Search

Few studies have quantified the multimorbidity burden in older adults with chronic obstructive pulmonary disease (COPD) using large and generalizable data. Such evidence is essential to inform evidence-based research, clinical care, and resource allocation. This retrospective cohort study used a nationally representative sample of Medicare beneficiaries aged 65 years or older with COPD and 1:1 matched (on age, sex, and race) non-COPD beneficiaries to: (1) quantify the prevalence of multimorbidity at COPD onset and one-year later; (2) quantify the rates [per 100 person-years (PY)] of newly diagnosed multimorbidity during in the year prior to and in the year following COPD onset; and (3) compare multimorbidity prevalence in beneficiaries with and without COPD. Among 739,118 eligible beneficiaries with and without COPD, the average number of multimorbidity was 10.0 (SD = 4.7) and 1.0 (SD = 3.3), respectively. The most prevalent multimorbidity at COPD onset and at one-year after, respectively, were hypertension (70.8% and 80.2%), hyperlipidemia (52.2% and 64.8%), anemia (42.1% and 52.0%), arthritis (39.8% and 47.7%), and congestive heart failure (CHF) (31.3% and 38.8%). Conditions with the highest newly diagnosed rates before and following COPD onset, respectively, included hypertension (39.8 and 32.3 per 100 PY), hyperlipidemia (22.8 and 27.6), anemia (17.8 and 20.3), CHF (16.2 and 13.2), and arthritis (12.9 and 13.2). COPD was significantly associated with increased odds of all measured conditions relative to non-COPD controls. This study updates existing literature with more current, generalizable findings of the substantial multimorbidity burden in medically complex older adults with COPD-necessary to inform patient-centered, multidimensional care.Supplemental data for this article is available online at https://doi.org/10.1080/15412555.2021.1968815 .

During the period from 1999 to 2016, more than 350,000 Americans died from overdoses related to the use of prescription opioids. To the extent that supply is directly related to overprescribing, policy interventions aimed at changing prescriber behavior, such as the recent Centers for Disease Control and Prevention guideline, are clearly warranted. Although these could plausibly reduce the prevalence of opioid overuse and dependency, little is known about their economic and health-related impacts. The aim of this study was to quantify the efficacy of a policy intervention aimed at reducing the length of initial opioid prescriptions. A Markov decision process model was fitted on a retrospective cohort of 827,265 patients, and patient cost and health trajectories were simulated over a 24-month period. The model’s parameters were based on patients who received short (≤ 3 days) or long (> 7 days) initial opioid prescriptions, matched using propensity score methods. All active-duty US Army soldiers from 2011 to 2014; the data contained detailed medical and administrative information on over 11 million soldier-months corresponding to 827,265 individual soldiers. Overall costs of a policy change, quality-adjusted life-years (QALYs) gained, and $/QALY gained. Over a 2-year horizon, a reassignment of 10,000 patients to short initial duration would generate a cost saving in the vicinity of $3.1 million (excluding program costs), and would also lead to an estimated 4451 additional opioid-free months, i.e. months without any opioid prescriptions. The analysis found that efforts to change prescriber behavior can be cost effective, and further studies into the implementation of such policies are warranted.

Author(s): Ren, Ai; Golden, Bruce; Alt, Frank; Wasil, Edward; Bjarnadottir, Margret; Hirshon, Jon Mark; Pimentel, Laura | Abstract: Introduction: On January 1, 2014, the State of Maryland implemented the Global Budget Revenue (GBR) program. We investigate the impact of GBR on length of stay (LOS) for inpatients in emergency departments (ED) in Maryland.Methods: We used the Hospital Compare data reports from the Centers for Medicare and Medicaid Services (CMS) and CMS Cost Reports Hospital Form 2552-10 from January 1, 2012–March 31, 2016, with GBR hospitals from Maryland and hospitals from West Virginia (WV), Delaware (DE), and Rhode Island (RI). We implemented difference-in-differences analysis and investigated the impact of GBR implementation on the LOS or ED1b scores of Maryland hospitals using a mixed-effects model with a state-level fixed effect, a hospital-level random effect, and state-level heterogeneity.Results: The GBR impact estimator was 9.47 (95% confidence interval [CI], 7.06 to 11.87, p-valuel0.001) for Maryland GBR hospitals, which implies, on average, that GBR implementation added 9.47 minutes per year to the time that hospital inpatients spent in the ED in the first two years after GBR implementation. The effect of the total number of hospital beds was 0.21 (95% CI, 0.089 to 0.330, p-value = 0 .001), which suggests that the bigger the hospital, the longer the ED1b score. The state-level fixed effects for WV were -106.96 (95% CI, -175.06 to -38.86, p-value = 0.002), for DE it was 6.51 (95% CI, -8.80 to 21.82, p-value=0.405), and for RI it was -54.48 (95% CI, -82.85 to -26.10, p-valuel0.001).Conclusion: Our results indicate that GBR implementation has had a statistically significant negative impact on the efficiency measure ED1b of Maryland hospital EDs from January 2014 to April 2016. We also found that the significant state-level fixed effect implies that the same inpatient might experience different ED processing times in each of the four states that we studied. [West J Emerg Med.

This case uses player evaluation and personnel decision making in Major League Baseball (MLB) to introduce many of the key steps of data analytics projects. The data analytics process is a unique c...

BACKGROUND CONTEXT Nonspecific neck and back pain are highly variable conditions with hundreds of treatment options. Using standard analytical techniques, it is difficult to understand how patients seek and receive health care across service locations. PURPOSE To build and display a comprehensive understanding of common utilization patterns for nonspecific neck and back pain. Also, to understand how observable patient characteristics (including insurance type, age and comorbidities) and initial sites of care (patient self-triage decisions to one of 9 initial sites of care including primary care, chiropractor and the emergency room) are associated with subsequent treatment and utilization patterns. STUDY DESIGN/SETTING This retrospective, observational study leveraged a US multipayer medical and pharmaceutical administrative claims dataset. Although not nationally representative, the data are comprised of commercial and Medicaid insured patients from 47 states and the District of Columbia. Each patient was eligible for at least 24 months in the database, enabling researchers to study utilization patterns over time and across service locations. PATIENT SAMPLE We identified a sample of 113,654 episodes for patients aged 18-64 who received care for nonspecific neck or back pain. Patients had a neck and back pain “clean” period of at least 12 months prior to initial presentation and were observed 6 months after the index claim. Subjects were excluded if their medical history and pain causes indicated potential deviation from conservative diagnostic and treatment practices. OUTCOME MEASURES The patient's initial site of care, the patient's subsequent diagnostic and treatment pattern (one of 14 clusters representing the 6 months after the index neck or back pain event), and common utilization metrics (eg, procedures, emergency care, and opioid prescription rates) within 6 months of the initial encounter. METHODS We applied a sequence alignment and density-based clustering methodology to identify treatment pathways based on the types and ordering of each patient's neck and back pain-related events for 6 months after the index claim. Using a series of multinomial models, we examined how observable patient characteristics were associated with the initial site of care and how the initial site of care was associated with the patient's subsequent treatment pathway. Prediction accuracy was evaluated using an 80% training, 20% testing division of the sample. We calculated health care utilization during the first 6 months after the index event. RESULTS The clustering methodology discovered 14 distinct patterns of how patients are diagnosed and treated for nonspecific neck and back pain in the first 6 months after the index encounter. The patterns ranged significantly in terms of utilization rates of advanced diagnostic imaging, opioid prescribing, and use of conservative therapy. The algorithm also identified 3.1% of episodes in which patients underwent uncommon treatment plans that included invasive procedures, multiple advanced images, and emergency or inpatient admissions. While some patient characteristics (including insurance type and geography) were associated with the patient's initial site of care, patient characteristics alone could not accurately predict this self-triage decision for an individual patient (47.9% vs 45.2% naive prediction). In contrast, the initial site of care is highly predictive of the subsequent treatment pathway (42.3% prediction accuracy using initial site of care alone vs 30.3% naive prediction). CONCLUSIONS The patient journey clustering methodology created a granular, data-driven understanding into how patients are diagnosed and treated after an index neck or back pain claim. The analyses suggest that the patient's self-triage decision to an initial site of care is difficult to predict using observable patient characteristics. However, once this initial site of care is chosen, the patient's subsequent treatment pattern and utilization within the first 6 months becomes more predictable. Follow-up studies on more generalizable samples are needed to use the data to proactively route patients toward high-quality, appropriate treatment plans. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.

In the United States, even though national guidelines for allocating scarce healthcare resources are lacking, 26 states, including New York State (NYS), have specific ventilator allocation guidelines to be invoked in case of a shortage. NYS developed these guidelines in 2015 as "pandemic influenza is a foreseeable threat, one that we cannot ignore." The primary objective of this study is to assess the existing procedures and priority rules in place for allocating/rationing scarce ventilator capacity and propose alternative (and improved) priority schemes. We first build machine learning models using inpatient records of COVID-19 patients admitted to New York-Presbyterian/Columbia University Irving Medical Center and an affiliated community health center to predict survival probabilities as well as ventilator length-of-use. Then, we use the resulting point estimators and their uncertainties as inputs for a multi-class priority queueing model with abandonments to assess three priority schemes: (i) SOFA-P, which most closely mimics the existing practice by prioritizing patients with sufficiently low Single Organ Failure Assessment (SOFA) scores, (ii) ISP, which assigns priority based on patient-level survival predictions, and (iii) ISP-LU, which takes into account survival predictions and resource use duration. Our findings highlight that our proposed priority scheme, ISP-LU, achieves a demonstrable improvement over the other two alternatives. Specifically, the expected number of survivals increases and death risk while waiting for ventilator use decreases. We also illustrate how priority schemes such as ISP with its sole focus on acute-phase survival odds may be discriminatory with respect to certain demographics, and highlight that ISP-LU allocates scarce healthcare capacity in a more equitable way.

Having accurate, unbiased prognosis information can help patients and providers make better decisions about what course of treatment to take. Using a comprehensive dataset of all colorectal cancer patients in California, we generate predictive models that estimate short‐term and medium‐term survival probabilities for patients based on their clinical and demographic information. Our study addresses some of the contradictions in the literature about survival rates and significantly improves predictive power over the performance of any model in previously published studies.

Bundled payments as a reimbursement mechanism have the potential to reduce health care expenditures and improve the quality of care by aligning the incentives of payers, providers and, most importantly, patients. The Centers for Medicare and Medicaid Services (CMS) launched the Bundled Payments for Care Improvement (BPCI) program in April 2013 and has set ambitious goals for adopting alternative payment models on a large scale. One of the crucial components for successful implementation of a bundled payment system is the identification of procedural homogeneous groups within an episode of care (a set of services needed to treat a medical condition), to which a flat reimbursement rate can be applied. In this study, we propose a data-driven clustering approach to automatically detect and explicitly represent homogeneous sub-groups of services for a given condition. Manual detection is slow and relies on consensus decisions, but automatic detection can serve as an important foundational input for bun...

When preparing administrative medical and pharmacy claims data for analysis, decisions about data clean up and analytical approach need to be made. However, information about the effects of various modelling decisions on adherence measures such as the medication possession ratio (MPR) is limited. We address this gap with this study. We utilized cross-sectional administrative claims data for commercially insured members filling at least two prescriptions for drugs within five classes of hypertension medication between 2008 and 2010. We divided nine modelling decisions into three categories: data scrubbing, study design, and MPR definition/calculations. We defined the base-case settings with commonly used values, varied each modelling decision singly and in combination, and measured the effects on the MPR. Claims data for 358,418 individuals were available for analysis. Two modelling decisions were found to be highly influential, each yielding a difference of over 25 percentage points from the base case: the decision of whether to use interval- or prescription-based study periods, and the decision of how to handle overlapping prescription claims. The effect of other decisions was smaller, with a difference of 1–9 percentage points from the base case. Some of the decisions considered had a large impact on the MPR. Therefore, it is important for researchers to standardize approaches for study period length and overlapping prescription claims. We also conclude that transparent reporting of modelling decisions will facilitate the interpretation of results and comparisons across studies.

Frequent emergency department (ED) users impose a significant burden on the healthcare system. Case management (CM) can target potential frequent users to reduce their ED utilization. As CM is cost...

Background Heart failure (HF) is associated with high mortality rates and high costs, and self-care is crucial in the management of the condition. Telehealth can promote patients’ self-care while providing frequent feedback to their health care providers about the patient’s compliance and symptoms. A number of technologies have been considered in the literature to facilitate telehealth in patients with HF. An important factor in the adoption of these technologies is their ease of use. Conversational agent technologies using a voice interface can be a good option because they use speech recognition to communicate with patients. Objective The aim of this paper is to study the engagement of patients with HF with voice interface technology. In particular, we investigate which patient characteristics are linked to increased technology use. Methods We used data from two separate HF patient groups that used different telehealth technologies over a 90-day period. Each group used a different type of voice interface; however, the scripts followed by the two technologies were identical. One technology was based on Amazon’s Alexa (Alexa+), and in the other technology, patients used a tablet to interact with a visually animated and voice-enabled avatar (Avatar). Patient engagement was measured as the number of days on which the patients used the technology during the study period. We used multiple linear regression to model engagement with the technology based on patients’ demographic and clinical characteristics and past technology use. Results In both populations, the patients were predominantly male and Black, had an average age of 55 years, and had HF for an average of 7 years. The only patient characteristic that was statistically different (P=.008) between the two populations was the number of medications they took to manage HF, with a mean of 8.7 (SD 4.0) for Alexa+ and 5.8 (SD 3.4) for Avatar patients. The regression model on the combined population shows that older patients used the technology more frequently (an additional 1.19 days of use for each additional year of age; P=.004). The number of medications to manage HF was negatively associated with use (−5.49; P=.005), and Black patients used the technology less frequently than other patients with similar characteristics (−15.96; P=.08). Conclusions Older patients’ higher engagement with telehealth is consistent with findings from previous studies, confirming the acceptability of technology in this subset of patients with HF. However, we also found that a higher number of HF medications, which may be correlated with a higher disease burden, is negatively associated with telehealth use. Finally, the lower engagement of Black patients highlights the need for further study to identify the reasons behind this lower engagement, including the possible role of social determinants of health, and potentially create technologies that are better tailored for this population.

We explore whether free agents in Major League Baseball meet the expectations set forth by newly signed contracts. The value and duration of these contracts are negotiated between the player and his agent and the signing team and are based primarily on the player's performance to date, projected future performance, and potential marketing value to the team. We develop two classes of models to explore this problem using a variety of regression- and tree-based machine learning algorithms. The market model uses player and team data to predict the market value of a player's performance i.e., average contract salary. The performance model uses the same data to predict wins above replacement as a surrogate for overall player performance. We translate this measure into dollars using position-based conversion factors. Analysis of these models demonstrates that the performance model more consistently predicts and assesses player value with respect to their free agent contracts. Together, these models can be used to target or avoid free agents or other players whose performance-based value differs significantly from their market value. © 2016 Wiley Periodicals, Inc. Statistical Analysis and Data Mining: The ASA Data Science Journal, 2016

We aim to explain the variation in provider charges based on Centers for Medicare and Medicaid Services’s (CMS) recently released data containing information on payments to and charges by health-care providers. The large data set includes operational, financial, and quality measures, as well as service identifiers. We evaluate how CMS’s published payment models explain payments and charges for inpatient and outpatient services, and employ additional models that incorporate available service- and provider-specific information. We found that the variation in payments is explained extremely well by CMS’s payment models, but these same models only explain provider charges to a limited extent. Efforts to include service- and provider-specific data and segment the data only marginally improve the fit of provider charges, leaving almost 30% of the variation of charges unexplained for all global models. Our analysis demonstrates that provider charges are highly variable and providers are using potentially diverse information and different methods to determine their prices.

Background: Optimization of health care delivery in different settings with the goal of improving quality of life and reduction of heart failure (HF) morbidity continues to be an unmet need. Digital engagement between patients and providers is becoming more prevalent for care coordination and management of heart failure.Aims: To design and test a customized and interactive heart failure disease specific functionality with activated technology (Amazon's Echo Dot or an Avatar) as a tool for management of HF patients at home. To identify the effect of a voice activated technology in the reduction of heart failure hospitalizations, improvement of medication adherence, improvement in health-related quality of life and increase in comfort with voice activated technologies. Methods We compared two types of conversational agent technologies to standard of care, using speech recognition, to evaluate patient engagement in HF patients. The first technologyutilized the Amazon Echo Dot Alexa technology and only communicated with the patient through audio and the second was an application on a tablet using an Avatar, combining audio and visual. The conversational agent asks the HF patients a series of 11 questions, divided into three components: compliance (questions 1-3), mild HF symptoms (questions 4-6) and moderate/severe HF symptoms (questions 7-11). Participants completed questions regarding their use and comfort with technology. Inclusion criteria: admitted or treated for HF;18 or older; and resides with access to wifi. Exclusion criteria: heart transplant or ventricular assist device recipients. Participants were monitored for three months.Results: Thirty patients were enrolled in each intervention arm of the two studies with the Alexa or Avatar. Mean age was 54 years in the Alexa group and 56 years in the Avatar group. There were predominately males 61% and 63%, Alexa and Avatar, respectively. In the Alexa and Avatar groups race was predominately black, 64% and 63% followed by white 21% and 30%. The Alexa group was statistically more likely to complete the three compliance factors of weight checking (p=.030), salt monitoring (p=0.016) and medication adherence (p=0.005). Initial results suggest that older people and individuals using the smart phone answer the questionnaire at higher rates, and individuals taking an increased number of medications to treat HF had lower participation.Conclusion: Voice enabled, internet connected devices are poised to have profound impacts on the quality of life for in-home, independent, and assisted-living patients by vastly increasing the connectivity to healthcare providers.

Case management is a $6 billion industry that employs over 34,000 people in the United States alone. Traditionally, case management has been utilized to help patients navigate the health care system and to coordinate care in the hope of lowering costs and achieving better health outcomes. However, since enrollment into these programs is typically either universal or limited to very sick patients, studies on the cost-effectiveness of case management programs find that their performance is mixed, at best. In this article we posit an opportunity to improve outcomes and lower costs by targeting certain patients for case management and early intervention. Utilizing modern data mining methods, we develop a methodology to identify these patients, who we describe as “jumpers” because their costs are currently low but will increase significantly in the near future. Given the performance of the prediction models, we also show that unless case management can prevent over 7.5% of health care cost increases, i...

Postmarketing drug surveillance is the process of monitoring the adverse events of pharmaceutical or medical devices after they are approved by the appropriate regulatory authorities. Historically, such surveillance was based on voluntary reports by medical practitioners, but with the widespread adoption of electronic medical records and comprehensive patient databases, surveillance systems that utilize such data are of considerable interest. Unfortunately, existing methods for analyzing the data in such systems ignore the open-ended exploratory nature of such systems that requires the assessment of multiple possible adverse events. In this article, we propose a method, SEQMEDS, that assesses the effect of a single drug on multiple adverse events by analyzing data that accumulate sequentially and explicitly captures interdependencies among the multiple events. The method continuously monitors a vector-valued test-statistic derived from the cumulative number of adverse events. It flags a potential adverse event once the test-statistic crosses a stopping boundary. We employ asymptotic analysis that assumes a large number of observations in a given window of time to show how to compute the stopping boundary by solving a convex optimization problem that achieves a desired Type I error and minimizes the expected time to detection under a pre-specified alternative hypothesis. We apply our method to a model in which the interdependency among the multiple adverse events is captured by a Cox proportional hazards model with time-dependent covariates and demonstrate that it provides an approximation of a fully sequential test for the maximum hazard ratio of the drug over multiple adverse events. A numerical study verifies that our method delivers Type I /II errors that are below pre-specified levels and is robust to distributional assumptions and parameter values.

Artificial Intelligence, especially deep learning based approaches, has been criticized for its \black-box" approach in recent years. The lack of interpretability of their performance makes people uncomfortable when applying these models in critical situations. In this study, we aim to contribute to the interpretability of the deep learning model in predicting healthcare costs. We propose to use long short-term memory based recurrent neural network to incorporate the sequential information for more accurate healthcare cost predictions. We first compare the performance of our deep learning approach with traditional machine learning methods including linear regression, lasso regression, ridge regression, and random forest. Among all the methods, deep learning shows the best performance. The superior performance of the deep learning model is further confirmed in subgroup analyses of patients. We then propose a novel interpretation method to examine how the deep learning model performs differently from other methods when facing fluctuations in the monthly costs. We find that while most traditional prediction models are getting worse with greater fluctuation in the data, the deep learning model can incorporate the fluctuation information and gain in prediction accuracy. Our work makes important contributions to the interpretability of deep learning models for more accurate prediction of healthcare costs.

INTRODUCTION The use of opioids has increased drastically over the past few years and decades. As a result, concerns have mounted over serious outcomes associated with chronic opioid use (COU), including dependency and death. A greater understanding of the factors that are associated with COU will be critical if prescribers are to navigate potentially competing objectives to provide compassionate care, while reducing the overall opioid use problem. In this study, we study pain levels and opioid prescription volumes and their effects on the risk of COU.This study leveraged passive data sources that support automated decision support systems (DSSs) currently employed in a large military population. The models presented compute monthly, person-specific, adjusted probability of subsequent COT and could potentially provide critical decision support for clinicians engaged in pain management. MATERIALS AND METHODS The study population included all outpatient presentations at military medical facilities worldwide among active duty United States Army soldiers during July 2011 to September 2014 (17,664,006 encounters; population N = 552,193). We conducted a retrospective cohort study of this population and employed longitudinal data and a discrete time multivariable logistic regression model to compute COT probability scores. The contribution of pain scores and opioid prescription quantities to the probability of COT represented analytic foci. RESULTS There were 13,891 subjects (2.5%) who experienced incident COT during the observed time period. Statistically significant interactions between pain scores and prescription quantity were present, in addition to effects of multiple other control variables. Counts of monthly opioid prescriptions and maximum stated pain scores per month were each positively associated with COT. A wide range in individual COT risk scores was evident. The effect of prescription volume on the COT risk was larger than the effect of the pain score, and the combined effect of larger pain scores and increased prescription quantity was moderated by the interaction term. CONCLUSIONS The results verified that passive data on the US Army can support a robust COT risk computation in this population. The individual, adjusted risk level requires statistical analyses to be fully understood. Because the same data sources drive current military DSSs, this work provides the potential basis for new, evidence-based decision support resources for military clinicians. The strong, independent impact of increasing opioid prescription counts on the COT risk reinforces the importance of exploring alternatives to opioids in pain management planning. It suggests that changing provider behavior through enhanced decision support could help reduce COT rates.

The rising cost of health care is one of the world's most important problems. Accordingly, predicting such costs with accuracy is a significant first step in addressing this problem. Since the 1980s, there has been research on the predictive modeling of medical costs based on (health insurance) claims data using heuristic rules and regression methods. These methods, however, have not been appropriately validated using populations that the methods have not seen. We utilize modern data-mining methods, specifically classification trees and clustering algorithms, along with claims data from over 800,000 insured individuals over three years, to provide rigorously validated predictions of health-care costs in the third year, based on medical and cost data from the first two years. We quantify the accuracy of our predictions using unseen (out-of-sample) data from over 200,000 members. The key findings are: (a) our data-mining methods provide accurate predictions of medical costs and represent a powerful tool for prediction of health-care costs, (b) the pattern of past cost data is a strong predictor of future costs, and (c) medical information only contributes to accurate prediction of medical costs of high-cost members.

Advanced computing capabilities and novel visual analytics tools now allow us to move beyond the traditional cross-sectional summaries to analyze longitudinal prescription patterns and the impact of study design decisions. For example, design decisions regarding gaps and overlaps in prescription fill data are necessary for measuring adherence using prescription claims data. However, little is known regarding the impact of these decisions on measures of medication possession (e.g., medication possession ratio). The goal of the study was to demonstrate the use of visualization tools for pattern discovery, hypothesis generation, and study design. We utilized EventFlow, a novel discrete event sequence visualization software, to investigate patterns of prescription fills, including gaps and overlaps, utilizing large-scale healthcare claims data. The study analyzes data of individuals who had at least two prescriptions for one of five hypertension medication classes: ACE inhibitors, angiotensin II receptor blockers, beta blockers, calcium channel blockers, and diuretics. We focused on those members initiating therapy with diuretics (19.2 %) who may have concurrently or subsequently take drugs in other classes as well. We identified longitudinal patterns in prescription fills for antihypertensive medications, investigated the implications of decisions regarding gap length and overlaps, and examined the impact on the average cost and adherence of the initial treatment episode. A total of 790,609 individuals are included in the study sample, 19.2 % (N = 151,566) of whom started on diuretics first during the study period. The average age was 52.4 years and 53.1 % of the population was female. When the allowable gap was zero, 34 % of the population had continuous coverage and the average length of continuous coverage was 2 months. In contrast, when the allowable gap was 30 days, 69 % of the population showed a single continuous prescription period with an average length of 5 months. The average prescription cost of the period of continuous coverage ranged from US$3.44 (when the maximum gap was 0 day) to US$9.08 (when the maximum gap was 30 days). Results were less impactful when considering overlaps. This proof-of-concept study illustrates the use of visual analytics tools in characterizing longitudinal medication possession. We find that prescription patterns and associated prescription costs are more influenced by allowable gap lengths than by definitions and treatment of overlap. Research using medication gaps and overlaps to define medication possession in prescription claims data should pay particular attention to the definition and use of gap lengths.

After the withdrawal of rofecoxib (known by the trade name Vioxx) from the US pharmaceutical market in 2004, post-approval drug safety and surveillance came under serious scrutiny. In 2008 the FDA announced the Sentinel Initiative, which includes an active surveillance system based on 100 million people’s health-care data. In this chapter we describe a number of challenges involved in active drug and vaccine surveillance and provide an overview of state-of-the-art surveillance methodologies. We also address the statistical tradeo-ffs involved in surveillance, highlight some areas for future research, and frame the policy issues that designers of surveillance systems will have to address.