Reliability and correlations among quality measures for lung, breast, and colorectal cancer

2073

Background: Alternative payment models for oncology seek to improve quality and reduce spending. Yet the ability to measure high-quality care across oncology practices remains uncertain. We characterized quality of care for oncology practices using registry and claims-based measures of processes, utilization, end-of-life care, and survival and assessed correlations of practice-level performance across measure type and cancers. Methods: Using SEER-Medicare data, we studied individuals with newly diagnosed lung (N = 95,635), breast (N = 78,736), or colorectal (CRC, N = 51,385) cancers in 2010-2015 treated in oncology practices with ≥20 patients (502, 492, and 347 practices, respectively). We measured receipt of guideline-recommended treatment and surveillance (processes), hospitalizations or emergency department visits during 6-month chemotherapy episodes (utilization), care intensity in the last month of life (EOL), and 12-month survival (lung and CRC only). We calculated summary process, utilization, and EOL measures for each patient (number of measures met divided by the number for which the patient was eligible). We used hierarchical linear models with practice-level random effects to estimate summary measures and survival for each practice. We calculated practice-level reliability (a measurement’s reproducibility) for each measure based on the between-measure variance, within-measure variance, and sample size. Results: Few practices had ≥20 patients eligible for most measures (38%, 37%, and 31% of practices had ≥20 patients for any lung, breast, and CRC measures, respectively). Measure reliability was low. Only 13%, 7%, and 20% of measures for lung, breast, and CRC, respectively, had a median reliability across practices ≥0.7. Among practices with ≥20 patients with summary measures of each type within cancer, correlations across measure types were low (all correlation coefficients (r)≤0.21 except a weak correlation of the CRC process summary measure with 1-year CRC survival, r = 0.38, p < 0.001). Summary process measures were minimally or not correlated across cancer type (lung, breast, CRC; all correlation coefficients ≤0.16). Conclusions: Claims-based measures of care processes, utilization, EOL care, and survival are limited by small numbers of fee-for-service Medicare patients across practices, even after pooling 6 years of data. Measures have poor reliability and are poorly correlated across measure or cancer type. Additional research is needed to identify reliable quality measures for practice-level alternate payment models.

U.S. trends and racial/ethnic disparities in opioid access among patients with poor prognosis cancer at the end of life (EOL)

7005

Background: Heightened US opioid regulations may limit advanced cancer patients’ access to effective pain management, particularly for racial/ethnic minority and other vulnerable populations. We examined trends in opioid access, disparities in access, and pain-related emergency department (ED) visits among cancer patients near end of life (EOL). Methods: Using a 20% random sample of Medicare FFS beneficiaries, we identified 243,124 patients with poor prognosis cancers who died between 2007-2016. We examined trends in outpatient opioid prescription fills and pain-related ED visits near EOL (30 days prior to death or hospice enrollment), for the overall cohort and by race (white, black, other). Per-capita opioid supply by state was obtained from the federal Drug Enforcement Agency ARCOS database. Geographic fixed-effects models examined predictors of opioid use near EOL, opioid dose in morphine milligram equivalents (MMEs), and pain-related ED visits, adjusted for patient demographic and clinical characteristics, state, opioid supply, and year. Results: From 2007-2016 the proportion of patients with poor prognosis cancers filling an opioid prescription near EOL fell from 41.7% to 35.7%, with greater decrements among blacks (39.3% to 29.8%) than whites (42.2% to 36.5%) and other races (38.2% to 32.4%). The proportion of patients receiving long-acting opioids near EOL fell from 17% to 12% overall (15% to 9% among blacks). Among patients receiving EOL opioids, the median daily dose fell from 40MMEs (IQR 16.5-98.0) to 30MMEs (IQR 15.0–78.8). In adjusted analyses, blacks were less likely than whites to receive EOL opioids (AOR 0.85; 95% CI, 0.80 to 0.91) and on average received 10MMEs less per day (b -9.9; 95% CI -15.7 to -4.2). Patients of other race were also less likely to receive EOL opioids (AOR 0.92; 95% CI, 0.85-0.95), although their dose did not differ significantly from whites. Rates of pain-related ED visits near EOL increased from 13.2% to 18.8% over the study period. In adjusted analyses, blacks were more likely than whites to have pain-related ED visits (AOR 1.29, 95% CI, 1.16-1.37) near death, as were those of other races (AOR 1.30; 95% CI, 1.17-1.37). Conclusions: While lawmakers have sought to mitigate the impact of opioid regulations upon cancer patients, access to EOL opioids have decreased substantially over time with concomitant increases in pain-related ED visits. There are significant racial/ethnic disparities in opioid access, with blacks receiving fewer opioids at lower doses and having more ED-based care for pain near EOL.

Physician and facility drivers of spending variation in locoregional prostate cancer

BACKGROUND

Prostate cancer is the most common male cancer, with a wide range of treatment options. Payment reform to reduce unnecessary spending variation is an important strategy for reducing waste, but its magnitude and drivers within prostate cancer are unknown.

METHODS

In total, 38,971 men aged ≥66 years with localized prostate cancer who were enrolled in Medicare fee-for-service and were included in the Surveillance, Epidemiology, and End Results-Medicare database from 2009 to 2014 were included. Multilevel linear regression with physician and facility random effects was used to examine the contributions of urologists, radiation oncologists, and their affiliated facilities to variation in total patient spending in the year after diagnosis within geographic region. The authors assessed whether spending variation was driven by patient characteristics, disease risk, or treatments. Physicians and facilities were sorted into quintiles of adjusted patient-level spending, and differences between those that were high-spending and low-spending were examined.

RESULTS

Substantial variation in spending was driven by physician and facility factors. Differences in cancer treatment modalities drove more variation across physicians than differences in patient and disease characteristics (72% vs 2% for urologists, 20% vs 18% for radiation oncologists). The highest spending physicians spent 46% more than the lowest and had more imaging tests, inpatient care, and radiotherapy spending. There were no differences across spending quintiles in the use of robotic surgery by urologists or the use of brachytherapy by radiation oncologists.

CONCLUSIONS

Significant differences were observed for patients with similar demographics and disease characteristics. This variation across both physicians and facilities suggests that efforts to reduce unnecessary spending must address decision making at both levels.

Spending variation among ACOs in the Medicare Shared Savings Program

OBJECTIVES

Understanding variation in spending across organizations, rather than across geographic areas, is important because care is delivered by organizations and interventions increasingly focus on organizations. Accountable care organizations (ACOs) are particularly important to study given their incentives to reduce spending. Analyzing spending differences across ACOs may help identify cost savings opportunities.

STUDY DESIGN

Cross-sectional analysis of Medicare claims.

METHODS

We stratified ACOs into quartiles based on the deviation between each ACO's risk-adjusted spending and average risk-adjusted fee-for-service spending in the same market (hospital referral region). We compared spending between top- and bottom-quartile ACOs on each of 7 major service categories and 10 clinical condition groups to identify areas of potential savings. We simulated spending reductions if ACOs with high adjusted spending reduced spending to the levels of lower-spending ACOs.

RESULTS

In 2016, geographically adjusted and risk-adjusted total per-beneficiary spending for the highest-spending quartile of ACOs was 14% higher than for ACOs in the lowest quartile. Variation between high- and low-spending ACOs was greatest, at 27%, in the use of skilled nursing facilities-a service category in which ACOs have reduced spending by the greatest percentage. Inpatient care was the largest driver of absolute dollar differences in spending, however, accounting for 37% of the total spread. If spending in ACOs above median adjusted spending were brought down to the median, savings would be 3% to 4%.

CONCLUSIONS

By extending the variations literature to focus on ACOs, we illustrated that meaningful further savings opportunities exist both within and across markets.

Disparities in outpatient visits for mental health and/or substance use disorders during the COVID surge and partial reopening in Massachusetts

OBJECTIVE

To examine changes in outpatient visits for mental health and/or substance use disorders (MH/SUD) in an integrated healthcare organization during the initial Massachusetts COVID-19 surge and partial state reopening.

METHODS

Observational study of outpatient MH/SUD visits January 1st-June 30th, 2018-2020 by: 1) visit diagnosis group, 2) provider type, 3) patient race/ethnicity, 4) insurance, and 5) visit method (telemedicine vs. in-person).

RESULTS

Each year, January-June 52,907-73,184 patients were seen for a MH/SUD visit. While non-MH/SUD visits declined during the surge relative to 2020 pre-pandemic (-38.2%), MH/SUD visits increased (9.1%)-concentrated in primary care (35.3%) and non-Hispanic Whites (10.5%). During the surge, MH visit volume increased 11.7% while SUD decreased 12.7%. During partial reopening, while MH visits returned to 2020 pre-pandemic levels, SUD visits declined 31.1%; MH/SUD visits decreased by Hispanics (-33.0%) and non-Hispanic Blacks (-24.6%), and among Medicaid (-19.4%) and Medicare enrollees (-20.9%). Telemedicine accounted for ~5% of MH/SUD visits pre-pandemic and 83.3%-83.5% since the surge.

CONCLUSIONS

MH/SUD visit volume increased during the COVID surge and was supported by rapidly-scaled telemedicine. Despite this, widening diagnostic and racial/ethnic disparities in MH/SUD visit volume during the surge and reopening suggest additional barriers for these vulnerable populations, and warrant continued monitoring and research.

Multimodality cancer care and implications for episode-based payments in cancer

Most patients receiving multimodality cancer care receive care from different practices. Therefore, episode-based payments in oncology must hold multiple providers accountable for costs and quality.

Early Findings From the Oncology Care Model Evaluation

PURPOSE:

The Oncology Care Model (OCM) is an alternative payment model administered by the Centers for Medicare & Medicaid Services (CMS) that is structured around 6-month chemotherapy treatment episodes. This report describes the CMS-sponsored OCM evaluation and summarizes early evaluation findings.

METHODS:

The OCM evaluation examines health care spending and use, quality of care, and patient experience during chemotherapy treatment episodes. Because OCM participation is voluntary, the evaluation compares participating physician practices with a propensity-matched group of nonparticipating practices by using a difference-in-differences approach. This report examines 6-month episodes initiated during the first OCM performance period (July 1, 2016, through January 1, 2017).

RESULTS:

During the first OCM performance period, there was no statistically significant impact of OCM on total episode payments. There were small declines in intensive care unit (ICU) admissions (7 per 1,000 episodes) and emergency department visits (15 per 1,000 episodes); there was no statistically significant impact on hospitalizations or 30-day readmissions. Analyses of care quality and end-of-life care showed statistically significant impacts of OCM on the proportion of patients with inpatient hospitalizations in the last 30 days of life (1.5% absolute decrease) and ICU admissions in the last 30 days of life (2.1% decrease). There was no significant OCM impact on measures of hospice use.

CONCLUSION:

Early findings from the OCM evaluation demonstrate modest program-related impacts on some acute care services and no change in total episode payments. Early findings may not reflect practice redesign efforts that were phased in after the beginning of OCM.

Association Between Clinical Practice Group Adherence to Quality Measures and Adverse Outcomes Among Adult Patients With Diabetes

IMPORTANCE

Clinical practice group performance on quality measures associated with chronic disease management has become central to reimbursement. Therefore, it is important to determine whether commonly used process and disease control measures for chronic conditions correlate with utilization-based outcomes, as they do in acute disease.

OBJECTIVE

To examine the associations among clinical practice group performance on diabetes quality measures, including process measures, disease control measures, and utilization-based outcomes.

DESIGN, SETTING, AND PARTICIPANTS

This retrospective, cross-sectional analysis examined commercial claims data from a national health insurance plan. A cohort of eligible beneficiaries with diabetes aged 18 to 65 years who were enrolled for at least 12 months from January 1, 2010, through December 31, 2014, was defined. Eligible beneficiaries were attributed to a clinical practice group based on the plurality of their primary care or endocrinology office visits. Data were analyzed from October 1, 2018, through April 30, 2019.

MAIN OUTCOMES AND MEASURES

For each clinical practice group, performance on current diabetes quality measures included 3 process measures (2 testing measures [hemoglobin A1c {HbA1c} and low-density lipoprotein {LDL} testing] and 1 drug use measure [statin use]) and 2 disease control measures (HbA1c <8% and LDL level <100 mg/dL). The rates of utilization-based outcomes, including hospitalization for diabetes and major adverse cardiovascular events (MACEs), were also measured.

RESULTS

In this cohort of 652 258 beneficiaries with diabetes from 886 clinical practice groups, 42.9% were aged 51 to 60 years, and 52.6% were men. Beneficiaries lived in areas that were predominantly white (68.1%). At the clinical practice group level, except for high correlation between the 2 testing measures, correlations among different quality measures were weak (r range, 0.010-0.244). Rate of HbA1c of less than 8% had the strongest correlation with hospitalization for MACE (r = -0.046; P = .03) and diabetes (r = -0.109; P < .001). Rates of HbA1c control at the clinical practice group level were not significantly associated with likelihood of hospitalization at the individual level. Performance on the process and disease control measures together explained 3.9% of the variation in the likelihood of hospitalization for a MACE or diabetes at the individual level.

CONCLUSIONS AND RELEVANCE

In this study, performance on utilization-based measures-intended to reflect the quality of chronic disease management-was only weakly associated with direct measures of chronic disease management, namely, disease control measures. This correlation should be considered when determining the degree of financial emphasis to place on hospitalization rates as a measure of quality in treatment of chronic diseases.

Factors Contributing To Geographic Variation In End-Of-Life Expenditures For Cancer Patients

Health care spending in the months before death varies across geographic areas but is not associated with outcomes. Using data from the prospective multiregional Cancer Care Outcomes Research and Surveillance Consortium (CanCORS) study, we assessed the extent to which such variation is explained by differences in patients' sociodemographic factors, clinical factors, and beliefs; physicians' beliefs; and the availability of services. Among 1,132 patients ages sixty-five and older who were diagnosed with lung or colorectal cancer in 2003-05, had advanced-stage cancer, died before 2013, and were enrolled in fee-for-service Medicare, mean expenditures in the last month of life were $13,663. Physicians in higher-spending areas reported less knowledge about and comfort with treating dying patients and less positive attitudes about hospice, compared to those in lower-spending areas. Higher-spending areas also had more physicians and fewer primary care providers and hospices in proportion to their total population than lower-spending areas did. Availability of services and physicians' beliefs, but not patients' beliefs, were important in explaining geographic variations in end-of-life spending. Enhanced training to better equip physicians to care for patients at the end of life and strategic resource allocation may have potential for decreasing unwarranted variation in care.

Classifying Stage IV Lung Cancer From Health Care Claims: A Comparison of Multiple Analytic Approaches

PURPOSE

Cancer stage is a key determinant of outcomes; however, stage is not available in claims-based data sources used for real-world evaluations. We compare multiple methods for classifying lung cancer stage from claims data.

METHODS

Our study used the linked SEER-Medicare data. The patient samples included fee-for-service Medicare beneficiaries diagnosed with lung cancer from 2010 to 2011 (development cohort) and 2012 to 2013 (validation cohort) who received chemotherapy. Classification algorithms considered Medicare Part A and B claims for care in the 3 months before and after chemotherapy initiation. We developed a clinical algorithm to predict stage IV (v I to III) cancer on the basis of treatment patterns (surgery, radiotherapy, chemotherapy). We also considered an ensemble of claims-based machine learning algorithms. Classification methods were trained in the development cohort, and performance was measured in both cohorts. The SEER data were the gold standard for cancer stage.

RESULTS

Development and validation cohorts included 14,760 and 14,620 patients with lung cancer, respectively. Validation analyses assessed clinical, random forest, and simple logistic regression algorithms. The best performing classifier within the development cohort was the random forests, but this performance was not replicated in validation analysis. Logistic regression had stable performance across cohorts. Compared with the clinical algorithm, the 14-variable logistic regression algorithm demonstrated higher accuracy in both the development (77% v 71%) and validation cohorts (77% v 73%), with improved specificity for stage IV disease.

CONCLUSION

Machine learning algorithms have potential to improve lung cancer stage classification but may be prone to overfitting. Use of ensembles, cross-validation, and external validation can aid generalizability. Degradation of accuracy between development and validation cohorts suggests the need for caution in implementing machine learning in research or care delivery.

The Comparative Advantage of Medicare Advantage

We ascertain the degree of service-level selection in Medicare Advantage (MA) using individual level data on the 100 most frequent HCC's or combination of HCC's from two national insurers in 2012-2013. We find differences in the distribution of beneficiaries across HCC's between TM and MA, principally in the smaller share of MA enrollees with no coded HCC, consistent with greater coding intensity in MA. Among those with an HCC code, absolute differences between MA and TM shares of beneficiaries are small, consistent with little service-level selection. Variation in HCC margins does not predict differences between an HCC's share of MA and TM enrollees, although one cannot a priori sign a relationship between margin and service-level selection. Margins are negatively associated with the importance of post-acute care in the HCC. Margins among common chronic disease classes amenable to medical management and typically managed by primary care physicians are larger than among diseases typically managed by specialists. These margin differences by disease are robust against a test for coding effects and suggest that the average technical efficiency of MA relative to TM may vary by diagnosis. If so, service-level selection on the basis of relative technical efficiency could be welfare enhancing.

Abstract 224: Recent Trends in Coronary Artery Disease Quality Performance and Implications for Clinical Practice in the Era of Alternative Payment Models

Background: In light of recent shifts away from fee-for-service and toward alternative payment models (APM), national trends in quality performance for common cardiac conditions, such as CAD, are important for identifying areas for quality improvement and also for determining physician/health system reimbursement.

Methods: Using Medicare data from 2010-2013, we created a cohort of patients with CAD using a combination of chronic condition warehouse (CCW) flags, ICD-9 and CPT codes. We the determined national performance trends for the 2011 ACC/AHA CAD performance measures. For drug use metrics, we used 80% of days covered after the index event as the threshold.

Results: From 2010-2013, performance trends for testing (annual LDL) and post-MI metrics (beta blocker use, P2Y12 use and cardiac rehab) were flat (p=ns). Among patients with CAD and another comorbidity such as heart failure or diabetes, neurohormonal therapy use increased (p<0.001, Figure 1 ).

Conclusion: The rate of neurohormonal therapy use in patients with CAD and another comorbidity improved while testing and post-MI performance in patients with CAD alone changed little. The reasons for this and may relate to an increased emphasis on complex, costly patients in APMs. Whether these trends impact longer-term patient outcomes should be explored.

Social Risk Adjustment of Quality Measures for Diabetes and Cardiovascular Disease in a Commercially Insured US Population

IMPORTANCE

Patients' social risk factors may be associated with physician group performance on quality measures.

OBJECTIVE

To examine the association of social risk with change in physician group performance on diabetes and cardiovascular disease (CVD) quality measures in a commercially insured population.

DESIGN, SETTING, AND PARTICIPANTS

In this cross-sectional study using claims data from 2010 to 2014 from a US national health insurance plan, the performance of 1400 physician groups (physicians billing under the same tax identification number) was estimated. After base adjustments for age and sex, changes in variation across groups and reordering of rankings resulting from additional adjustments for clinical, social, or both clinical and social risk factors were analyzed. In all models, only within-group associations were adjusted to distinguish the association of patients' social risk factors with outcomes while excluding physician groups' distinct characteristics that could also change observed performance. Data analysis was conducted between April and July 2018.

MAIN OUTCOMES AND MEASURES

Process measures (hemoglobin A1c [HbA1c] testing, low-density lipoprotein cholesterol [LDL-C] testing, and statin use), disease control measures (HbA1c and LDL-C level control), and use-based outcome measures (hospitalizations for ambulatory-sensitive conditions) were calculated with base adjustment (age and sex), clinical adjustment, social risk factor adjustment, and both clinical and social adjustments. Quality variance in physician group performance and changes in rankings following these adjustments were measured.

RESULTS

This study identified 1 684 167 enrollees (859 618 [51%] men) aged 18 to 65 years (mean [SD] age, 50 [10.7] years) with diabetes or CVD. Performance rates were high for HbA1c and LDL-C level testing (mean ranged from 79.5% to 87.2%) but lower for statin use (54.7% for diabetes cohort and 44.2% for CVD cohort) and disease control measures (57.9% on LDL-C control for diabetes cohort and 40.0% for CVD cohort). On average, only 8.8% of enrollees with diabetes and 1.0% of enrollees with CVD in a group were hospitalized. The addition of clinical and social risk factors to base adjustment reduced variance across physician groups for most measures (percentage change in SD ranged from -13.9% to 1.6%). Although overall agreement between performance scores with base vs full adjustment was high, there was still substantial reordering for some measures. For example, social risk adjustment resulted in reordering for disease control in the diabetes cohort. Of the 1400 physician groups, 330 (23.6%) had performance rankings for HbA1c control that increased or decreased by at least 10 percentile points after adding social risk factors to age and sex. Both clinical and social risk adjustment affected rankings on hospital admissions.

CONCLUSIONS AND RELEVANCE

Accounting for social risk may be important to mitigate adverse consequences of performance-based payments for physician groups serving socially vulnerable populations.

What drives variation in spending for breast cancer patients within geographic regions?

OBJECTIVE

To estimate and describe factors driving variation in spending for breast cancer patients within geographic region.

DATA SOURCE

Surveillance, Epidemiology, and End Results (SEER)-Medicare database from 2009-2013.

STUDY DESIGN

The proportion of variation in monthly medical spending within geographic region attributed to patient and physician factors was estimated using multilevel regression models with individual patient and physician random effects. Using sequential models, we estimated the contribution of differences in patient and disease characteristics or use of cancer treatment modalities to patient-level and physician-level variance in spending. Services associated with high spending physicians were estimated using linear regression.

DATA EXTRACTION METHOD

A total of 20 818 women with a breast cancer diagnosis in 2010-2011.

PRINCIPAL FINDINGS

We observed substantial between-patient and between-provider variation in spending following diagnosis and at the end-of-life. Immediately following diagnosis, 48% of between-patient and 31% of between-physician variation were driven by differences in delivery of cancer treatment modalities to similar patients. At the end-of-life, patients of high spending physicians had twice as many inpatient days, double the chemotherapy spending, and slightly more hospice days.

CONCLUSIONS

Similar patients receive very different treatments, which yield significant differences in spending. Efforts to reduce unwanted variation may need to target treatment choices within patient-doctor discussions.

Comparison of Approaches for Aggregating Quality Measures in Population-based Payment Models

OBJECTIVE

To assess the impact of alternative methods of aggregating individual quality measures on Accountable Care Organization (ACO) overall scores.

DATA SOURCE

2014 quality scores for Medicare ACOs.

STUDY DESIGN

We compare ACO overall scores derived using CMS' aggregation approach to those derived using alternative approaches to grouping and weighting measures.

PRINCIPAL FINDINGS

Alternative grouping and weighting methods based on statistical criteria produced overall quality scores similar to those produced using CMS' approach (κ = 0.80 to 0.95). Scores derived from giving specific domains greater weight were less similar (κ = 0.51 to 0.93).

CONCLUSIONS

How measures are grouped into domains and how these domains are weighted to generate overall scores can have important implications for ACO's shared savings payments.

Synthesis Of Research On Patient-Centered Medical Homes Brings Systematic Differences Into Relief

The patient-centered medical home (PCMH) model emphasizes comprehensive, coordinated, patient-centered care, with the goals of reducing spending and improving quality. To evaluate the impact of PCMH initiatives on utilization, cost, and quality, we conducted a meta-analysis of methodologically standardized findings from evaluations of eleven major PCMH initiatives. There was significant heterogeneity across individual evaluations in many outcomes. Across evaluations, PCMH initiatives were not associated with changes in the majority of outcomes studied, including primary care, emergency department, and inpatient visits and four quality measures. The initiatives were associated with a 1.5 percent reduction in the use of specialty visits and a 1.2 percent increase in cervical cancer screening among all patients, and a 4.2 percent reduction in total spending (excluding pharmacy spending) and a 1.4 percent increase in breast cancer screening among higher-morbidity patients. These associations were significant. Identification of the components of PCMHs likely to improve outcomes is critical to decisions about investing resources in primary care.

Abstract 35: Trends in Statin Use and Adherence and the Impact of the 2013 Cholesterol Guidelines

Background: In 2013, the ACC/AHA updated the cholesterol treatment guidelines. At the time, it was estimated that an additional 13 million Americans would quality for statin therapy. To date, the real-world implications of this guideline change have not been well studied. This study aims to better understand trends in statin use and adherence, by gender, and the impact of guideline change.

Methods: This is a retrospective, observational study using medical and pharmacy claims from 2009 to 2014 from a large, national, commercial insurer. Considering all beneficiaries aged 18-65 with ≥1 year of continuous enrollment, we created annual cross sectional populations of statin-eligible patients and divided them into 3 statin benefit groups (SBG). In descending order of risk, the groups were: (1) atherosclerotic cardiovascular disease (ASCVD); (2) diabetes and (3) hyperlipidemia. Patients were assigned to the highest risk group that they qualified for.

Results: Statin use rates among those with ASCVD increased until 2012 and then plateaued (Figure 1a ). Use rates among those with diabetes, were flat until 2011 and then increased. Use rates among those with hyperlipidemia steadily rose from 2009-2014. Statin adherence rates among those with ASCVD increased from 2009-2014 ( Figure 1b ). Adherence rates among those with diabetes, decreased from 2009-2011 and then rose significantly from 2011-2014. Adherence rates among those with hyperlipidemia also rose steadily from 2009-2014. The most significant gender gap in treatment, for both use and adherence, was between men and women with ASCVD. There was with little change in this treatment gap, in any risk group, over the time period observed.

Conclusion: The 2013 cholesterol guidelines have not yet had a significant effect on statin use or adherence. Recently improving trends in statin use and adherence, especially among patients with diabetes, appear to predate the 2013 guideline change. A significant gender gap in statin treatment remains, especially among those in the highest risk group.

Classifying Lung Cancer Severity with Ensemble Machine Learning in Health Care Claims Data

Research in oncology quality of care and health outcomes has been limited by the difficulty of identifying cancer stage in health care claims data. Using linked cancer registry and Medicare claims data, we develop a tool for classifying lung cancer patients receiving chemotherapy into early vs. late stage cancer by (i) deploying ensemble machine learning for prediction, (ii) establishing a set of classification rules for the predicted probabilities, and (iii) considering an augmented set of administrative claims data. We find our ensemble machine learning algorithm with a classification rule defined by the median substantially outperforms an existing clinical decision tree for this problem, yielding full sample performance of 93% sensitivity, 92% specificity, and 93% accuracy. This work has the potential for broad applicability as provider organizations, payers, and policy makers seek to measure quality and outcomes of cancer care and improve on risk adjustment methods.

An administrative stage inference algorithm for use in patients receiving chemotherapy for colorectal cancer

e18121

Background: Claims data are often used for evaluating cancer care outcomes; however, such data lack information about cancer stage. We developed a claims-based stage inference algorithm to classify chemotherapy-treated patients with localized vs. metastatic colorectal cancer (CRC). Methods: We used the SEER-Medicare linked data (2010-‘11) to develop and validate an algorithm to predict cancer stage (localized vs. metastatic) among patients receiving chemotherapy within 6 months of CRC diagnosis. We used claims to identify treatments received (surgery, radiation, and chemotherapy agents) during the 6 months before and after the first dose of chemotherapy. The sample was split 1:1 into development and validation cohorts. After testing in the development cohort, the final algorithm was evaluated in the validation cohort. SEER data served as the gold standard for cancer stage. Results: We identified 25,258 patients with fee-for-service Medicare and a new diagnosis of CRC. 6,907 patients (27%) received chemotherapy for CRC within 6 months of diagnosis. The median age of chemotherapy-treated patients was 73, 49% were female, and 76% were white; 69% had localized cancer (AJCC stage 1-3) and 31% had metastasis at diagnosis (stage 4). Split-sample validation of the final classification algorithm demonstrated sensitivity and specificity of 87% (95% CI 86-89%) and 76% (73-78%) for localized cancer and 73% (70-75%) and 91% (90-92%) for metastatic cancer. The overall accuracy of classification was 83%. Stratified analyses demonstrated preserved algorithm performance across subgroups of age, sex, race, geography, and comorbidity. Misclassification was most common among patients with metastatic disease who were treated with surgery followed by fluoropyrimidine chemotherapy with or without oxaliplatin. 2-year overall survival was 79.8% (stage 1-3) and 35.4% (stage 4) for SEER stage groups, vs. 79.6% and 35.8% for predicted stage groups. Conclusions: A claims-based algorithm can classify extent of disease in chemotherapy-treated CRC patients with an accuracy of 83%. This algorithm will allow more clinically-relevant patient stratification for claims-based evaluations of cancer care outcomes.