Effective discharge decision-making in intensive care units (ICUs) is crucial for optimizing patient outcomes and ensuring efficient utilization of limited resources, a challenge exacerbated during the COVID-19 pandemic. Discharge decision-making involves a tradeoff: early discharge increases the risk of ICU readmission and mortality, whereas prolonged ICU stays raise costs and strain capacity. We develop a data-driven framework that integrates offline reinforcement learning (offline RL) with constrained reinforcement learning (constrained RL) to derive personalized ICU discharge policies. We consider the decision-making process with multiple ICU readmissions, and readmission count is incorporated into the state space. The framework also explicitly incorporates clinical and operational constraints, including ICU length of stay (LOS) and readmission risk. Using real-world medical data from the MIMIC-IV database, we train and evaluate the framework with sensitivity and robustness analysis . The learned policiesreduce esimated mortality risk while satisfying predefined thresholds on LOS and readmission risk. Numerical results also show that the learned policies outperform observed clinical practices in mortality estimates and demonstrate conservative decision behaviors for high-risk patients. Overall, this work highlights the feasibility and value of combining offline and constrained RL for critical care decisoin support.

During pandemics, understanding how population mobility and connectivity amongst regions impact the spread of infections is crucial in evaluating the effectiveness of non-pharmaceutical interventions. This paper presents a novel endemic–epidemic (EE) infectious disease model that utilizes vehicle trajectory mobility data to identify the geographic coupling weights and better quantify the impacts of mobility on disease progression. Specifically, a new hierarchical clustering method is proposed for segmenting origin-destination (OD) flow rate matrix time-series data, which in turn is used to estimate geographic coupling weights of the EE model. Simulation experiments showed that EE models using the proposed coupling weight estimation approach provide significantly improved predictions of infection counts. A real-data case study covering the COVID-19 pandemic shows that the proposed movement rate approach provides meaningful insights into the impacts of the neighbor effects on disease spread.

Analysis of transient performance of patient flow in hospitals is critical for healthcare delivery, which could offer valuable insights into building and environmental design, medical resource allocation, and traffic planning. Despite extensive research on hospital operation management, effective analytical models to study dynamic behaviors of patient flow remain scarce. Due to the time-varying and stochastic features in patient arrivals and care services, transient modeling, rather than steady-state analysis, could provide insights into short-term system dynamics, which is strongly needed, but the study is limited. To bridge this gap, this study introduces a dynamic queueing network model to analyze the transient performance of patient flow, under the scenario of a large public hospital. Specifically, the transient behavior of patient service processes, characterized by multiple units with complex structures, is analyzed using a history-buffer-based iterative method derived from dynamic equations. To address cumulative errors in long-term analysis, a real-time correction method is proposed. Extensive numerical experiments demonstrate that the proposed approach can provide transient estimation of patient flow with acceptable accuracy. In addition, the influence of parameter settings and system properties are investigated. Such a model provides an effective analytical tool for real-time management to improve care delivery.

With the high prevalence of disability and multiple chronic conditions, older adults often experience multidimensional and highly heterogeneous functional limitations and declines, generating diverse needs for various healthcare providers over time. To support effective healthcare resource preparation and planning for an ever-growing aging population with diverse needs, it is crucial to accurately model the multidimensional performance degradation while explicitly capturing degradation heterogeneity at the subpopulation level. Many existing degradation performance modeling approaches for older adults have restrictive assumptions, such as assuming a homogeneous population, modeling univariate degradation performance, and/or prespecifying the number of subpopulations in advance. To relax these modeling assumptions and provide deeper healthcare insights into population aging, we propose a subpopulation-level heterogeneity modeling framework to jointly characterize multiple and correlated functional degradation performance over time for a heterogeneous population of older adults. Specifically, a data-driven semi-parametric model is formulated to relax the conventional assumption of pre-specifying a fixed number of subpopulations. An effective model estimation algorithm further jointly identifies the number of subpopulations and estimates parameters for each subpopulation across multiple performance degradation dimensions. Based on the identified subpopulations with distinct aging patterns, additional quantitative analyses were conducted to generate healthcare-relevant insights, including identifying and quantifying key risk and protective factors associated with subpopulation membership, assessing the impact of degradation heterogeneity on patterns of healthcare service utilization, and evaluating cost implications to inform targeted health promotion strategies aimed at improving outcomes and potentially reducing expenditures for high-risk subpopulations across various care settings, such as hospitals and nursing homes. To illustrate the proposed work and highlight its benefits from both modeling and healthcare insight perspectives, a real-world case study characterizing both physical and cognitive degradation performance of community-dwelling older adults was provided.

Diabetic Retinopathy (DR), a complication of diabetes affecting the small blood vessels, stands as the primary reason for vision impairment in adults of working age. This condition often advances to critical stages due to inadequate screening participation and the prohibitive costs of necessary diagnostic technology. The widespread availability of Electronic Health Records (EHRs) offers a significant opportunity to overcome these challenges through the use of machine learning techniques. This paper proposes an MB-TabNet model for interpretable data-driven detection of DR using EHR data. By leveraging the Attentive Interpretable Tabular Learning (TabNet) architecture, along with the multi-branching technique, our MB-TabNet model aims to not only tackle the prevalent issue of imbalanced class distribution in medical datasets, but also enhance the interpretability of deep learning models for DR detection. Experimental results demonstrate that our method not only enhances feature interpretability but also outperforms existing data-driven approaches in current practice. This innovative approach has the potential to enable confident recommendations for personalized ophthalmic exams and screening frequencies, particularly targeting at-risk patients.

Predicting drug demand is of great importance for hospital pharmacy managers. Drug expenditures per hospital in the US averages over 7 million dollars per year, so it is important to manage these costs effectively. Due to erratic demand patterns, however, predicting demand for drugs is not an easy task. One of the patterns that occur when predicting drug demand are demand patterns with changepoints. These non-stationary patterns suddenly change with respect to their mean, which might occur due to drug authority approvals or events such as a pandemic. In this paper, we study demand patterns with changepoints, which we found to affect 25.9% of the top 500 drugs at our partnering hospital. We propose to use a Bayesian Online Changepoint Detection model. The overall performance of this method is at least as good as other methods under regular circumstances, but outperforms other commonly used forecasting tools especially right after the changepoint occurs, showing its ability to quickly adapt to changepoints. Furthermore, changepoints are more likely to occur for pricier drugs. In particular, drugs within the third and fourth quantile price range are almost three times more likely to have a changepoint compared to inexpensive drugs in the first quantile. This implies that especially for pricier drugs, specialized changepoint models may be useful in predicting drug demand.

Robotic-assisted surgery (RAS) is a safe, effective, and rapidly growing minimally invasive surgical technology. This research investigates the average treatment outcomes of RAS and the factors that may influence its effectiveness by a novel optimization-based approach. Traditional causal effect estimation methods fall short of capturing the changing inclination of surgeons and its impact during the early stages of RAS adoption. By applying our optimization-based method, we uncovered several important findings: First, our analysis of Inguinal Hernia Repair, Ventral/Incisional Hernia Repair, Hysterectomy, Nephrectomy, and Colectomy procedures and other minimally invasive surgical procedures in the Southeast U.S. found no significant difference in treatment effect between RAS and non-RAS procedures. In addition, we identified worse treatment outcomes within subgroups undergoing RAS, such as older adults, certain racial groups, and people with lower socioeconomic status. Statistical tests for effect modification revealed that these factors negatively impact RAS outcomes. Faster adoption of RAS technology positively affects treatment outcomes, validating the potential interventions for improved RAS treatment effects through increased insurance coverage and professional training.

The United States (US) opioid crisis contributed to 81,806 fatalities in 2022. It has strained hospitals, treatment facilities, and law enforcement agencies due to the enormous resources and procedures needed to respond to the crisis. As a result, many individuals who use opioids never receive or finish the treatment they need and instead have many interactions with hospitals or the criminal justice system. This paper introduces a discrete event simulation model that evaluates three opioid use disorder treatment policies: arrest diversion, reentry case management, and overdose diversion. Publicly available data from 2011 to 2019 in Dane County, Wisconsin, was used to forecast opioid-related outcomes through 2032. Through analyzing a variety of policy-mix implementations, the study offers a versatile framework for evaluating policies at various implementation levels. The results demonstrate that treatment policies that create new pathways and programming by utilizing treatment services and successfully divert at least 20% of eligible individuals can lead to more opioid-resilient communities. The benefits increase when more policies are enacted and/or offered to more individuals, with the largest impact from overdose diversion, followed by reentry case management, and the smallest impact from arrest diversion. The statistically significant 10-year cumulative total reduction in societal costs from 2023 through 2032 ranges from $39 M (USD) to $584 M (USD), excluding implementation costs of policies. To reverse the opioid crisis within a community, treatment policies may need to be combined with other strategies, such as harm reduction, supply reduction, and use prevention.

Liver cancer is a leading cause of mortality worldwide, and accurate computed tomography (CT)-based tumor segmentation is essential for diagnosis and treatment. Manual delineation is time-intensive, prone to variability, and highlights the need for reliable automation. While deep learning has shown promise for automated liver segmentation, precise liver tumor segmentation remains challenging. We present SmoothSegNet, a novel deep learning framework that addresses this challenge with the three key designs: (1) a novel knowledge-informed label smoothing technique that distills knowledge from clinical data to generate smooth labels, which are used to regularize model training, reducing the overfitting risk and enhancing model performance; (2) a global and local segmentation framework that breaks down the main task into two simpler subtasks, allowing optimized preprocessing and training for each; and (3) pre- and post-processing pipelines customized to the challenges of each subtask aimed to enhance tumor visibility and refines tumor boundaries. We apply the proposed model on a challenging HCC-TACE-Seg dataset and show that SmoothSegNet outperformed various benchmarks in segmentation performance, particularly at smaller tumors (<10 cm). Our ablation studies show that the three design components complementarily contribute to the model improved performance. Code for the proposed method are available at https://github.com/lingchm/medassist-liver-cancer.

The use of electronic health record (EHR) tabular data for clustering has garnered global attention for its potential to uncover meaningful clusters, improve patient care and advance healthcare research. Applications of EHR data extend beyond healthcare to include financial modeling, fraud detection, differential diagnosis and rare disease identification. However, critical challenges persist, including privacy concerns, limited access to real EHR data, underexplored generative methods and inadequate validation frameworks. This study aims to generate and validate synthetic EHR tabular data and demonstrate its utility in clustering, addressing existing gaps in privacy, data quality and validation. Advanced deep learning architectures, including Generative Adversarial Networks (GANs), were developed to generate high-quality synthetic datasets. We implemented robust validation frameworks that incorporate fidelity, utility and privacy-preserving mechanisms. Dirichlet Process Mixture Models (DPMMs) used synthetic data to group similar items and measured the differences between real and synthetic data distributions by estimating divergence. GANs effectively replicated complex datasets while balancing fidelity, utility and privacy. Validation confirmed the robustness of DPMMs in identifying latent structures. Cluster analysis revealed significant multimorbidity patterns, facilitating personalized treatment strategies, resource allocation and preventive care. This study demonstrates the transformative potential of synthetic EHR data in clustering, advancing healthcare delivery, planning and research.