Prevalence of lifestyle-related behavioral information in claims data in the U.S.

OBJECTIVE: Given their association with varying health risks, lifestyle-related behaviors are essential to consider in population-level disease prevention. Health insurance claims are a key source of information for population health analytics, but the availability of lifestyle information within claims data is unknown. Our goal was to assess the availability and prevalence of data items that describe lifestyle behaviors across several domains within a large U.S. claims database. METHODS: We conducted a retrospective, descriptive analysis to determine the availability of the following claims-derived lifestyle domains: nutrition, eating habits, physical activity, weight status, emotional wellness, sleep, tobacco use, and substance use. To define these domains, we applied a serial review process with three physicians to identify relevant diagnosis and procedure codes within claims for each domain. We used enrollment files and medical claims from a large national U.S. health plan to identify lifestyle relevant codes filed between 2016 and 2020. We calculated the annual prevalence of each claims-derived lifestyle domain and the proportion of patients by count within each domain. RESULTS: Approximately half of all members within the sample had claims information that identified at least one lifestyle domain (2016 = 41.9%; 2017 = 46.1%; 2018 = 49.6%; 2019 = 52.5%; 2020 = 50.6% of patients). Most commonly identified domains were weight status (19.9-30.7% across years), nutrition (13.3-17.8%), and tobacco use (7.9-9.8%). CONCLUSION: Our study demonstrates the feasibility of using claims data to identify key lifestyle behaviors. Additional research is needed to confirm the accuracy and validity of our approach and determine its use in population-level disease prevention.

Assessing racial bias in healthcare predictive models: Practical lessons from an empirical evaluation of 30-day hospital readmission models.

OBJECTIVE: Despite increased availability of methodologies to identify algorithmic bias, the operationalization of bias evaluation for healthcare predictive models is still limited. Therefore, this study proposes a process for bias evaluation through an empirical assessment of common hospital readmission models. The process includes selecting bias measures, interpretation, determining disparity impact and potential mitigations. METHODS: This retrospective analysis evaluated racial bias of four common models predicting 30-day unplanned readmission (i.e., LACE Index, HOSPITAL Score, and the CMS readmission measure applied as is and retrained). The models were assessed using 2.4 million adult inpatient discharges in Maryland from 2016 to 2019. Fairness metrics that are model-agnostic, easy to compute, and interpretable were implemented and apprised to select the most appropriate bias measures. The impact of changing model's risk thresholds on these measures was further assessed to guide the selection of optimal thresholds to control and mitigate bias. RESULTS: Four bias measures were selected for the predictive task: zero-one-loss difference, false negative rate (FNR) parity, false positive rate (FPR) parity, and generalized entropy index. Based on these measures, the HOSPITAL score and the retrained CMS measure demonstrated the lowest racial bias. White patients showed a higher FNR while Black patients resulted in a higher FPR and zero-one-loss. As the models' risk threshold changed, trade-offs between models' fairness and overall performance were observed, and the assessment showed all models' default thresholds were reasonable for balancing accuracy and bias. CONCLUSIONS: This study proposes an Applied Framework to Assess Fairness of Predictive Models (AFAFPM) and demonstrates the process using 30-day hospital readmission model as the example. It suggests the feasibility of applying algorithmic bias assessment to determine optimized risk thresholds so that predictive models can be used more equitably and accurately. It is evident that a combination of qualitative and quantitative methods and a multidisciplinary team are necessary to identify, understand and respond to algorithm bias in real-world healthcare settings. Users should also apply multiple bias measures to ensure a more comprehensive, tailored, and balanced view. The results of bias measures, however, must be interpreted with caution and consider the larger operational, clinical, and policy context.

Evaluating Algorithmic Bias in 30-Day Hospital Readmission Models: Retrospective Analysis.

BACKGROUND: The adoption of predictive algorithms in health care comes with the potential for algorithmic bias, which could exacerbate existing disparities. Fairness metrics have been proposed to measure algorithmic bias, but their application to real-world tasks is limited. OBJECTIVE: This study aims to evaluate the algorithmic bias associated with the application of common 30-day hospital readmission models and assess the usefulness and interpretability of selected fairness metrics. METHODS: We used 10.6 million adult inpatient discharges from Maryland and Florida from 2016 to 2019 in this retrospective study. Models predicting 30-day hospital readmissions were evaluated: LACE Index, modified HOSPITAL score, and modified Centers for Medicare & Medicaid Services (CMS) readmission measure, which were applied as-is (using existing coefficients) and retrained (recalibrated with 50% of the data). Predictive performances and bias measures were evaluated for all, between Black and White populations, and between low- and other-income groups. Bias measures included the parity of false negative rate (FNR), false positive rate (FPR), 0-1 loss, and generalized entropy index. Racial bias represented by FNR and FPR differences was stratified to explore shifts in algorithmic bias in different populations. RESULTS: The retrained CMS model demonstrated the best predictive performance (area under the curve: 0.74 in Maryland and 0.68-0.70 in Florida), and the modified HOSPITAL score demonstrated the best calibration (Brier score: 0.16-0.19 in Maryland and 0.19-0.21 in Florida). Calibration was better in White (compared to Black) populations and other-income (compared to low-income) groups, and the area under the curve was higher or similar in the Black (compared to White) populations. The retrained CMS and modified HOSPITAL score had the lowest racial and income bias in Maryland. In Florida, both of these models overall had the lowest income bias and the modified HOSPITAL score showed the lowest racial bias. In both states, the White and higher-income populations showed a higher FNR, while the Black and low-income populations resulted in a higher FPR and a higher 0-1 loss. When stratified by hospital and population composition, these models demonstrated heterogeneous algorithmic bias in different contexts and populations. CONCLUSIONS: Caution must be taken when interpreting fairness measures' face value. A higher FNR or FPR could potentially reflect missed opportunities or wasted resources, but these measures could also reflect health care use patterns and gaps in care. Simply relying on the statistical notions of bias could obscure or underplay the causes of health disparity. The imperfect health data, analytic frameworks, and the underlying health systems must be carefully considered. Fairness measures can serve as a useful routine assessment to detect disparate model performances but are insufficient to inform mechanisms or policy changes. However, such an assessment is an important first step toward data-driven improvement to address existing health disparities.

Patterns of Morbidity Across the Lifespan: A Population Segmentation Framework for Classifying Health Care Needs for All Ages.

BACKGROUND: Classification systems to segment such patients into subgroups for purposes of care management and population analytics should balance administrative simplicity with clinical meaning and measurement precision. OBJECTIVE: To describe and empirically apply a new clinically relevant population segmentation framework applicable to all payers and all ages across the lifespan. RESEARCH DESIGN AND SUBJECTS: Cross-sectional analyses using insurance claims database for 3.31 Million commercially insured and 1.05 Million Medicaid enrollees under 65 years old; and 5.27 Million Medicare fee-for-service beneficiaries aged 65 and older. MEASURES: The "Patient Need Groups" (PNGs) framework, we developed, classifies each person within the entire 0-100+ aged population into one of 11 mutually exclusive need-based categories. For each PNG segment, we documented a range of clinical and resource endpoints, including health care resource use, avoidable emergency department visits, hospitalizations, behavioral health conditions, and social need factors. RESULTS: The PNG categories included: (1) nonuser, (2) low-need child, (3) low-need adult, (4) low-complexity multimorbidity, (5) medium-complexity multimorbidity, (6) low-complexity pregnancy, (7) high-complexity pregnancy, (8) dominant psychiatric/behavioral condition, (9) dominant major chronic condition, (10) high-complexity multimorbidity, and (11) frailty. Each PNG evidenced a characteristic age-related trajectory across the full lifespan. In addition to offering clinically cogent groupings, large percentages (29%-62%) of patients in two pregnancy and high-complexity multimorbidity and frailty PNGs were in a high-risk subgroup (upper 10%) of potential future health care utilization. CONCLUSIONS: The PNG population segmentation approach represents a comprehensive measurement framework that captures and categorizes available electronic health care data to characterize individuals of all ages based on their needs.

Early relaxation of community mitigation policies and risk of COVID-19 resurgence in the United States.

This study aimed to assess the impact of coronavirus disease (COVID-19) prevalence in the United States in the week leading to the relaxation of the stay-at-home orders (SAH) on future prevalence across states that implemented different SAH policies. We used data on the number of confirmed COVID-19 cases as of August 21, 2020 on county level. We classified states into four groups based on the 7-day change in prevalence and the state's approach to SAH policy. The groups included: (1) High Change (19 states; 7-day prevalence change ≥50th percentile), (2) Low Change (19 states; 7-day prevalence change <50th percentile), (3) No SAH (11 states: did not adopt SAH order), and (4) No SAH End (2 states: did not relax SAH order). We performed regression modeling assessing the association between change in prevalence at the time of SAH order relaxation and COVID-19 prevalence days after the relaxation of SAH order for four selected groups. After adjusting for other factors, compared to the High Change group, counties in the Low Change group had 33.8 (per 100,000 population) fewer cases (standard error (SE): 19.8, p < 0.001) 7 days after the relaxation of SAH order and the difference was larger by time passing. On August 21, 2020, the No SAH End group had 383.1 fewer cases (per 100,000 population) than the High Change group (SE: 143.6, p < 0.01). A measured, evidence-based approach is required to safely relax the community mitigation strategies and practice phased-reopening of the country.

Differential impact of mitigation policies and socioeconomic status on COVID-19 prevalence and social distancing in the United States.

BACKGROUND: The spread of COVID-19 has highlighted the long-standing health inequalities across the U.S. as neighborhoods with fewer resources were associated with higher rates of COVID-19 transmission. Although the stay-at-home order was one of the most effective methods to contain its spread, residents in lower-income neighborhoods faced barriers to practicing social distancing. We aimed to quantify the differential impact of stay-at-home policy on COVID-19 transmission and residents' mobility across neighborhoods of different levels of socioeconomic disadvantage. METHODS: This was a comparative interrupted time-series analysis at the county level. We included 2087 counties from 38 states which both implemented and lifted the state-wide stay-at-home order. Every county was assigned to one of four equally-sized groups based on its levels of disadvantage, represented by the Area Deprivation Index. Prevalence of COVID-19 was calculated by dividing the daily number of cumulative confirmed COVID-19 cases by the number of residents from the 2010 Census. We used the Social Distancing Index (SDI), derived from the COVID-19 Impact Analysis Platform, to measure the mobility. For the evaluation of implementation, the observation started from Mar 1st 2020 to 1 day before lifting; and, for lifting, it ranged from 1 day after implementation to Jul 5th 2020. We calculated a comparative change of daily trends in COVID-19 prevalence and Social Distancing Index between counties with three highest disadvantage levels and those with the least level before and after the implementation and lifting of the stay-at-home order, separately. RESULTS: On both stay-at-home implementation and lifting dates, COVID-19 prevalence was much higher among counties with the highest or lowest disadvantage level, while mobility decreased as the disadvantage level increased. Mobility of the most disadvantaged counties was least impacted by stay-at-home implementation and relaxation compared to counties with the most resources; however, disadvantaged counties experienced the largest relative increase in COVID-19 infection after both stay-at-home implementation and relaxation. CONCLUSIONS: Neighborhoods with varying levels of socioeconomic disadvantage reacted differently to the implementation and relaxation of COVID-19 mitigation policies. Policymakers should consider investing more resources in disadvantaged counties as the pandemic may not stop until most neighborhoods have it under control.

Primary care practices' ability to predict future risk of expenditures and hospitalization using risk stratification and segmentation.

BACKGROUND: Patients with complex health care needs may suffer adverse outcomes from fragmented and delayed care, reducing well-being and increasing health care costs. Health reform efforts, especially those in primary care, attempt to mitigate risk of adverse outcomes by better targeting resources to those most in need. However, predicting who is susceptible to adverse outcomes, such as unplanned hospitalizations, ED visits, or other potentially avoidable expenditures, can be difficult, and providing intensive levels of resources to all patients is neither wanted nor efficient. Our objective was to understand if primary care teams can predict patient risk better than standard risk scores. METHODS: Six primary care practices risk stratified their entire patient population over a 2-year period, and worked to mitigate risk for those at high risk through care management and coordination. Individual patient risk scores created by the practices were collected and compared to a common risk score (Hierarchical Condition Categories) in their ability to predict future expenditures, ED visits, and hospitalizations. Accuracy of predictions, sensitivity, positive predictive values (PPV), and c-statistics were calculated for each risk scoring type. Analyses were stratified by whether the practice used intuition alone, an algorithm alone, or adjudicated an algorithmic risk score. RESULTS: In all, 40,342 patients were risk stratified. Practice scores had 38.6% agreement with HCC scores on identification of high-risk patients. For the 3,381 patients with reliable outcomes data, accuracy was high (0.71-0.88) but sensitivity and PPV were low (0.16-0.40). Practice-created scores had 0.02-0.14 lower sensitivity, specificity and PPV compared to HCC in prediction of outcomes. Practices using adjudication had, on average, .16 higher sensitivity. CONCLUSIONS: Practices using simple risk stratification techniques had slightly worse accuracy in predicting common outcomes than HCC, but adjudication improved prediction.

Performance of a Predictive Model versus Prescription-Based Thresholds in Identifying Patients at Risk of Fatal Opioid Overdose.

Background: Prescription Drug Monitoring Programs (PDMPs) collect controlled substance prescriptions dispensed within a state. Many PDMP programs perform targeted outreach (i.e., "unsolicited reporting") for patients who exceed numerical thresholds, however, the degree to which patients at highest risk of fatal opioid overdose are identified has not been compared with one another or with a predictive model. Methods: A retrospective analysis was performed using statewide PDMP data for Maryland residents aged 18 to 80 years with an opioid fill between April to June 2015. The outcome was opioid-related overdose death in 2015 or 2016. A multivariable logistic regression model and three PDMP thresholds were evaluated: (1) multiple provider episodes; (2) high daily average morphine milligram equivalents (MME); and (3) overlapping opioid and benzodiazepine prescriptions. Results: The validation cohort consisted of 170,433 individuals and 244 deaths. The predictive model captured more individuals who died (46.3% of total deaths) and had a higher death rate (7.12 per 1000) when the risk score cutoff (0.0030) was selected for a comparable size of high-risk individuals (n = 15,881) than those meeting the overlapping opioid/benzodiazepine prescriptions (n = 17,440; 33.2% of total deaths; 4.64 deaths per 1000) and high MME (n = 14,675; 24.6% of total deaths; 4.09 deaths per 1000) thresholds. Conclusions: The predictive model identified more individuals at risk of fatal opioid overdose as compared with PDMP thresholds commonly used for unsolicited reporting. PDMP programs could improve their targeting of unsolicited reports to reach more individuals at risk of overdose by using predictive models instead of simple threshold-based approaches.

Internet Access, Social Risk Factors, and Web-Based Social Support Seeking Behavior: Assessing Correlates of the "Digital Divide" Across Neighborhoods in The State of Maryland.

We aimed to empirically measure the degree to which there is a "digital divide" in terms of access to the internet at the small-area community level within the State of Maryland and the City of Baltimore and to assess the relationship and association of this divide with community-level SDOH risk factors, community-based social service agency location, and web-mediated support service seeking behavior. To assess the socio-economic characteristics of the neighborhoods across the state, we calculated the Area Deprivation Index (ADI) using the U.S. Census, American Community Survey (5-year estimates) of 2017. To assess the digital divide, at the community level, we used the Federal Communications Commission (FCC) data on the number of residential fixed Internet access service connections. We assessed the availability of and web-based access to community-based social service agencies using data provided by the "Aunt Bertha" information platform. We performed community and regional level descriptive and special analyses for ADI social risk factors, connectivity, and both the availability of and web-based searches for community-based social services. To help assess potential neighborhood linked factors associated with the rates of web-based social services searches by individuals in need, we applied logistic regression using generalized estimating equation modeling. Baltimore City contained more disadvantaged neighborhoods compared to other areas in Maryland. In Baltimore City, 20.3% of neighborhoods (defined by census block groups) were disadvantaged with ADI at the 90th percentile while only 6.6% of block groups across Maryland were in this disadvantaged category. Across the State, more than half of all census tracts had 801-1000 households (per 1000 households) with internet subscription. In contrast, in Baltimore City about half of all census tracts had only 401-600 of the households (per 1000 households) with internet subscriptions. Most block groups in Maryland and Baltimore City lacked access to social services facilities (61% of block groups at the 90th percentile of disadvantage in Maryland and 61.3% of block groups at the 90th percentile of disadvantage in Baltimore City). After adjusting for other variables, a 1% increase in the ADI measure of social disadvantage, resulting in a 1.7% increase in the number of individuals seeking social services. While more work is needed, our findings support the premise that the digital divide is closely associated with other SDOH factors. The policymakers must propose policies to address the digital divide on a national level and also in disadvantaged communities experiencing the digital divide in addition to other SDOH challenges.

The Impact of Various Risk Assessment Time Frames on the Performance of Opioid Overdose Forecasting Models.

BACKGROUND: An individual's risk for future opioid overdoses is usually assessed using a 12-month "lookback" period. Given the potential urgency of acting rapidly, we compared the performance of alternative predictive models with risk information from the past 3, 6, 9, and 12 months. METHODS: We included 1,014,033 Maryland residents aged 18-80 with at least 1 opioid prescription and no recorded death in 2015. We used 2015 Maryland prescription drug monitoring data to identify risk factors for nonfatal opioid overdoses from hospital discharge records and investigated fatal opioid overdose from medical examiner data in 2016. Prescription drug monitoring program-derived predictors included demographics, payment sources for opioid prescriptions, count of unique opioid prescribers and pharmacies, and quantity and types of opioids and benzodiazepines filled. We estimated a series of logistic regression models that included 3, 6, 9, and 12 months of prescription drug monitoring program data and compared model performance, using bootstrapped C-statistics and associated 95% confidence intervals. RESULTS: For hospital-treated nonfatal overdose, the C-statistic increased from 0.73 for a model including only the fourth quarter to 0.77 for a model with 4 quarters of data. For fatal overdose, the area under the curve increased from 0.80 to 0.83 over the same models. The strongest predictors of overdose were prescription fills for buprenorphine and Medicaid and Medicare as sources of payment. CONCLUSIONS: Models predicting opioid overdose using 1 quarter of data were nearly as accurate as models using all 4 quarters. Models with a single quarter may be more timely and easier to identify persons at risk of an opioid overdose.

Predictors of Overdose Death Among High-Risk Emergency Department Patients With Substance-Related Encounters: A Data Linkage Cohort Study.

STUDY OBJECTIVE: Persons with substance use disorders frequently utilize emergency department (ED) services, creating an opportunity for intervention and referral to addiction treatment and harm-reduction services. However, EDs may not have the appropriate tools to distinguish which patients are at greatest risk for negative outcomes. We link hospital ED and medical examiner mortality databases in one state to identify individual-level risk factors associated with overdose death among ED patients with substance-related encounters. METHODS: This retrospective cohort study linked Maryland statewide ED hospital claims records for adults with nonfatal overdose or substance use disorder encounters in 2014 to 2015 with medical examiner mortality records in 2015 to 2016. Logistic regression was used to identify factors in hospital records associated with risk of opioid overdose death. Predicted probabilities for overdose death were calculated for hypothetical patients with different combinations of overdose and substance use diagnostic histories. RESULTS: A total of 139,252 patients had substance-related ED encounters in 2014 to 2015. Of these patients, 963 later experienced an opioid overdose death, indicating a case fatality rate of 69.2 per 10,000 patients, 6 times higher than that of patients who used the ED for any cause. Factors most strongly associated with death included having both an opioid and another substance use disorder (adjusted odds ratio 2.88; 95% confidence interval 2.04 to 4.07), having greater than or equal to 3 previous nonfatal overdoses (adjusted odds ratio 2.89; 95% confidence interval 1.54 to 5.43), and having a previous nonfatal overdose involving heroin (adjusted odds ratio 2.24; 95% confidence interval 1.64 to 3.05). CONCLUSION: These results highlight important differences in overdose risk among patients receiving care in EDs for substance-related conditions. The findings demonstrate the potential utility of incorporating routine data from patient records to assess risk of future negative outcomes and identify primary targets for initiation and linkage to lifesaving care.

Healthcare costs and utilization associated with high-risk prescription opioid use: a retrospective cohort study.

BACKGROUND: Previous studies on high-risk opioid use have only focused on patients diagnosed with an opioid disorder. This study evaluates the impact of various high-risk prescription opioid use groups on healthcare costs and utilization. METHODS: This is a retrospective cohort study using QuintilesIMS health plan claims with independent variables from 2012 and outcomes from 2013. We included a population-based sample of 191,405 non-elderly adults with known sex, one or more opioid prescriptions, and continuous enrollment in 2012 and 2013. Three high-risk opioid use groups were identified in 2012 as (1) persons with 100+ morphine milligram equivalents per day for 90+ consecutive days (chronic users); (2) persons with 30+ days of concomitant opioid and benzodiazepine use (concomitant users); and (3) individuals diagnosed with an opioid use disorder. The length of time that a person had been characterized as a high-risk user was measured. Three healthcare costs (total, medical, and pharmacy costs) and four binary utilization indicators (the top 5% total cost users, the top 5% pharmacy cost users, any hospitalization, and any emergency department visit) derived from 2013 were outcomes. We applied a generalized linear model (GLM) with a log-link function and gamma distribution for costs while logistic regression was employed for utilization indicators. We also adopted propensity score weighting to control for the baseline differences between high-risk and non-high-risk opioid users. RESULTS: Of individuals with one or more opioid prescription, 1.45% were chronic users, 4.81% were concomitant users, and 0.94% were diagnosed as having an opioid use disorder. After adjustment and propensity score weighting, chronic users had statistically significant higher prospective total (40%), medical (3%), and pharmacy (172%) costs. The increases in total, medical, and pharmacy costs associated with concomitant users were 13%, 7%, and 41%, and 28%, 21% and 63% for users with a diagnosed opioid use disorder. Both total and pharmacy costs increased with the length of time characterized as high-risk users, with the increase being statistically significant. Only concomitant users were associated with a higher odds of hospitalization or emergency department use. CONCLUSIONS: Individuals with high-risk prescription opioid use have significantly higher healthcare costs and utilization than their counterparts, especially those with chronic high-dose opioid use.

Assessing the Impact of Body Mass Index Information on the Performance of Risk Adjustment Models in Predicting Health Care Costs and Utilization.

BACKGROUND: Using electronic health records (EHRs) for population risk stratification has gained attention in recent years. Compared with insurance claims, EHRs offer novel data types (eg, vital signs) that can potentially improve population-based predictive models of cost and utilization. OBJECTIVE: To evaluate whether EHR-extracted body mass index (BMI) improves the performance of diagnosis-based models to predict concurrent and prospective health care costs and utilization. METHODS: We used claims and EHR data over a 2-year period from a cohort of continuously insured patients (aged 20-64 y) within an integrated health system. We examined the addition of BMI to 3 diagnosis-based models of increasing comprehensiveness (ie, demographics, Charlson, and Dx-PM model of the Adjusted Clinical Group system) to predict concurrent and prospective costs and utilization, and compared the performance of models with and without BMI. RESULTS: The study population included 59,849 patients, 57% female, with BMI class I, II, and III comprising 19%, 9%, and 6% of the population. Among demographic models, R improvement from adding BMI ranged from 61% (ie, R increased from 0.56 to 0.90) for prospective pharmacy cost to 29% (1.24-1.60) for concurrent medical cost. Adding BMI to demographic models improved the prediction of all binary service-linked outcomes (ie, hospitalization, emergency department admission, and being in top 5% total costs) with area under the curve increasing from 2% (0.602-0.617) to 7% (0.516-0.554). Adding BMI to Charlson models only improved total and medical cost predictions prospectively (13% and 15%; 4.23-4.79 and 3.30-3.79), and also improved predicting all prospective outcomes with area under the curve increasing from 3% (0.649-0.668) to 4% (0.639-0.665; and, 0.556-0.576). No improvements in prediction were seen in the most comprehensive model (ie, Dx-PM). DISCUSSION: EHR-extracted BMI levels can be used to enhance predictive models of utilization especially if comprehensive diagnostic data are missing.

Defining and Assessing Geriatric Risk Factors and Associated Health Care Utilization Among Older Adults Using Claims and Electronic Health Records.

BACKGROUND: Using electronic health records (EHRs), in addition to claims, to systematically identify patients with factors associated with adverse outcomes (geriatric risk) among older adults can prove beneficial for population health management and clinical service delivery. OBJECTIVE: To define and compare geriatric risk factors derivable from claims, structured EHRs, and unstructured EHRs, and estimate the relationship between geriatric risk factors and health care utilization. RESEARCH DESIGN: We performed a retrospective cohort study of patients enrolled in a Medicare Advantage plan from 2011 to 2013 using both administrative claims and EHRs. We defined 10 individual geriatric risk factors and a summary geriatric risk index based on diagnosed conditions and pattern matching techniques applied to EHR free text. The prevalence of geriatric risk factors was estimated using claims, structured EHRs, and structured and unstructured EHRs combined. The association of geriatric risk index with any occurrence of hospitalizations, emergency department visits, and nursing home visits were estimated using logistic regression adjusted for demographic and comorbidity covariates. RESULTS: The prevalence of geriatric risk factors increased after adding unstructured EHR data to structured EHRs, compared with those derived from structured EHRs alone and claims alone. On the basis of claims, structured EHRs, and structured and unstructured EHRs combined, 12.9%, 15.0%, and 24.6% of the patients had 1 geriatric risk factor, respectively; 3.9%, 4.2%, and 15.8% had ≥2 geriatric risk factors, respectively. Statistically significant association between geriatric risk index and health care utilization was found independent of demographic and comorbidity covariates. For example, based on claims, estimated odds ratios for having 1 and ≥2 geriatric risk factors in year 1 were 1.49 (P<0.001) and 2.62 (P<0.001) in predicting any occurrence of hospitalizations in year 1, and 1.32 (P<0.001) and 1.34 (P=0.003) in predicting any occurrence of hospitalizations in year 2. CONCLUSIONS: The results demonstrate the feasibility and potential of using EHRs and claims for collecting new types of geriatric risk information that could augment the more commonly collected disease information to identify and move upstream the management of high-risk cases among older patients.

Comparing Population-based Risk-stratification Model Performance Using Demographic, Diagnosis and Medication Data Extracted From Outpatient Electronic Health Records Versus Administrative Claims.

BACKGROUND: There is an increasing demand for electronic health record (EHR)-based risk stratification and predictive modeling tools at the population level. This trend is partly due to increased value-based payment policies and the increasing availability of EHRs at the provider level. Risk stratification models, however, have been traditionally derived from claims or encounter systems. This study evaluates the challenges and opportunities of using EHR data instead of or in addition to administrative claims for risk stratification. METHODS: This study used the structured EHR records and administrative claims of 85,581 patients receiving outpatient care at a large integrated provider system. Common data elements for risk stratification (ie, age, sex, diagnosis, and medication) were extracted from outpatient EHR records and administrative claims. The performance of a validated risk-stratification model was assessed using data extracted from claims alone, EHR alone, and claims and EHR combined. RESULTS: EHR-derived metrics overlapped considerably with administrative claims (eg, number of chronic conditions). The accuracy of the model, when using EHR data alone, was acceptable with an area under the curve of ∼0.81 for hospitalization and ∼0.85 for identifying top 1% utilizers using the concurrent model. However, when using EHR data alone, the predictive model explained a lower amount of variation in utilization-based outcomes compared with administrative claims. DISCUSSION: The results show a promising performance of models predicting cost and hospitalization using outpatient EHR's diagnosis and medication data. More research is needed to evaluate the benefits of other EHR data types (eg, lab values and vital signs) for risk stratification.

Evaluating the Impact of Prescription Fill Rates on Risk Stratification Model Performance.

BACKGROUND: Risk adjustment models are traditionally derived from administrative claims. Prescription fill rates-extracted by comparing electronic health record prescriptions and pharmacy claims fills-represent a novel measure of medication adherence and may improve the performance of risk adjustment models. OBJECTIVE: We evaluated the impact of prescription fill rates on claims-based risk adjustment models in predicting both concurrent and prospective costs and utilization. METHODS: We conducted a retrospective cohort study of 43,097 primary care patients from HealthPartners network between 2011 and 2012. Diagnosis and/or pharmacy claims of 2011 were used to build 3 base models using the Johns Hopkins ACG system, in addition to demographics. Model performances were compared before and after adding 3 types of prescription fill rates: primary 0-7 days, primary 0-30 days, and overall. Overall fill rates utilized all ordered prescriptions from electronic health record while primary fill rates excluded refill orders. RESULTS: The overall, primary 0-7, and 0-30 days fill rates were 72.30%, 59.82%, and 67.33%. The fill rates were similar between sexes but varied across different medication classifications, whereas the youngest had the highest rate. Adding fill rates modestly improved the performance of all models in explaining medical costs (improving concurrent R by 1.15% to 2.07%), followed by total costs (0.58% to 1.43%), and pharmacy costs (0.07% to 0.65%). The impact was greater for concurrent costs compared with prospective costs. Base models without diagnosis information showed the highest improvement using prescription fill rates. CONCLUSIONS: Prescription fill rates can modestly enhance claims-based risk prediction models; however, population-level improvements in predicting utilization are limited.

Identifying Consistent High-cost Users in a Health Plan: Comparison of Alternative Prediction Models.

BACKGROUND: High-cost users in a period may not incur high-cost utilization in the next period. Consistent high-cost users (CHUs) may be better targets for cost-saving interventions. OBJECTIVES: To compare the characteristics of CHUs (patients with plan-specific top 20% medical costs in all 4 half-year periods across 2008 and 2009) and point high-cost users (PHUs) (top users in 2008 alone), and to build claims-based models to identify CHUs. RESEARCH DESIGN: This is a retrospective cohort study. Logistic regression was used to predict being CHUs. Independent variables were derived from 2007 claims; 5 models with different sets of independent variables (prior costs, medications, diagnoses, medications and diagnoses, medications and diagnoses and prior costs) were constructed. SUBJECTS: Three-year continuous enrollees aged from 18 to 62 years old from a large administrative database with $100 or more yearly costs (N=1,721,992). MEASURES: Correlation, overlap, and characteristics of top risk scorers derived from 5 CHUs models were presented. C-statistics, sensitivity, and positive predictive value were calculated. RESULTS: CHUs were characterized by having increasing total and pharmacy costs over 2007-2009, and more baseline chronic and psychosocial conditions than PHUs. Individuals' risk scores derived from CHUs models were moderately correlated (∼0.6). The medication-only model performed better than the diagnosis-only model and the prior-cost model. CONCLUSIONS: Five models identified different individuals as potential CHUs. The recurrent medication utilization and a high prevalence of chronic and psychosocial conditions are important in differentiating CHUs from PHUs. For cost-saving interventions with long-term impacts or focusing on medication, CHUs may be better targets.

Patient sharing and quality of care: measuring outcomes of care coordination using claims data.

BACKGROUND: With the goal of improving clinical efficiency and effectiveness, programs to enhance care coordination are a major focus of health care reform. OBJECTIVE: To examine whether "care density"--a claims-based measure of patient sharing by office-based physicians--is associated with measures of quality. Care density is a proxy measure that may reflect how frequently a patient's doctors collaborate. RESEARCH DESIGN: Cohort study using administrative databases from 3 large commercial insurance plans. SUBJECTS: A total of 1.7 million adult patients; 31,675 with congestive heart failure, 78,530 with chronic obstructive pulmonary disease, and 240,378 with diabetes. MEASURES: Care density was assessed in 2008. Prevention Quality Indicators (PQIs), 30-day readmissions, and Healthcare Effectiveness Data and Information Set quality indicators were measured in the following year. RESULTS: Among all patients, we found that patients with the highest care density density--indicating high levels of patient sharing among their office-based physicians--had significantly lower rates of adverse events measured as PQIs compared with patients with low-care density (odds ratio=0.88; 95% confidence interval, 0.85-0.92). A significant association between care density and PQIs was also observed for patients with diabetes mellitus but not congestive heart failure or chronic obstructive pulmonary disease. Diabetic patients with higher care density scores had significantly lower odds of 30-day readmissions (odds ratio=0.68, 95% confidence interval, 0.48-0.97). Significant associations were observed between care density and Healthcare Effectiveness Data and Information Set measures although not always in the expected direction. CONCLUSION: In some settings, patients whose doctors share more patients had lower odds of adverse events and 30-day readmissions.

The wired patient: patterns of electronic patient portal use among patients with cardiac disease or diabetes.

BACKGROUND: As providers develop an electronic health record-based infrastructure, patients are increasingly using Web portals to access their health information and participate electronically in the health care process. Little is known about how such portals are actually used. OBJECTIVE: In this paper, our goal was to describe the types and patterns of portal users in an integrated delivery system. METHODS: We analyzed 12 months of data from Web server log files on 2282 patients using a Web-based portal to their electronic health record (EHR). We obtained data for patients with cardiovascular disease and/or diabetes who had a Geisinger Clinic primary care provider and were registered "MyGeisinger" Web portal users. Hierarchical cluster analysis was applied to longitudinal data to profile users based on their frequency, intensity, and consistency of use. User types were characterized by basic demographic data from the EHR. RESULTS: We identified eight distinct portal user groups. The two largest groups (41.98%, 948/2258 and 24.84%, 561/2258) logged into the portal infrequently but had markedly different levels of engagement with their medical record. Other distinct groups were characterized by tracking biometric measures (10.54%, 238/2258), sending electronic messages to their provider (9.25%, 209/2258), preparing for an office visit (5.98%, 135/2258), and tracking laboratory results (4.16%, 94/2258). CONCLUSIONS: There are naturally occurring groups of EHR Web portal users within a population of adult primary care patients with chronic conditions. More than half of the patient cohort exhibited distinct patterns of portal use linked to key features. These patterns of portal access and interaction provide insight into opportunities for electronic patient engagement strategies.

Development of a hospital coding intensity measure based on sepsis diagnoses.

Significant variation in coding intensity among hospitals has been observed and can lead to reimbursement inequities and inadequate risk adjustment for quality measures. Reliable tools to quantify hospital coding intensity are needed. We hypothesized that coded sepsis rates among patients hospitalized with common infections may serve as a useful surrogate for coding intensity and derived a hospital-level sepsis coding intensity measure using prevalence of "sepsis" primary diagnoses among patients hospitalized with urinary tract infection, cellulitis, and pneumonia. This novel measure was well correlated with the hospital mean number of discharge diagnoses, which has historically been used to quantify hospital-level coding intensity. However, it has the advantage of inferring hospital coding intensity without the strong association with comorbidity that the mean number of discharge diagnoses has. Our measure may serve as a useful tool to compare coding intensity across institutions.