Poorly separated data and imbalanced data present major challenges to classifiers which often result in lower accuracy and reliability in making predictions. In this paper, we introduce the multi-stage classification construct in which ‘difficult-to-classify’ observations are placed into a reserved judgment region for delayed future classification. Such a design is well-suited for poorly separated data that are difficult to classify without committing a high percentage of misclassification errors. The misclassification constraints within the classifier can be finetuned to allow management of imbalanced data, guiding minority entities into the reserved judgement region to avoid being misclassified into the majority groups. We wrap the classifier with a fast feature selection heuristic based on particle swarm optimization. An exact combinatorial branch-and-bound algorithm is also implemented to measure the quality of the heuristic solutions. We apply this multi-stage multi-group machine learning framework to two real-life medical problems: (a) multi-site treatment outcome prediction for best practice discovery in cardiovascular disease, and (b) uncovering patient characteristics that predict optimal response to intra-articular injections of hyaluronic acid for the treatment of knee osteoarthritis. Both problems involve poorly separated data and imbalanced groups in which traditional classifiers yield low prediction accuracy (47% - 66%). The multi-stage BB-PSO/DAMIP manages the poorly separated and imbalanced data well and returns interpretable results with 82% - 97% blind prediction accuracy.

More than 14 million new cases of cancer are diagnosed globally each year, with approximately 50% of all cancer patients receiving radiation therapy during the course of their disease. Advanced technology using a multi-leaf collimator can deliver non-uniform radiation doses that best conform to the tumor shape while sparing the surrounding normal tissues. Combinatorically, selecting beam angles is computationally taxing and remains a major challenge. Consequently, clinical systems opt to use preselected angles without integrating a combinatorial computational engine within the fluence map optimization process.In this paper, we offer a mixed integer quadratic programming optimization approach to simultaneously determine optimal beam angles and fluence maps. For computational efficiency, a dynamic restricted branch-and-price method is derived to solve these large-scale patient instances. Clinical analyses of prostate and head-and-neck cancer patient cases show that our approach yields treatment plans that are far superior to current best practice in the clinical environment. Specifically, the resulting plans achieve better conformity and homogeneity to the tumor while reducing radiation exposure to the critical structures. Computationally, besides drastically reducing the solution time, the dynamic restricted branch-and-price method provides an additional bonus: every node consists of only a relatively small number of beam variables in the branch-and-bound framework. This implies that in the middle of the solution process, a feasible and clinically acceptable treatment plan is available, and can be obtained by taking a few branching steps on the beam variables that are in the active set.

This study establishes interoperability among electronic medical records from 800 clinical sites and uses machine learning for best practice discovery. A novel extraction-mapping algorithm is designed that accurately extracts, summarizes, and maps free text and content to concise structured medical concepts. Multiple concepts are mapped, including patient diagnoses, laboratory results, medications, and procedures, which allow shared characterization and hierarchical comparison. In addition, clinical and decision processes are established. These integrated data can be accessed through a secure web-based portal. A classification machine learning model (DAMIP) is then leveraged to establish predictive rules by uncovering relatively small subsets of discriminatory features that can predict the quality of treatment outcomes. We demonstrate system usability by analyzing treatment outcome for cardiovascular diseases, diabetes, hypertension, and chronic kidney disease. DAMIP yields good blind prediction accuracies of 89% – 97%. Moreover, for each disease the best practice was only used at fewer than 5% of the clinical sites, offering an excellent opportunity for knowledge sharing and rapid learning. Our findings also led to implementation of a new treatment policy for chronic kidney disease management. This resulting policy offers a better outcome for patients, saves lives, improves the quality of life for patients, and reduces 35% of treatment costs. Thus, this work has critical health practice implications. Through data-driven interoperability and predictive analytics, we have established and identified practice characteristics that result in good outcomes. The system is scalable and generalizable to different hospital settings and health conditions.

SARs, bird flu, H1N1, Ebola crisis in W. Africa, Zika and current SARS-CoV-2 underscore the critical importance of emergency response and medical preparedness. Such needs are wide-spread as globalization and air transportation facilitate rapid disease spread across the world. Computational modeling of infectious disease outbreaks and epidemics offer insights in propagation patterns and facilitate policy makers to synthesize potential interventions. Current models include inclined decay with an exponential adjustment, SEIR (susceptible, exposed, infectious, recovered) compartmental model, discrete time stochastic processes, and transmission tree. While many of these models incorporate contact tracing to predict spread pattern, key elements on optimal usage of scarce resources and effective and efficient process performance (e.g., diagnostics and screening, non-pharmaceutical interventions, trained personnel/robots for treatment, decontamination) have not been included. This is particularly critical in the fight of COVID-19 containment due to lack of testing kits and the prevalence of asymptomatic transmission, and the long period of hospitalization required by severely sick patients.This work focuses on designing a system computational decision modeling framework that simultaneously i) captures disease spread characteristics, ii) incorporates day-to-day hospital and home care processes and resource usage, iii) explores non-pharmaceutical intervention, social and human behavior and iv) allows for system optimization to minimize infection and mortality under time and labor constraints.

We develop a theoretical framework for a two-sided market structure to model the competition between a Preferred Provider organization (PPO) and a Health Maintenance organization (HMO). Both health plans compete to attract policyholders and providers. Our game-theoretical framework examines the consequences of risk segmentation on providers and the network effect on policyholders based on market information and network size. The outcome of the competition depends mostly on two effects: a market share effect and an adverse selection effect, captured by policyholders' surplus expectations on both policies and copayments. If the adverse selection effect is strong enough, the HMO plan gets higher profits. On the other hand, if the market share effect dominates, the PPO profit is higher despite the unfavorable risk segmentation and higher premium. The twosided nature of the health insurance market can explain the preference of higher flexibility that has been observed during the last 15 years in the US health insurance market. However, PPO plans were drastically reduced in HealthCare.gov's 2016 lineup since many insurers claimed to be losing money on these plans and opted not to offer them in the health insurance exchange market. Our analysis shows that to be profitable, the PPO has to maintain an appropriate ratio between in-network copayment and out-of-network copayment. One conjecture is that HMO can offer more attractive remuneration to the physicians' side and premium to the policyholders' side. This may explain the gradual gain in the HMO market.

Recent studies have shown that ensemble feature selection approaches can improve the robustness and stability of final classification models. Existing methods for aggregating feature lists from different methods require use of arbitrary thresholds for selecting the top ranked features and are often based on metrics independent of the classification accuracy while selecting the optimal set. In this paper, we develop the OptSelect tool for ensemble feature selection and stability assessment of individual features for improved biomarker discovery. The software tool is packaged in R for broad dis-semination. OptSelect is a multi agent-based stochastic optimization tool designed for ensemble feature selection. Stage one involves function perturbation, where ranked list of features is generated using multiple feature selection methods. Stage two in-volves data perturbation, where feature selection is performed within randomly selected learning sets of the training da-ta. The agents are assigned to different behavior states and move according to a binary PSO algorithm. A multi-objective fitness function is used to evaluate the classification accuracy of the agents. We evaluate OptSelect system performance using the random probe method testing on five publicly available microarray datasets. The performance is compared with single feature selection techniques and existing aggregation methods. The results show that OptSelect improves classification accuracy when compared to both individual and existing rank aggregation methods. The PSO algorithm is able to uncover important discriminatory features for predicting COVID-19 disease severity, demonstrating its important role within the optSelect tool. The algorithm is incorporated into an R package and disseminated via GitHub: https://github.com/kuppa12/optSe1ect.

The rapidly increasing size of biomedical databases such as Medline requires the use of intelligent data mining methods for information extraction and summarization. Existing biomedical text-mining tools have limited capabilities for incorporating citation information during document ranking and for inferring topological and network relationships between biomedical terms. Often too much is returned during summarization leading to information overload. Furthermore, literature-based discoveries could be hard to interpret if the network is too complex. SEACOIN2.0 can incorporate citation information during document ranking and uses a unique association rule mining algorithm to generate multi-level k-ary trees. The multi-level trees facilitate efficient information retrieval, visual data exploration, summarization, and hypothesis generation. The system presents graphical summarization via multiple dynamic visualization panels and an interactive word cloud. LexRank algorithm is used to identify salient sentences in top abstracts related to the query. An average F-measure of 94% was achieved for document retrieval, and an average precision of 88% was obtained for identification of top co-occurrence terms. SEACOIN2.0 was also used to replicate previously published findings using the literature-based discovery and EMR-based PheWAS approaches. We present herein SEACOIN2.0 (https://newton.isye.gatech.edu/SEACOIN2/), an interactive visual mining tool for improved information retrieval, automated multi-level summarization of Medline abstracts, and literature-based discovery. SEACOIN2.0 addresses the problem of “information overload” and allows clinicians and biomedical researchers to meet their information needs.

Strong social networks can encourage healthy behaviors. In this paper, we introduce a sociology-based computational framework for influence networks. The model construct is generic and is applicable to diverse social network analysis. We demonstrate its usage in calibrating the positive influence of church clergy in spreading HIV/AIDs information in a large metropolitan city. Five experiments are designed to contrast influence with respect to the interaction style between clergy and churchgoers. Competitive and non-competitive knowledge dissemination are also analyzed. The results show that when only one set of information exists, the spreading scope is directly proportional to the product of population size and the disease infection rate. When competing information is present, the importance of clergy in spreading the information decreases when the original propagation sources are ample. However, if sufficient interaction and trust are present among the clergy and the participants, the clergy's positive influence remains significant despite pre-existing knowledge. The generalized framework requires minimal regional data to establish the influence network. It provides useful policy insights for decision makers to determine effective avenues for information dissemination through community influencers.

We consider the problem of partitioning clinical services in hospitals into groups with the goal of efficiently allocating existing inpatient beds. At the strategic level, there are two major possibilities: pooling versus focusing. Pooling the bed capacity allows one to achieve an overall high occupancy level for a fixed number of beds. On the other hand, focusing by dividing the capacity into groups with restricted access may offer increased efficiency and better resource utilization. We derive a two-stage schema to address the 3-fold problem: 1) how many groups of services to form; 2) how many beds to allocate to each group; and 3) how to partition services among the groups. Stage 1 uses cluster analysis utilizing the similarity principle for possible advantages of economies of scale, coupled with queueing-based optimization models to obtain a set of candidate groups. Stage 2 incorporates utility/benefit functions to optimize the partitions and allocation of beds. Three full-scale examples demonstrate the flexibility and diverse application of our framework with managerial insights for different utility optimization goals and queueing systems.

Remote patient monitoring systems (RMS) have gained increasing popularity in recent years. RMS have great potential to improve medical services by providing more affordable, timely, and accessible care. This paper describes an effective low-cost RMS that is readily deployable. The system targets chronic disease patients and attempts to reduce patient visits to the hospital and healthcare costs. The system is comprised of three modules: (1) an application for data acquisition, processing, and transmission, (2) an adaptable set of “personalized” sensors for measuring vitals and reporting emergency situations, and (3) a secure communication module for remote patient-physician interactions. The users interface with the RMS through an application installed on a mobile device. Using a return of investment (ROI) cost-benefit analysis and a cohort of 2.7 million patients, we estimate that through the implementation of such a system, the patients and the healthcare system would see benefits within one year.

We consider the problem of partitioning clinical services in hospitals into groups with the goal of efficiently allocating existing inpatient beds. At the strategic level, there are two major possibilities: pooling versus focusing. Pooling the bed capacity allows one to achieve an overall high occupancy level for a fixed number of beds. On the other hand, focusing by dividing the capacity into groups with restricted access may offer increased efficiency and better resource utilization. We derive a two-stage schema to address the 3-fold problem: 1) how many groups of services to form; 2) how many beds to allocate to each group; and 3) how to partition services among the groups. Stage 1 uses cluster analysis utilizing the similarity principle for possible advantages of economies of scale, coupled with queueing-based optimization models to obtain a set of candidate groups. Stage 2 incorporates utility/benefit functions to optimize the partitions and allocation of beds. Three full-scale examples demonstrate the flexibility and diverse application of our framework with managerial insights for different utility optimization goals and queueing systems.

Risk and decision models and predictive analytics have long been cornerstones for advancement of business analytics in industrial, government, and military applications. In particular, multi-source data system modeling and big data analytics and technologies play an increasingly important role in modern business enterprise. Many problems arising in these domains can be formulated into mathematical models and can be analyzed using sophisticated optimization, decision analysis, and computational techniques. In this talk, we will share some of our successes in healthcare, defense, and service sector applications through innovation in predictive and big data analytics through the modeling and computational advances in integer programming. Specifically, the first model is a discrete support vector machine predictive model that incorporates comprehensive factors related to demographics and socio-economic status, clinical and hospital resources, operations and utilization, and patient complaints and risk factors for global prediction of readmission and treatment outcome of patients. The second model describes an outcome-driven personalized treatment planning model for cancer patients.

This study focuses on reducing SSI for coronary artery bypass graft(CABG) patients at a large safety-net hospital. The SSI rate in 2010–11 hovered around 20%. The overall goal is to reduce CABG-SSI incidence by 25% (to < 15%) and measure its sustainability. A system-approach is employed which takes into account the interdependency of preoperative, intraoperative and postoperative processes. A decision tree model and a simulation-optimization model are developed to identify critical infection factors. Implemented changes involve pre-op sterilization, aggressive nasal cleaning, proper hair-clipping, and optimized antibiotics prophylaxis timing and dosage. E-alerts are also implemented for documentation to facilitate compliance and training. The site realized a drop of 65% in SSI (from 23% to 8%) in the first six months. It achieved zero percentage thereafter and sustained that rate for 18 months. The system-approach is generalizable and is currently being explored for rectal-colon cancer and hysterectomy, the second most common elective surgery among American women.

The study aims to reduce the incidence of central line associated bloodstream infections (CLABSI). We design a computer model that comprises the entire process of central line insertion and maintenance. The model attempts to capture all major events in patient care from entrance to the hospital through the time at which the central line is ultimately removed. CLABSI data is analyzed to identify areas of potential increased risk of patient death from CLABSI. Specifically, we attempt to predict death among patients affected with CLABSI while minimizing Type II (false negative) error. By crafting the model to prioritize against this type of error, and thus identifying patients most likely to die, providers will have the best chance of intervening to reduce CLABSI-related deaths. The study led to implementation of reminders and protocols that result in a reduction of 18% of CLABSI over a period of 12 months. The predictive analysis identifies high-risk individuals to allow for proper intervention to prevent CLABSI-related deaths.

Medication errors significantly impact patient health and quality of care. In this study, systems modeling and simulation are used to analyze the medication workflow in a pediatric setting. The resulting simulation system allows for derivation and validation of effective intervention strategies for error mitigation. The analysis focused on ‘High Alert’ medications, which are more likely to produce serious patient harm when errors occur. Our study shows that strategic process interventions (e.g., independent intervention check points) can significantly reduce error occurrence (by as much as 50.25%) and the cost savings could be significant. We analyze the cost-savings versus intervention effectiveness versus the cost of the extra resource needed to perform the intervention, and identify the break-even point. The estimated errors from our model remain consistently close to the reported error statistics (within 5%). This work offers a system-decision support framework for analysis of medication workflow and helps in understanding error propagation mechanisms and process interdependencies. The framework allows users to derive intervention strategies and evaluate their overall mitigation effectiveness.

Solution methods are presented for a mixed-integer program (MIP) associated with a method for constrained discrimination. In constrained discrimination, one wishes to maximize the probability of correct classification subject to intergroup misclassification limits. The misclassification limits are satisfied by allowing the placement of observations in a reserved judgment group. The approach investigated here involves modifying a standard classification rule by solving an optimization problem. A polynomial-time algorithm for solving the problem is given for two-group discrimination. The decision problem upon which the optimization problem is based is shown to be NP complete for a general number of groups. For three or more groups, an MIP is used to solve the problem. Solution methods incorporating cutting planes from conflict graphs are presented for solving instances in a branch-and-bound framework. These methods are used to enhance industry-standard software, and are shown to provide as much as a 20-fold reduction in computational time over the software alone. Computational experiments illustrate the tradeoff between misclassification rates and reserved judgment rates. Some base classifiers are not well suited to be modified to a constrained discrimination rule. The method for constrained discrimination studied here performs particularly well in the presence of class imbalance. For certain other data sets, however, the method is outperformed by a simple centroid method.

A parallel implementation of a disjunctive cutting-plane algorithm in a distributed memory environment is described. Guided by a selection of difficult instances from MIPLIB and real instances obtained from brain-tumor research, various strategies of cut synchronization are considered, and their influence on speedup, communication overhead, load balance, and effectiveness in closing the integrality gap are studied. The parallel cutting-plane algorithm is coupled with an LP-based heuristic to assist in returning a good quality integer feasible solution upon termination of the parallel process. The parallel implementation is sufficiently coarse-grained to yield an average of less than 6% of the total time performing tasks associated with communication overhead, and it provides reasonable speedup when executing in parallel. Noticeable differences in load-balance scores are observed, depending on the number of processors used, the synchronization scheme used, and the structure of the MIP problem instance. Nevertheless, the synergism of the combined collection of cuts generated locally on each processor is effective in closing the integrality gap in all cases, and there is minimal variability in the amount of the gap closed as the number of processors varies. In particular, the degree of decentralization, as governed by the synchronization schemes, has little effect on the overall quality of the cuts generated.

A linear-programming model is proposed for deriving discriminant rules that allow allocation of entities to a reserved-judgment region. The size of the reserved-judgment region, which can be controlled by varying parameters within the model, dictates the level of aggressiveness (cautiousness) of allocating (misallocating) entities to groups. Results of simulation experiments for various configurations of normal and contaminated normal three-group populations are reported for a variety of parameter selections. Results of cross-validation experiments using real data sets are also reported. Both the simulation and cross-validation experiments include comparison with other discriminant analysis techniques. The results demonstrate that the proposed model is useful for deriving discriminant rules that reduce the chances of misclassification, while maintaining a reasonable level of correct classification.

We describe a parallel, linear programming and implication-based heuristic for solving set partitioning problems on distributed memory computer architectures. Our implementation is carefully designed to exploit parallelism to greatest advantage in advanced techniques like preprocessing and probing, primal heuristics, and cut generation. A primaldual subproblem simplex method is used for solving the linear programming relaxation, which breaks the linear programming solution process into natural phases from which we can exploit information to find good solutions on the various processors. Implications from the probing operation are shared among the processors. Combining these techniques allows us to obtain solutions to large and difficult problems in a reasonable amount of computing time.