Advertisement

Sociodemographic Determinants of Oral Anticoagulant Prescription in Patients with Atrial Fibrillations: Findings from the PINNACLE Registry Using Machine Learning

Open AccessPublished:November 23, 2022DOI:https://doi.org/10.1016/j.hroo.2022.11.004

      ABSTRACT

      Background

      Current risk scores that are solely based on clinical factors have shown modest predictive ability for understanding of factors associated with gaps in real-world prescription of oral anticoagulants (OAC) in patients with atrial fibrillation (AF).

      Objective

      In this study we sought to identify the role of social and geographic determinants, beyond clinical factors associated with variation in OAC prescriptions using a large national registry of ambulatory patients with AF.

      Methods

      Between January 2017 and June 2018, we identified patients with AF from the American College of Cardiology’s (ACC) Practice Innovation and Clinical Excellence (PINNACLE) Registry. We examined associations between patient and site-of-care factors and prescription of OAC across U.S. counties. Several ML methods were used to identify factors associated with OAC prescription.

      Results

      Amongst 864,339 patients with AF, 586,560 (68%) were prescribed OAC. County OAC prescription rates ranged from 26.8% to 93%, with higher OAC use in the Western US. Supervised ML analysis in predicting likelihood of OAC prescriptions and identified a rank order of patient features associated with OAC prescription. In the ML models, in addition to clinical factors, medication use (aspirin, antihypertensives, antiarrhythmic agents, lipid modifying agents), and age; household income, clinic size and US regions were amongst the most important predictors of OAC prescription.

      Conclusions

      In a contemporary, national cohort of patients with AF underuse of OAC remains high, with notable geographic variation. Our results demonstrated the role of several important demographic and socioeconomic factors in underutilization of OAC in patients with atrial fibrillation.

      KEY WORDS

      ABBREVIATIONS:

      ACC (American College of Cardiology), AF (atrial fibrillation), AUC (area under the receiver operating characteristic curve), CI (confidence interval), DOAC (direct oral anticoagulant), GFR (glomerular filtration rate), INR (international normalized ratio), ML (machine learning), OAC (oral anticoagulation), PINNACLE (Practice Innovation and Clinical Excellence)

      INTRODUCTION

      Oral anticoagulation (OAC) reduces the risk of stroke and systemic embolism in patients with atrial fibrillation (AF). Yet, use of OAC in patients with AF has historically been suboptimal [1,2]. Previous analyses involving the American College of Cardiology’s (ACC) Practice Innovation and Clinical Excellence (PINNACLE) Registry from 2008 to 2014 have documented OAC prescription rates ranging between 45% and 61%
      • Hsu J.C.
      • et al.
      Oral Anticoagulant Therapy Prescription in Patients With Atrial Fibrillation Across the Spectrum of Stroke Risk: Insights From the NCDR PINNACLE Registry.
      ,
      • Marzec L.N.
      • et al.
      Influence of Direct Oral Anticoagulants on Rates of Oral Anticoagulation for Atrial Fibrillation.
      . Factors contributing to the observed care gaps are numerous and include those at the patient, clinician, and health system level.
      The CHA2DS2-VASc score (Congestive heart failure, Hypertension, Age ≥75 years [doubled], Diabetes, Stroke [doubled], Vascular disease, Age 65 to 74 years, and Sex [female]) has been extensively validated to estimate the risk for the development of stroke based on specific demographic and clinical risk factors and is used to inform treatment decisions about OAC
      • Lip G.Y.
      • Nieuwlaat R.
      • Pisters R.
      • Lane D.A.
      • Crijns H.J.
      Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the euro heart survey on atrial fibrillation.
      . This clinical risk score along with other similar scores have shown modest predicting ability to predict outcomes. One explanation of the modest predictive performance of these scores may be that these scores only include clinical factors and do not consider socio-demographic and geographical variations that are known to be important predictors of cardiovascular outcomes and anticoagulation use
      • Havranek E.P.
      • et al.
      Social Determinants of Risk and Outcomes for Cardiovascular Disease.
      ,
      • Hernandez I.
      • Saba S.
      • Zhang Y.
      Geographic Variation in the Use of Oral Anticoagulation Therapy in Stroke Prevention in Atrial Fibrillation.
      . The 2019 American Heart Association/ACC/Heart Rhythm Society guidelines recommend OAC for all AF patients based upon a qualifying CHA2DS2-VASc score
      • January C.T.
      • et al.
      2019 AHA/ACC/HRS Focused Update of the 2014 AHA/ACC/HRS Guideline for the Management of Patients With Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society.
      . Preference is given to direct oral anticoagulants (DOACs) in most patients with AF
      • January C.T.
      • et al.
      2019 AHA/ACC/HRS Focused Update of the 2014 AHA/ACC/HRS Guideline for the Management of Patients With Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society.
      • Yao X.
      • et al.
      Effect of Adherence to Oral Anticoagulants on Risk of Stroke and Major Bleeding Among Patients With Atrial Fibrillation.
      • Jackevicius C.A.
      • et al.
      Early non-persistence with dabigatran and rivaroxaban in patients with atrial fibrillation.
      due to their ease of administration and therapeutic advantages compared to vitamin K antagonists
      • Patel M.R.
      • et al.
      Rivaroxaban versus warfarin in nonvalvular atrial fibrillation.
      • Granger C.B.
      • et al.
      Apixaban versus warfarin in patients with atrial fibrillation.
      • Connolly S.J.
      • et al.
      Dabigatran versus warfarin in patients with atrial fibrillation.
      • Giugliano R.P.
      • et al.
      Edoxaban versus warfarin in patients with atrial fibrillation.
      • January C.T.
      • et al.
      2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: executive summary: a report of the American College of Cardiology/American Heart Association Task Force on practice guidelines and the Heart Rhythm Society.
      . The current OAC practice patterns for AF patients remain incompletely characterized, with substantial opportunity to better understand geographic, clinical, and socioeconomic determinants of guideline-directed OAC use.
      Machine learning (ML) is a branch of artificial intelligence that leverages data analysis to identify relationships between variables directly from the data. ML encompasses supervised (e.g. predicting an outcome) and unsupervised (e.g. clustering) methods, which can be used to process complex, high-volume datasets. Such techniques may complement traditional statistical approaches by identifying non-intuitive features or combined patient and site-of-care variables (signatures) to gain insight into patterns and predictors of OAC prescription among patients with AF.
      Using the PINNACLE Registry, we sought to describe role of social determinants of health and geographic differences in contemporary OAC prescription practices among patients with AF. We further leveraged a clinical intelligence data platform using ML algorithms to identify predictors of OAC care gaps.

      METHODS

      Data Source

      We analyzed data from the PINNACLE Registry, which includes 829 practices throughout the US. Details related to this registry, along with available data elements (e.g., patient demographics, comorbidities, vital signs, medications, laboratory values, and recent hospitalizations) have been previously described
      • Hsu J.C.
      • et al.
      Oral Anticoagulant Therapy Prescription in Patients With Atrial Fibrillation Across the Spectrum of Stroke Risk: Insights From the NCDR PINNACLE Registry.
      ,
      • Chan P.S.
      • Maddox T.M.
      • Tang F.
      • Spinler S.
      • Spertus J.A.
      Practice-level variation in warfarin use among outpatients with atrial fibrillation (from the NCDR PINNACLE program).
      ,
      • Messenger J.C.
      • et al.
      The National Cardiovascular Data Registry (NCDR) Data Quality Brief: the NCDR Data Quality Program in 2012.
      . Waiver of written informed consent and authorization for this study was granted by Chesapeake Research Review Incorporated due to the use of de-identified, retrospective data.

      Clinical Intelligence Engine

      For this study, we leveraged CLINTTM, an analytic engine from the ACC’s innovation collaborator, HealthPals Incorporated (Millbrae, CA). CLINTTM has a comprehensive array of codified ACC cardiometabolic guidelines that map best practices to individual patient data, allowing for efficient identification of care gaps. Available fields in the PINNACLE Registry (now operated by Veradigm and comprising ∼360 structured data elements per patient encounter) were integrated into CLINTTM and used to derive a CHA2DS2-VASc score for each patient. The CHA2DS2-VASc score was used to identify care gaps, defined as the percentage of patients who have a Class I indication for OAC according to the 2019 American Heart Association/ACC/Heart Rhythm Society Atrial Fibrillation guideline update
      • January C.T.
      • et al.
      2019 AHA/ACC/HRS Focused Update of the 2014 AHA/ACC/HRS Guideline for the Management of Patients With Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society.
      . These evidence-based care gaps were then aggregated into an interactive population dashboard. We also used CLINTTM to efficiently perform cohort selection and train and evaluate ML models on the PINNACLE data.

      Study Population

      The study population consisted of patients enrolled in the PINNACLE Registry from January 2017 to June 2018. Eligible patients included those with a diagnosis of non-valvular (or ‘unspecified’) AF at any encounter within the 18-month survey period and a recorded gender, which is necessary for calculating the CHA2DS2-VASc score.

      Outcome

      The primary outcome was OAC prescription, defined as the presence of at least one anticoagulant (apixaban, dabigatran, edoxaban, rivaroxaban or warfarin) in the most recent three months of each patient’s record. Prescription rates were calculated by dividing the number of patients prescribed OAC by the total number of OAC-eligible patients with AF. To identify geographic gaps in guideline adherence, we grouped patients with AF into counties based on their clinic’s street address. Prescription rates were then calculated for each US county. To ensure sufficient data quality, rates were only calculated for counties with at least 40 patients in the study.

      Patient Characteristics

      Variables were extracted from the PINNACLE Registry. For each patient, information was collected from the quarter (three months) during which the patient’s most recent outpatient encounter occurred. In addition to demographic variables, variables representing pre-selected cardiometabolic comorbidities (dyslipidemia, chronic kidney disease, chronic liver disease, thyroid disease, hemodialysis, prior kidney transplant, sleep apnea, stable and unstable angina) and medications prescribed (antihypertensives, antiarrhythmic agents, lipid modifying therapies, aspirin, antiplatelets other than aspirin [clopidogrel, prasugrel, vorapaxar, ticagrelor], and blood glucose regulation agents) were created by searching for any mention of each within the three-month period. Laboratory variables (international normalized ratio [INR], glomerular filtration rate [GFR], lipid levels) were also collected; if a patient had multiple values from the same lab test in the three-month period, the values were averaged. Vital signs (heart rate, blood pressure, weight) and insurance information (commercial, Medicaid, Medicare, or other) were included. Clinic information included US census region (South/Midwest/Northeast/West), urbanicity (urban/suburban/rural), and number of patients seen by the clinic. Mean household income was determined using data from the 2016 US Census and the ZIP code in which the patient’s clinic was located. The CHA2DS2-VASc score was calculated based on PINNACLE data fields as previously described
      • Thompson L.E.
      • et al.
      Sex Differences in the Use of Oral Anticoagulants for Atrial Fibrillation: A Report From the National Cardiovascular Data Registry (NCDR(®)) PINNACLE Registry.
      . Variables were used at the patient level to determine associations and develop models and were aggregated at the county level to explore geographic trends.

      Machine Learning Analysis

      Machine learning analyses were completed in Python 3.6 using the Scikit-learn package, version 0.21.2. In order to identify the most important drivers of guideline-adherent OAC prescriptions, several ML binary classifiers, including logistic regression, LASSO-penalized logistic regression, random forests, and extreme gradient boosting (XGBoost)
      • Breiman L.
      Random Forests.
      ,

      Chen, T. & Guestrin, C. Xgboost: A scalable tree boosting system. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining 785-794 (2016).

      , were trained on variables derived from the PINNACLE dataset. These classifiers were chosen for their ability to effectively incorporate many variables into the models. The tree-based ML classifiers [random forests and XGBoost] aggregate the predictions of many independent decision trees. This allows these models to capture complex variable interactions while simultaneously minimizing variance.
      As a baseline comparison, we analyzed the ability of the CHA2DS2-VASc score to directly predict which patients would receive an OAC prescription
      • Hsu J.C.
      • et al.
      Oral Anticoagulant Therapy Prescription in Patients With Atrial Fibrillation Across the Spectrum of Stroke Risk: Insights From the NCDR PINNACLE Registry.
      . To compare directly with the CHA2DS2-VASc score, we trained the four ML models to predict OAC prescription using only the variables comprising the CHA2DS2-VASc score: gender, age, heart failure, hypertension, diabetes, peripheral artery disease, peripheral vascular disease, prior myocardial infarction, coronary artery bypass graft surgery, percutaneous coronary intervention, ischemic stroke, and transient ischemic attack.
      As ML models (particularly tree-based models) are able to effectively use a large number of variables which can be correlated with each other, we next included a wide range of clinical, demographic, and geographic variables in “enhanced” variants of the ML models. In addition to the CHA2DS2-VASc variables, clinic information (region, urbanicity, number of patients), demographic factors (race/ethnicity, mean household income), medical comorbidities, medications prescribed, vital signs, laboratory data, and insurance information were used to train and evaluate the enhanced ML classifiers. For variables with continuous values, missing fields were imputed using the median value of that variable across the dataset.
      Data were split into training (80%) and testing (20%) sets. Within the training set, 5-fold cross validation was used to tune hyperparameters, such as regularization parameters, maximum tree depth, and number of trees. Hyperparameters were tuned to control the models’ complexities and prevent overfitting. Models were compared based on the area under the receiver operating characteristic curve (AUC), also known as the C-statistic. Once the best hyperparameters were selected for each ML classifier, models with these hyperparameters were retrained on the entire training set. The final AUCs, model accuracy, precision, recall, and area under the precision-recall curve (PRAUC) of both regular and “enhanced” ML models, as well as the CHA2DS2-VASc score, were reported on the testing set.
      We also assessed the feature importance of the model with the highest testing-set AUC in order to understand how much weight the model places on each of the expanded set of covariates in determining OAC prescription probability. This analysis was only performed on the testing set. Traditionally, feature importance for random forests is reported using the decrease in the Gini impurity in the training process, but this has a number of shortcomings
      • Strobl C.
      • Boulesteix A.L.
      • Zeileis A.
      • Hothorn T.
      Bias in random forest variable importance measures: illustrations, sources and a solution.
      , particularly in the presence of correlated variables and variables of mixed types (binary/categorical/continuous). As such, we ranked variables using permutation importance
      • Altmann A.
      • Toloşi L.
      • Sander O.
      • Lengauer T.
      Permutation importance: a corrected feature importance measure.
      , a method which randomly permutes values in columns of the test data and measures decrease in performance. Notably, while the permutation importance represents the overall magnitude of influence for each feature on OAC prescription rate, the polarity of influence of any given feature (e.g. CHA2DS2-VASc score) may be a combination of positive and negative statistical associations that may depend upon the numerical value of the feature itself or other variable inputs. To offer insight into the polarity of variable influence, we plot OAC rates against the most important variables (Supplementary Figure 1).

      Data Availability

      The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information files.
      The linked data used in this analysis were deidentified, so the study was exempt from the requirement for review board approval and informed consent. The research reported in this paper adhered to Helsinki Declaration guidelines for human research.

      RESULTS

      Descriptive Patterns in OAC Prescription

      Between January 1, 2017 and June 30, 2018, there were 864,339 patients with AF in the registry (Figure 1). Table 1 shows patient-level characteristics for the study population by OAC use. A total of 586,554 (68%) received OAC, of which, 69% (n=401,953) were prescribed a DOAC. Most of the AF patients (85%, n=734,288) met contemporary Class I indications for an OAC prescription; approximately 70% of these patients (n=520,909) were prescribed OAC.
      Figure thumbnail gr1
      Figure 1Title: Cohort diagram. Caption: Flowchart detailing the inclusion and exclusion criteria used for identifying this study population from the PINNACLE dataset. Abbreviations: PINNACLE: Practice Innovation and Clinical Excellence; AF: atrial fibrillation
      Table 1Characteristics of AF patients.
      TotalOAC PrescriptionNo OAC PrescriptionP Value
      N Patients864,339586,554277,785
      Mean Household Income ($1K)69.43 ± 13.1569.50 ± 13.1269.28 ± 13.21<.001
      Age (years)73.54 ± 11.4374.48 ± 10.3871.57 ± 13.15<.001
      Gender Male490,102 (57)332,949 (57)157,153 (57)0.096
      Weight (kg)88.47 ± 24.3889.52 ± 24.6886.23 ± 23.56<.001
      CHA2DS2-VASc Score3.60 ± 1.713.74 ± 1.643.30 ± 1.81<.001
      Eligible for OAC*734,288 (85)520,909 (89)213,379 (77)<.001
      Race/Ethnicity
      Hispanic25,532 (3)16,834 (2.9)8,698 (3.1)<.001
      Non-Hispanic White572,814 (66)393,402 (67)179,412 (65)
      Non-Hispanic Black37,833 (4.4)24,671 (4.2)13,162 (4.7)
      Other13,268 (1.5)8,864 (1.5)4,404 (1.6)
      Missing214,892 (25)142,783 (24)72,109 (26)
      Clinic Location
      West161,060 (19)111,685 (19)49,375 (18)<.001
      Northeast182,001 (21)124,672 (21)57,329 (21)
      Midwest131,658 (15)89,046 (15)42,612 (15)
      South436,030 (50)292,118 (50)143,912 (52)
      Urban182,659 (21)123,413 (21)59,246 (21)
      Suburban136,431 (16)96,446 (16)39,985 (14)
      Rural32,420 (3.8)22,109 (3.8)10,311 (3.7)
      Size of Clinic (N patients)19,705.51 ± 16,583.8119,994.35 ± 16,805.2519,095.63 ± 16,089.24<.001
      Insurance Type
      Private500,882 (58)340,187 (58)160,695 (58)<.001
      Medicaid54,517 (6.3)35,652 (6.1)18,865 (6.8)
      Medicare526,572 (61)375,615 (64)150,957 (54)
      State7,839 (0.91)5,155 (0.88)2,684 (0.97)
      Other35,070 (4.1)24,250 (4.1)10,820 (3.9)
      None3,225 (0.37)2,147 (0.37)1,078 (0.39)
      Clinic and Lab Values
      Heart Rate (beats per minute)72.20 ± 13.5172.60 ± 13.7271.41 ± 13.03<.001
      Systolic BP (mmHg)127.97 ± 17.20127.77 ± 17.12128.40 ± 17.37<.001
      Diastolic BP (mmHg)73.57 ± 10.4073.45 ± 10.3473.82 ± 10.52<.001
      Total Cholesterol (mg/dL)158.82 ± 40.99156.63 ± 39.95163.63 ± 42.81<.001
      HDL Cholesterol (mg/dL)50.13 ± 16.8149.78 ± 16.6150.90 ± 17.22<.001
      LDL Cholesterol (mg/dL)86.83 ± 34.6585.40 ± 34.0390.00 ± 35.77<.001
      Triglyceride (mg/dL)125.49 ± 70.64125.13 ± 69.96126.27 ± 72.100.008
      INR2.14 ± 2.102.23 ± 1.971.75 ± 2.54<.001
      GFR (mL/min/1.73 m2)63.67 ± 22.9662.74 ± 22.1666.11 ± 24.76<.001
      LVEF (%)54.98 ± 12.9454.34 ± 13.1856.48 ± 12.23<.001
      Comorbidities
      Hypertension664,713 (77)463,872 (79)200,841 (72)<.001
      Dyslipidemia517,586 (60)360,239 (61)157,347 (57)<.001
      Heart Failure238,781 (28)177,488 (30)61,293 (22)<.001
      Stable Angina87,749 (10)56,458 (9.6)31,291 (11)<.001
      Unstable Angina26,105 (3)15,994 (2.7)10,111 (3.6)<.001
      Transient Ischemic Attack56,425 (6.5)40,649 (6.9)15,776 (5.7)<.001
      Ischemic Stroke67,588 (7.8)48,479 (8.3)19,109 (6.9)<.001
      Coronary Artery Disease375,678 (43)257,035 (44)118,643 (43)<.001
      Myocardial Infarction55,161 (6.4)35,313 (6)19,848 (7.1)<.001
      Peripheral Artery Disease99,234 (11)68,325 (12)30,909 (11)<.001
      Peripheral Vascular Disease74,430 (8.6)51,771 (8.8)22,659 (8.2)<.001
      Coronary Artery Bypass Graft63,543 (7.4)40,949 (7)22,594 (8.1)<.001
      Percutaneous Coronary Intervention76,282 (8.8)49,994 (8.5)26,288 (9.5)<.001
      Type 2 Diabetes217,996 (25)157,247 (27)60,749 (22)<.001
      Chronic Kidney Disease89,262 (10)63,272 (11)25,990 (9.4)<.001
      Chronic Liver Disease90,557 (10)63,957 (11)26,600 (9.6)<.001
      Hemodialysis2,370 (0.27)1,492 (0.25)878 (0.32)<.001
      Kidney Transplant590 (0.068)404 (0.069)186 (0.067)0.783
      Hyperthyroidism8,024 (0.93)5,511 (0.94)2,513 (0.9)0.117
      Hypothyroidism55,172 (6.4)37,933 (6.5)17,239 (6.2)<.001
      Sleep Apnea82,975 (9.6)60,563 (10)22,412 (8.1)<.001
      Medications
      Antiplatelets487,204 (56)291,608 (50)195,596 (70)<.001
      Antiplatelets (without aspirin)105,056 (12)64,248 (11)40,808 (15)<.001
      Antiarrhythmic Agents327,998 (38)253,181 (43)74,817 (27)<.001
      Lipid Modifying Agents533,753 (62)390,012 (66)143,741 (52)<.001
      Blood Glucose Regulation Agents167,735 (19)126,588 (22)41,147 (15)<.001
      Antihypertensives775,798 (90)553,571 (94)222,227 (80)<.001
      Aspirin464,821 (54)276,320 (47)188,501 (68)<.001
      Prasugrel6,409 (0.74)4,232 (0.72)2,177 (0.78)0.002
      Ticagrelor7,104 (0.82)4,249 (0.72)2,855 (1)<.001
      Clopidogrel97,316 (11)59,874 (10)37,442 (13)<.001
      Vorapaxar78 (0.009)34 (0.0058)44 (0.016)<.001
      Anticoagulants
      Warfarin232,538 (27)232,538 (40)<.001
      Apixaban232,720 (27)232,720 (40)<.001
      Dabigatran54,136 (6.3)54,136 (9.2)<.001
      Rivaroxaban147,594 (17)147,594 (25)<.001
      Edoxaban4,291 (0.5)4,291 (0.73)<.001
      DOACs401,953 (47)401,953 (69)<.001
      Values are presented as [mean] ± [standard deviation] or [number of patients] ([percent]). Patients were stratified by whether they received an OAC prescription.
      *Class I indication for OAC was determined by whether a patient had an elevated CHA2DS2-VASc score as specified by contemporary guidelines (CHA2DS2-VASc score ≥2 for men, ≥3 for women) (4).
      Abbreviations: OAC: oral anticoagulation; N: number; BP: blood pressure; HDL: high-density lipoprotein LDL: low-density lipoprotein; INR: international normalized ratio; GFR: glomerular filtration rate; LVEF: left ventricular ejection fraction; DOAC: direct oral anticoagulant
      Patients who were prescribed OAC were more likely to reside in the Western US and in suburban counties, were visited in greater clinic size, be non-Hispanic White, be older, had greater household income, and be insured through Medicare.
      Moreover, these patients were more likely to have greater BMI, have a history of hypertension, heart failure, stroke, diabetes, chronic kidney disease, or sleep apnea, and be treated with antihypertensive, antiarrhythmic, lipid modifying, or blood glucose regulating medications.
      Several continuous variables including weight were not available for over 50% of patients, for which imputation methods were employed as described above; completeness information for continuous variables is available in Supplementary Table 1. OAC prescription rates increased with an increasing CHA2DS2-VASc score from 0 to 4, with a slight decrease among those with CHA2DS2-VASc scores >4 (Supplementary Figure 1).

      Geographic Patterns of OAC Use

      Figure 2 displays the geographic patterns of OAC prescription rates by county. Wide variation was observed, with county-level OAC prescription rates ranging from 26.8% to 93.2%. The counties with the lowest OAC prescription rates (<60% OAC coverage, in red) tended to be in urban areas and were more common in Iowa, Florida, Louisiana, Texas, and Virginia. In contrast, nearly all counties in Arizona, Oklahoma, Connecticut, Vermont, and Maine had OAC prescription rates above 60%. Patient characteristics by each quartile of OAC prescription rates are further detailed in Supplementary Table 2.
      Figure thumbnail gr2
      Figure 2Title: OAC Prescription Rates by County. Caption: OAC prescription rates are shown at a population level, split into four quartiles. Prescription rates are defined as: [AF patients with OAC prescriptions]/[AF patients]. The circle size denotes the number of PINNACLE patients treated within that area. The OAC prescription rates are shown by color, with red/orange areas indicating worse rates, and green indicating better rates. A histogram is also shown, indicating how many counties fall in each bin of OAC adherence. Abbreviations: OAC: oral anticoagulation; AF: atrial fibrillation; PINNACLE: Practice Innovation and Clinical Excellence

      Machine Learning Insights: Determinants of OAC Prescription

      The enhanced XGBoost model performed best in its ability to identify whether patients did or did not receive OAC with a test AUC of 0.811 (95% confidence interval [CI]: 0.809–0.813). This significantly surpassed the predictive performance of the CHA2DS2-VASc score (AUC: 0.571, 95% CI: 0.569–0.574, Figure 3). Every enhanced ML model outperformed all versions of the ML models which only relied on the CHA2DS2-VASc variables (Table2). Figure 4 shows features in order of permutation importance within the enhanced XGBoost model. The most predictive patient features include a) use of aspirin, antihypertensives, antiarrhythmic agents, lipid modifying agents, or antiplatelets b) age, c) mean household income, d) INR values, e) clinic size, f) patient weight, and g) US region. Beyond age, variables included in the CHA2DS2-VASc score had low importance in the enhanced random forest model (ranking 12th, 21st, 24th, 25th, 30th, and lower). Supplementary Figure 1 displays the different positive and negative associations of each feature in more detail.
      Figure thumbnail gr3
      Figure 3Title: ROC Curves for Identifying OAC Prescription: Machine Learning vs CHA2DS2-VASc Score. Caption: Receiver operating characteristic curves for 1) the CHA2DS2-VASc score; 2) four ML models (XGBoost, Random Forest, Logistic Regression, LASSO Regression) using only covariates considered in the CHA2DS2-VASc score (Congestive heart failure, Hypertension, Age, Diabetes, Stroke, Vascular disease, and Sex); and 3) four “enhanced” ML models which were trained on additional clinical comorbidities, medication usage, vital signs, laboratory data, insurance information, and socio- and geo-demographic variables. Metrics were calculated on held-out test data. Abbreviations: ROC: receiver operating characteristic; OAC: oral anticoagulation; AUC: area under the receiver operating characteristic curve; CI: confidence interval
      Table 2Summary of model performances for predicting OAC Prescription in training (5-fold cross-validation) and test sets.
      Regular Model CHA2DS2-VASc componentsEnhanced ML Model CHA2DS2-VASc components + new features
      AccuracyAUCPRAUCPrecisionRecallAccuracyAUCPRAUCPrecisionRecall
      XGBOOSTTest Set0.690.620.760.700.960.770.810.890.790.89
      Cross validation0.700.640.770.700.950.780.830.900.800.89
      Logistic RegressionTest Set0.690.600.730.700.960.730.750.860.770.87
      Cross validation0.690.600.730.700.960.740.760.860.760.88
      Random ForestTest Set0.680.590.730.680.990.760.790.850.790.88
      Cross validation0.750.780.880.760.930.990.990.850.990.99
      LASSO-penalized logistic regressionTest Set0.690.600.730.700.960.740.760.880.760.89
      Cross validation0.690.600.730.700.960.740.760.990.760.89
      Figure thumbnail gr4
      Figure 4Title: Predictive importance of individual clinical features on the likelihood of being prescribed an OAC. Caption: The highest-performing ML model, XGBoost, was used to determine the rank order of features associated with OAC prescriptions. This feature importance was measured using the “permutation importance” metric. With the fully trained model, independent variables were randomly shuffled, removing the relationships learned by the ML model, and the decrease in model performance was assessed. The average decrease in performance across 5 independent runs, and the standard deviation of those runs (black error bars), is shown for each variable. Abbreviations: OAC: oral anticoagulation; BP: blood pressure; INR: international normalized ratio; GFR: glomerular filtration rate; LVEF: left ventricular ejection fraction

      DISCUSSION

      In a contemporary cohort of US patients with AF, 68% were treated with OAC. Of the remaining third of AF patients not on OAC, most met a Class I indication for OAC use by contemporary guidelines. Significant geographic variation in OAC use was observed between counties, with highest rates among patients dwelling in suburban settings and in the Western US. Supervised ML analyses outperformed the CHA2DS2-VASc score at predicting OAC use and identified a rank order of associated patient’s sociodemographic features beyond clinical factors. The strongest associations were the use of aspirin, antihypertensives, antiarrhythmic agents, lipid modifying agents, and INR values, as well as the features of age, mean household income, clinic size, patient weight, and geographic region. Our results are largely consistent with prior findings of disparities in OAC prescription by patient characteristics, site of care, and geographic region. One such analysis 21found that, compared with those prescribed OAC, patients with AF prescribed aspirin as their sole antithrombotic therapy were more often located in the South and West, in nonurban settings, and in practices with larger patient volumes.
      While other studies have demonstrated an association between greater burden of clinical comorbidities
      • Savarese G.
      • Sartipy U.
      • Friberg L.
      • Dahlström U.
      • Lund L.H.
      Reasons for and consequences of oral anticoagulant underuse in atrial fibrillation with heart failure.
      and lower likelihood of OAC
      • Savarese G.
      • Sartipy U.
      • Friberg L.
      • Dahlström U.
      • Lund L.H.
      Reasons for and consequences of oral anticoagulant underuse in atrial fibrillation with heart failure.
      , sleep apnea has been associated with increased OAC use

      Johnson, K.G. & Johnson, D.C. Obstructive sleep apnea is a risk factor for stroke and atrial fibrillation. Chest 138, 239; author reply 239-240 (2010).

      . Our analysis revealed that the largest contributions to the predictive model of OAC use were prescriptions related to other comorbid conditions, including hypertension, congestive heart failure, diabetes, and stroke. Similar to other analyses

      Johnson, K.G. & Johnson, D.C. Obstructive sleep apnea is a risk factor for stroke and atrial fibrillation. Chest 138, 239; author reply 239-240 (2010).

      , we also found that use of aspirin and other forms of antiplatelet therapy was associated with lower rates of OAC prescriptions
      • Lubitz S.A.
      • et al.
      Predictors of oral anticoagulant non-prescription in patients with atrial fibrillation and elevated stroke risk.
      . This is possibly due to the increased risk of bleeding in these patients. OAC use was greater in patients who were on antiarrhythmic agents, which may be explained by the increased recurrence and severity of AF in patients on antiarrhythmic therapy.
      Our results also demonstrated geographical variation in OAC prescription, where the counties with the lowest OAC prescription rates (<60% OAC coverage, in red) tended to be in urban areas and were more common in Iowa, Florida, Louisiana, Texas, and Virginia. In contrast, nearly all counties in Arizona, Oklahoma, Connecticut, Vermont, and Maine had OAC prescription rates above 60%. Similar results were obtained in the study by Hernandez et al
      • Hernandez I.
      • Saba S.
      • Zhang Y.
      Geographic Variation in the Use of Oral Anticoagulation Therapy in Stroke Prevention in Atrial Fibrillation.
      , reporting large geographical variations in use of OAC for stroke prevention in patients with atrial fibrillation. In this study Midwest and Northwest had higher likelihood of OAC initiation compared to the South that had lowest likelihood of OAC use and higher risk of stroke.
      One important finding of our analysis was the role of social determinants of health including household income, and clinic size in OAC prescription and adherence. Cost of medication and follow-up of OAC can influence the prescription rates
      • Dalmau Llorca M.R.
      • et al.
      Gender and Socioeconomic Inequality in the Prescription of Direct Oral Anticoagulants in Patients with Non-Valvular Atrial Fibrillation in Primary Care in Catalonia (Fantas-TIC Study).
      ,
      • Teppo K.
      • et al.
      Association of income and educational levels with adherence to direct oral anticoagulant therapy in patients with incident atrial fibrillation: A Finnish nationwide cohort study.
      . Unequal access to OAC in socioeconomically disadvantaged patients and different geographical areas have been shown in previous studies
      • Teppo K.
      • et al.
      Association of income and educational levels with adherence to direct oral anticoagulant therapy in patients with incident atrial fibrillation: A Finnish nationwide cohort study.
      . In a study by Llorca et al
      • Dalmau Llorca M.R.
      • et al.
      Gender and Socioeconomic Inequality in the Prescription of Direct Oral Anticoagulants in Patients with Non-Valvular Atrial Fibrillation in Primary Care in Catalonia (Fantas-TIC Study).
      , those living in more socioeconomic deprived and rural areas had lower OAC prescription rates. Moreover, previous studies by Essien et al.
      • Essien U.R.
      • et al.
      Association of Race and Ethnicity and Anticoagulation in Patients With Atrial Fibrillation Dually Enrolled in Veterans Health Administration and Medicare: Effects of Medicare Part D on Prescribing Disparities.
      ,
      • Essien U.R.
      • et al.
      Association of Race/Ethnicity With Oral Anticoagulant Use in Patients With Atrial Fibrillation: Findings From the Outcomes Registry for Better Informed Treatment of Atrial Fibrillation II.
      showed lower initiation of OAC for Black patients and DOAC use for Black and Hispanic patients. This was also evident in our descriptive results; however, these didn’t emerge as high ranking features in our ML models, which may be due to the low sample sizes of these populations in our database. Therefore, there is a need to condition prescription patterns by sociodemographic factors besides clinical risk factors.
      In the absence of other prediction models to estimate OAC use, we utilized the CHA2DS2-VASc score. Even though the CHA2DS2-VASc score was initially designed to predict thromboembolic risk, prior work has demonstrated increased odds of OAC prescribing with increasing CHA2DS2-VASc score [1]. As such, we hypothesized that it would be modestly predictive of OAC use. We instead found that the performance of CHA2DS2-VASc score was only slightly above chance. Furthermore, the enhanced ML models (which incorporated additional social, geographic, and clinical variables), significantly outperformed the ML models that were limited to risk factors in the CHA2DS2-VASc score. This finding suggests that a range of social and clinical determinants of health likely underlie much of the observed variation in OAC guideline adherence
      • Lubitz S.A.
      • et al.
      Predictors of oral anticoagulant non-prescription in patients with atrial fibrillation and elevated stroke risk.
      .
      Our study extends and complements prior work by leveraging ML methods to identify important patient-level predictors of OAC prescriptions. Exploring the variables selected by ML adds unique insights beyond what traditional regression analysis provides. In real-world datasets, a number of variables may be unavailable for large numbers of patients. When these variables are present, they display markedly nonlinear, and even non-monotonic trends. Additive models such as the CHA2DS2-VASc, which assume equal weights for all risk factors, are unable to fully capture these associations. Furthermore, the rank order of a given feature’s influence on OAC prescribing, considering all other possible permutations of other concomitant features, would not be uncovered by less sophisticated models. While traditional multivariate approaches, such as Bayesian hierarchical linear models, are adept at identifying independent associations between a given patient feature and an outcome, ML enables the integration of potentially hundreds of different features, with varying levels of missingness, to determine the collective associations with clinical outcomes. These observations on an established, longitudinal patient registry demonstrate that ML offers unique additive value—and should continue to be leveraged—to identify non-linear associations between patient features and clinical management practices.

      Clinical Implications

      It is important to translate these findings into actionable value for clinicians and care teams to close critical gaps in medical care. One proposed approach is to utilize an analytics platform to apply guideline-driven insights, both longitudinally and in real-time, for every patient record within a health system at once. In the case of OAC use in eligible AF patients, this “precision population health” engine may alert care teams to focus their attention on specific patient cohorts with confirmed “care gaps” for OAC, or, upon patient cohorts who are at highest risk of developing an OAC care gap, as may be the case with the predictive model presented here.
      This patient-centered novel approach will provide an accurate tool for clinical decision making not only by incorporating clinical factors considered in the previous risk scores, but also including social determinants of health and geographical variations for risk profiling of patients with atrial fibrillation.

      Limitations

      Our results should be interpreted in the context of several limitations. Since our data included patients enrolled predominantly within outpatient cardiology practices, OAC prescribing patterns may not be generalized to non-cardiology practices. Incomplete or missing data may also have impacted our findings. For example, if a feature was not reported in the EHR for a patient (e.g. a history of stroke), it was interpreted in this analysis as the absence of stroke in the calculation of that patient’s CHA2DS2-VASc score. Patients had differing numbers of recorded encounters in the data, which led to differing levels of data completeness. Since this is a voluntary registry, sites that participate in the PINNACLE registry may not be nationally representative and some regions are not well-represented. It is likely that the noted sociodemographic disparities may be greater in other non-registry participating sites. Unlike a randomized controlled trial, the inference of causation is not possible due to the many uncontrolled factors not recorded
      • Hsu J.C.
      • et al.
      Oral Anticoagulant Therapy Prescription in Patients With Atrial Fibrillation Across the Spectrum of Stroke Risk: Insights From the NCDR PINNACLE Registry.
      . There are other confounding factors related to underprescription of OAC including utilization of LAAO devices which were not captured in our study. Moreover, increased risk of bleeding is an important factor associated with lower prescription rates. Future studies assessing this risk using related clinical risk scores (HAS-BLED) are warranted. As this study was based on data in 2017-18, there may be lower than expected DOAC use and/or higher than expected ASA use for a CHA2DS2VASc=1 population given that this would be based on the 2014 guidelines. Future work will explore the specific determinants of DOAC usage versus warfarin usage in the OAC groups, as current guidelines recommend DOACs for most AF patients. Future investigations will also aim to transform these insights into a population health-focused strategy to better target evidence-based interventions that promote closure of gaps in care.

      CONCLUSIONS

      In a contemporary national cohort of patients with AF, almost a third of patients with AF failed to receive OAC, with significant geographic practice variations. Specific, ML-derived predictors of OAC prescription were identified and offer complementary information to traditional analytic methods. Our results demonstrated the role of several important demographic and socioeconomic factors in underutilization of OAC in patients with atrial fibrillation. Therefore, by combining large, representative real-world datasets with ML techniques, features beyond clinical factors contributing to OAC underuse may be identified to inform targets for quality improvement.

      Uncited reference

      • Hsu J.C.
      • et al.
      Aspirin Instead of Oral Anticoagulant Prescription in Atrial Fibrillation Patients at Risk for Stroke.
      .

      References

        • Hsu J.C.
        • et al.
        Oral Anticoagulant Therapy Prescription in Patients With Atrial Fibrillation Across the Spectrum of Stroke Risk: Insights From the NCDR PINNACLE Registry.
        JAMA cardiology. 2016; 1: 55-62
        • Marzec L.N.
        • et al.
        Influence of Direct Oral Anticoagulants on Rates of Oral Anticoagulation for Atrial Fibrillation.
        Journal of the American College of Cardiology. 2017; 69: 2475-2484
        • Lip G.Y.
        • Nieuwlaat R.
        • Pisters R.
        • Lane D.A.
        • Crijns H.J.
        Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the euro heart survey on atrial fibrillation.
        Chest. 2010; 137: 263-272
        • Havranek E.P.
        • et al.
        Social Determinants of Risk and Outcomes for Cardiovascular Disease.
        Circulation. 2015; 132: 873-898
        • Hernandez I.
        • Saba S.
        • Zhang Y.
        Geographic Variation in the Use of Oral Anticoagulation Therapy in Stroke Prevention in Atrial Fibrillation.
        Stroke. 2017; 48: 2289-2291
        • January C.T.
        • et al.
        2019 AHA/ACC/HRS Focused Update of the 2014 AHA/ACC/HRS Guideline for the Management of Patients With Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society.
        Journal of the American College of Cardiology. 2019; 74: 104-132
        • Yao X.
        • et al.
        Effect of Adherence to Oral Anticoagulants on Risk of Stroke and Major Bleeding Among Patients With Atrial Fibrillation.
        Journal of the American Heart Association. 2016; 5
        • Jackevicius C.A.
        • et al.
        Early non-persistence with dabigatran and rivaroxaban in patients with atrial fibrillation.
        Heart (British Cardiac Society). 2017; 103: 1331-1338
        • Patel M.R.
        • et al.
        Rivaroxaban versus warfarin in nonvalvular atrial fibrillation.
        The New England journal of medicine. 2011; 365: 883-891
        • Granger C.B.
        • et al.
        Apixaban versus warfarin in patients with atrial fibrillation.
        The New England journal of medicine. 2011; 365: 981-992
        • Connolly S.J.
        • et al.
        Dabigatran versus warfarin in patients with atrial fibrillation.
        The New England journal of medicine. 2009; 361: 1139-1151
        • Giugliano R.P.
        • et al.
        Edoxaban versus warfarin in patients with atrial fibrillation.
        The New England journal of medicine. 2013; 369: 2093-2104
        • January C.T.
        • et al.
        2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: executive summary: a report of the American College of Cardiology/American Heart Association Task Force on practice guidelines and the Heart Rhythm Society.
        Circulation. 2014; 130: 2071-2104
        • Chan P.S.
        • Maddox T.M.
        • Tang F.
        • Spinler S.
        • Spertus J.A.
        Practice-level variation in warfarin use among outpatients with atrial fibrillation (from the NCDR PINNACLE program).
        The American journal of cardiology. 2011; 108: 1136-1140
        • Messenger J.C.
        • et al.
        The National Cardiovascular Data Registry (NCDR) Data Quality Brief: the NCDR Data Quality Program in 2012.
        Journal of the American College of Cardiology. 2012; 60: 1484-1488
        • Thompson L.E.
        • et al.
        Sex Differences in the Use of Oral Anticoagulants for Atrial Fibrillation: A Report From the National Cardiovascular Data Registry (NCDR(®)) PINNACLE Registry.
        Journal of the American Heart Association. 2017; 6
        • Breiman L.
        Random Forests.
        Machine Learning. 2001; 45: 5-32
      1. Chen, T. & Guestrin, C. Xgboost: A scalable tree boosting system. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining 785-794 (2016).

        • Strobl C.
        • Boulesteix A.L.
        • Zeileis A.
        • Hothorn T.
        Bias in random forest variable importance measures: illustrations, sources and a solution.
        BMC bioinformatics. 2007; 8: 25
        • Altmann A.
        • Toloşi L.
        • Sander O.
        • Lengauer T.
        Permutation importance: a corrected feature importance measure.
        Bioinformatics (Oxford, England). 2010; 26: 1340-1347
        • Hsu J.C.
        • et al.
        Aspirin Instead of Oral Anticoagulant Prescription in Atrial Fibrillation Patients at Risk for Stroke.
        Journal of the American College of Cardiology. 2016; 67: 2913-2923
        • Savarese G.
        • Sartipy U.
        • Friberg L.
        • Dahlström U.
        • Lund L.H.
        Reasons for and consequences of oral anticoagulant underuse in atrial fibrillation with heart failure.
        Heart (British Cardiac Society). 2018; 104: 1093-1100
      2. Johnson, K.G. & Johnson, D.C. Obstructive sleep apnea is a risk factor for stroke and atrial fibrillation. Chest 138, 239; author reply 239-240 (2010).

        • Lubitz S.A.
        • et al.
        Predictors of oral anticoagulant non-prescription in patients with atrial fibrillation and elevated stroke risk.
        American heart journal. 2018; 200: 24-31
        • Dalmau Llorca M.R.
        • et al.
        Gender and Socioeconomic Inequality in the Prescription of Direct Oral Anticoagulants in Patients with Non-Valvular Atrial Fibrillation in Primary Care in Catalonia (Fantas-TIC Study).
        International journal of environmental research and public health. 2021; 18
        • Teppo K.
        • et al.
        Association of income and educational levels with adherence to direct oral anticoagulant therapy in patients with incident atrial fibrillation: A Finnish nationwide cohort study.
        Pharmacology Research & Perspectives. 2022; 10e00961
        • Essien U.R.
        • et al.
        Association of Race and Ethnicity and Anticoagulation in Patients With Atrial Fibrillation Dually Enrolled in Veterans Health Administration and Medicare: Effects of Medicare Part D on Prescribing Disparities.
        Circulation: Cardiovascular Quality and Outcomes. 2022; 15e008389
        • Essien U.R.
        • et al.
        Association of Race/Ethnicity With Oral Anticoagulant Use in Patients With Atrial Fibrillation: Findings From the Outcomes Registry for Better Informed Treatment of Atrial Fibrillation II.
        JAMA cardiology. 2018; 3: 1174-1182