduling.
Many systems involve dynamic interactions between human and non-human elements to accomplish shared objectives, a concept referred to as Distributed Situation Awareness (DSA). These sociotechnical systems emphasize the collaborative interplay between social and technical elements. Network analysis can be used to represent communication and information exchanges among system components, making it easier to visualize interconnections within the system. Using the Event Analysis of Systemic Teamwork (EAST) Framework, this study constructs three networks-social, knowledge, and task-to represent relationships within a sample dermatology department, based on observational data. Based on these networks, the EAST Framework is extended into an analytical model to test relevant design implications. Based on this extension, five models were constructed using Agent-Based Simulation (ABS) to simulate patient flow within the department. The five models - baseline, social-based, task-based, knowledge-based, and combined- were evaluated using both operational and sociotechnical measures. Overall, the combined model performed best across the selected performance metrics from both cognitive and operational standpoints. Hence, when designing workflow and spatial arrangement of sociotechnical systems, using this extension of the EAST Framework is pivotal for operational success and team dynamics.
Purpose: Accurate glucose forecasting is vital for personalized diabetes management and proactive clinical decisions. This study aims to enhance the prediction horizon(PH) by developing BGDeepAR-TA, a probabilistic autoregressive network integrating multi-head temporal attention to prioritize clinically relevant glucose patterns dynamically. Methods: The standard DeepAR model was adapted to glucose time series via custom weight initialization (BGDeepAR) and further augmented with a temporal attention network to create BGDeepAR-TA. Experiments used the OhioT1DM and CGM-T2DM datasets, evaluating split ratios, context lengths, and initialization strategies. Results: For the OhioT1DM dataset, BGDeepAR-TA achieved 100% clinical accuracy for all prediction horizons, and it reduced average RMSE of 60-min PH by 36% (13.08 +/- 1.38 mg/dL vs. state-of-the-art). For CGM-T2DM, BGDeepAR-TA achieved 97-100% clinical accuracy for PHs upto 90 min, and the average RMSE improved by 20% (15.84 +/- 0.98 mg/dL) and 11% (31.27 +/- 1.38 mg/dL) for 30 and 60 min, respectively. For longer PHs of 75 and 90 min, superior performance was exhibited by the proposed models. Conclusion: The proposed models advance clinical deployability for closed-loop insulin systems and early warning tools. In conclusion, the proposed work shows significant promise for improving glucose time series forecasting by integrating the temporal modeling capabilities of DeepAR with attention mechanisms to enhance precision medicine in diabetes management.
Precision medicine traditionally identifies patient subgroups using genomic and molecular profiling, but such approaches are costly and complex, limiting widespread implementation. Electronic health records (EHRs) provide a scalable and accessible alternative for identifying clinically meaningful subgroups. We introduce a Simple Structure (SS) method, which partitions patients based on the ease of predicting outcomes using simple models, rather than relying on unsupervised clustering or feature-based stratification. Two datasets are analyzed: one predicting COVID-19 morbidity and another predicting heart failure diagnosis. For the COVID-19 data, an ensemble of logistics regression models trained on simple structures outperforms the benchmarks, and the features used by the ensemble model are more diverse than those used by the global model, thus revealing subgroups of patients where different comorbidities are important. Similarly, in the heart failure dataset the SS-identified subgroups discover patient subgroups with distinct risk factor profiles. This suggests that rather than a single global model with a fixed set of features, different patient subgroups exhibit unique predictors of heart failure. The ensemble of simple models, leveraging these subgroup-specific risk factors, outperforms simple global models while maintaining interpretability, offering a step toward precision healthcare.
Opioid overdose remains a major public health concern in the United States, claiming thousands of lives each year. In response, public health agencies have intensified efforts to expand access to life-saving interventions, with particular emphasis on the opioid antagonist naloxone. This article presents a data-driven inventory modeling framework that systematically integrates cost efficiency, responsiveness, and access into the design of naloxone distribution strategies. Specifically, we propose a parameterized lateral transshipment policy and demonstrate its potential to enhance inventory management system performance. Building on this, we incorporate deep reinforcement learning to develop dynamic and adaptive policies capable of further improving these outcomes across diverse operational conditions. The framework is validated through a case study grounded in real-world data from the state of Illinois.
As hospitals face growing pressure to deliver high-quality, efficient care with limited resources, effective patient admission scheduling has become increasingly vital-it plays a crucial role in optimizing bed utilization, managing case mix, and balancing resource supply and demand in downstream inpatient processes. To address the challenges caused by siloed and myopic decision-making in current practice, which often leads to unbalanced bed usage and misaligned resource allocation, this study proposes a case mix inspired patient admission scheduling model that not only integrates admission scheduling with case mix optimization but also explicitly incorporates preoperative examination resource alignment. An integrated approach combining prediction, optimization, and simulation is then established, in which machine learning forecasts the preoperative examination resource needs of patients, an integer programming model optimizes admission scheduling by jointly considering case mix and resource constraints, and a discrete-event simulation evaluates the model performance using real hospital data. Simulation experiments comparing five scheduling strategies demonstrate that the proposed model significantly improves preoperative resource matching, reduces waiting time, and enhances overall scheduling efficiency. Sensitivity analyses further indicate that the current prediction performance is sufficient to support these improvements, with only limited marginal gains from further prediction enhancements. Consequently, the model demonstrates robustness to prediction inaccuracies and practical applicability, underscoring its potential for real-world deployment in optimizing admission scheduling through case mix integration.
Hospital length of stay (LOS) drives healthcare costs and resource utilization, yet most research focuses on patient-level predictors while overlooking institutional and regional influences. This study quantifies the relative contributions of patient, hospital, and regional factors to LOS variation using hierarchical modeling. We analyzed 20,000 inpatient admissions from 2,812 hospitals across four U.S. census regions in the 2019 National Inpatient Sample, estimating Generalized Additive Mixed Models with random effects for variance decomposition. Results reveal that system-level factors account for 54% of LOS variation, with regional context (34%) exceeding hospital-specific factors (20%), whereas patient characteristics explain 46%. Among institutional characteristics, only hospital ownership remains significant after hierarchical adjustment, with private hospitals showing 1.4 to 1.8% shorter stays than government facilities. Notably, bed size and teaching status, which appear significant in conventional models, become nonsignificant when within-hospital clustering is addressed. These findings suggest that regional coordination strategies may yield greater efficiency gains than institution-focused interventions, offering evidence-based guidance for healthcare policy and resource allocation.
Cancer is defined as mutations or abnormal changes in genes that regulate cell growth and maintain their health. When these altered cells continue to divide uncontrollably, they can generate similar cells, ultimately forming a tumor. The interactions among genes can influence the rate of cancer development. Gene expression technology allows simultaneous measurement of thousands of genes in a single experiment. However, analyzing the data is challenging because of its complexity, heterogeneity, and intricate interconnections among genes. Network science offers a promising framework for studying genetic data, by representing gene-gene relationships explicitly, enabling structured information sharing, noise averaging, and improved generalization. Using breast cancer gene expression data, this study systematically constructs gene networks and investigates how different network topologies and properties influence the predictive performance of a Graph Convolutional Neural Network (GCNN) model. Network science analysis inspired the development of two network-aware feature selection approaches, leveraging more sophisticated mathematical techniques based on network science to reduce model complexity. The findings indicate that a correlation threshold of 0.5 acts as an optimal point, consistently yielding the best results for GCNN models. Model performance is influenced by the network topology. While stronger connectivity can improve performance, there may be a threshold beyond which excessive connections actually impede the flow of information.
Alarm fatigue is a prevalent issue in the hospital setting that is caused by numerous non-actionable alarms that contribute to increased stress, desensitization and heavy workload in nurses. This study's purpose was to use the Systems Engineering for Patient Safety (SEIPS) model to develop a simulation and test an alarm management intervention. 36 nurses participated in a simulation and were placed in either a control or intervention condition. The control condition matched the current PCU alarm system and the intervention condition contained improvement options. Participates updated fabricated patient records while responding to alarms that occurred at intervals within the 30-minute trial. Participants completed surveys before and after to assess their physiological and physical conditions before and after. The results show that the intervention condition yielded 62% higher alarm update number and accuracy and higher accuracy with alarms with an average 37% increase. Physiological factors improved as well with emotional stress, frustration, and confusion decreased by 23%, 42%, and 31% respectively. The findings suggest a correlation between the intervention design and improved work performance and decreased fatigue symptoms. The SEIPS method for designing a physical simulation is also applicable to similar alarm management problems. Future studies will need to examine the iterative nature of the studies and create future sustainability plans.
Many healthcare systems are facing a severe shortage of rehabilitation resources. In this paper, an analytical framework for image-based automatic evaluation and intervention in motor function rehabilitation is introduced. The framework uses a regular camera to capture patient movements, adopts machine learning methods to evaluate scale results representing the status of a patient, considers the impacts of different clothing, and finally, generates personalized intervention plan. Such a framework overcomes the time-consuming limitation of traditional rehabilitation evaluation that requires extensive sessions and physician's presence. Moreover, the intervention plans generated by the framework demonstrate consistency with those developed by physicians. By integrating evaluation and intervention, the proposed method can serve as a quantitative tool to reduce the burden of physicians, improve system efficiency and patient outcomes in rehabilitation processes, and help satisfy the growing need for accessible and efficient rehabilitation services.
Architecture transferability is crucial for advancing computer-aided diagnosis (CAD), especially in complex tasks such as skin lesion diagnosis, where data variability and scarcity can challenge the effectiveness of traditional diagnostic models. This study introduces a novel approach to improving knowledge transfer in skin lesion diagnosis by leveraging progressive differentiable architecture search (P-DARTS). Our approach discovers neural network architectures that adapt well across diverse datasets, enhancing the model's ability to generalize and transfer learned neural architecture to new contexts. The optimized architecture was evaluated using the publicly available ISIC-2019, PAD-UFES-20, and DermaMNIST datasets, demonstrating competitive performance against established models such as InceptionNet, ResNet, and DenseNet. P-DARTS produced high-performing, transferable models with an average diagnostic accuracy of 63.70% in multiclass classification and 97.71% in binary classification for PAD-UFES-20, underscoring its potential for creating robust artificial intelligence (AI) tools for accurate and timely skin lesion diagnosis. In addition, we observed that evaluation depth had more significance than the search-evaluation depth gap and that architectures with a greater number of pooling operations resulted in better performance. To ensure robustness and reliability of the evaluation, we extended the analysis across multiple image resolutions and incorporated clinically relevant metrics, including weighted F1-score, true positive rate, true negative rate, and area under the ROC curve. The results confirmed the consistency of architecture transferability across different evaluation settings. This approach addresses a key limitation in knowledge transfer for skin lesion diagnosis and fills the need for adaptable models that are capable of learning from limited, domain-specific data.
