Orfanoudaki, Agni

Tackling hospital emergency department (ED) overcrowding is a paramount challenge for healthcare systems. To combat this issue, an innovative approach is to identify patients who can be served vertically (i.e., in a seated position) and route them to a dedicated area termed the vertical processing pathway (VPP). Successfully implementing this design requires understanding which patients should be routed to the VPP and when. Currently, the decision to leverage the VPP is made in an ad hoc fashion. To assist our partner hospital and other EDs in capturing the value of the VPP, we develop a machine learning model that provides personalized risk scores predicting whether each arriving patient needs an ED bed. We use these scores as input to a stochastic patient flow model and analytically characterize the optimal VPP policy that minimizes length of stay. We then combine our results to derive an interpretable VPP patient streaming protocol suitable for implementation in practice and conduct a before-and-after experiment in which we leverage empirical analyses to evaluate the impact of integrating it into practice. The implemented protocol led to an 11-minute (4.2%) reduction in ED length of stay in our partner hospital without any adverse effect on quality-of-care outcomes. This effect was statistically significant and remained robust after controlling for various confounding factors and potential sources of endogeneity. Our VPP protocol can also be utilized in other EDs, offering operational improvements without requiring additional resources. This paper was accepted by Stefan Scholtes, healthcare management. Funding: S. Saghafian received financial support from the NSF [Grant CMMI-15 62645], “Data-Driven Management of Post-Transplant Medications,” which partially enabled this work, as well as funding and support from the Harvard Data Science Initiative and the Harvard Mossavar-Rahmani Center for Business and Government. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2024.07517 .

Background/Objectives: Recent advances in artificial intelligence (AI) and machine learning (ML) enable targeted optimization of emergency department (ED) operations. We examine how reworking an ED’s vertical processing pathway (VPP) using AI- and ML-driven recommendations affected patient throughput. Methods: We trained a non-linear ML model using triage data from 49,350 ED encounters to generate a personalized risk score that predicted whether an incoming patient is suitable for vertical processing. This model was integrated into a stochastic patient flow framework using queueing theory to derive an optimized VPP design. The resulting protocol prioritized a vertical assessment for patients with Emergency Severity Index (ESI) scores of 4 and 5, as well as 3 when the chief complaints involved skin, urinary, or eye issues. In periods of ED saturation, our data-driven protocol suggested that any waiting room patient should become VPP eligible. We implemented this protocol during a 13-week prospective trial and evaluated its effect on ED performance using before-and-after data. Results: Implementation of the optimized VPP protocol reduced the average ED length of stay (LOS) by 10.75 min (4.15%). Adjusted analyses controlling for potential confounders during the study period estimated a LOS reduction between 7.5 and 11.9 min (2.89% and 4.60%, respectively). No adverse effects were observed in the quality metrics, including 72 h ED revisit or hospitalization rates. Conclusions: A personalized, data-driven VPP protocol, enabled by ML predictions, significantly improved the ED throughput while preserving care quality. Unlike standard fast-track systems, this approach adapts to ED saturation and patient acuity. The methodology is customizable to patient populations and ED operational characteristics, supporting personalized patient flow optimization across diverse emergency care settings.

AbstractIntroductionInpatient hyperglycemia is an established independent risk factor among several patient cohorts for hospital readmission. This has not been studied after kidney transplantation. Nearly one‐third of patients who have undergone a kidney transplant reportedly experience 30‐day readmission.MethodsData on first‐time solitary kidney transplantations were retrieved between September 2015 and December 2018. Information was linked to the electronic health records to determine diagnosis of diabetes mellitus and extract glucometric and insulin therapy data. Univariate logistic regression analysis and the XGBoost algorithm were used to predict 30‐day readmission. We report the average performance of the models on the testing set on bootstrapped partitions of the data to ensure statistical significance.ResultsThe cohort included 1036 patients who received kidney transplantation; 224 (22%) experienced 30‐day readmission. The machine learning algorithm was able to predict 30‐day readmission with an average area under the receiver operator curve (AUC) of 78% with (76.1%, 79.9%) 95% confidence interval (CI). We observed statistically significant differences in the presence of pretransplant diabetes, inpatient‐hyperglycemia, inpatient‐hypoglycemia, minimum and maximum glucose values among those with higher 30‐day readmission rates. The XGBoost model identified the index admission length of stay, presence of hyper‐ and hypoglycemia, the recipient and donor body mass index (BMI) values, presence of delayed graft function, and African American race as the most predictive risk factors of 30‐day readmission. Additionally, significant variations in the therapeutic management of blood glucose by providers were observed.ConclusionsSuboptimal glucose metrics during hospitalization after kidney transplantation are associated with an increased risk for 30‐day hospital readmission. Optimizing hospital blood glucose management, a modifiable factor, after kidney transplantation may reduce the risk of 30‐day readmission.