XML Template (2014) [23.7.2014–9:12am] //blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/CPRJ/Vol00000/140099/APPFile/SG-CPRJ140099.3d
(CPR)
[1–8] [PREPRINTER stage]
EURO PEAN SO CIETY O F CARDIOLOGY ®
Original scientific paper
Performance of the CHARGE-AF risk model for incident atrial fibrillation in the EPIC Norfolk cohort
European Journal of Preventive Cardiology 0(00) 1–8 ! The European Society of Cardiology 2014 Reprints and permissions: sagepub.co.uk/journalsPermissions.nav DOI: 10.1177/2047487314544045 ejpc.sagepub.com
Roman Pfister1,*, Johannes Bra¨gelmann1,*, Guido Michels1, Nick J Wareham2, Robert Luben3 and Kay-Tee Khaw3
Abstract Background: Identification of individuals at risk for developing atrial fibrillation (AF) will help to target screening and preventive interventions. We aimed to validate the CHARGE-AF model (including variables age, race, height, weight, blood pressure, smoking, antihypertensive medication, diabetes, myocardial infarction and heart failure) for prediction of five-year incident AF in a representative European population with a wide age range. Methods and results: The CHARGE-AF model was calculated in 24,020 participants of the population-based EPIC Norfolk study with 236 cases of hospitalization with diagnosis of AF within five years. The model showed good discrimination (c-statistic 0.81, 95% confidence interval (CI) 0.75–0.85), but weak calibration (Chi2-statistic 142) with an almost two-fold overestimation of AF incidence. A recalibration to characteristics of the European Prospective Investigation into Cancer and Nutrition (EPIC) Norfolk cohort improved calibration considerably (Chi2-statistic 13.3), with acceptable discrimination in participants both >65 and 65 years of age (c-statistics 0.70, 95% CI 0.61–0.77 and 0.83, 95% CI 0.74–0.88). The recalibrated model also showed good discrimination in participants free of cardiovascular disease (c-statistics 0.80, 95% CI 0.75–0.84). Categories of predicted risk (5%) showed good concordance with observed five-year AF incidence of 0.62%, 3.49% and 8.74% (log rank test p < 0.001), respectively. Conclusion: A recalibration of the CHARGE-AF model is necessary for accurate predictions of five-year risk of AF in the EPIC Norfolk population. The recalibrated model showed good discrimination across a wide age range and in individuals free of cardiovascular disease, and hence is broadly applicable in primary care to identify people at risk for development of AF.
Keywords Atrial fibrillation, score, risk prediction Received 30 April 2014; accepted 28 June 2014
Introduction Atrial fibrillation (AF) is the most frequent sustained arrhythmia in the general population1 and recent projections expect that from 2010 to 2060 the number of adults with AF will more than double.2 Notably, AF is associated with an increased morbidity3,4 and mortality5 and thus places a significant burden in terms of quality of life and health care costs. Increasing emphasis is put on a better understanding of AF epidemiology and development of primary prevention interventions. Recently, an expert panel proposed the development of incident AF risk prediction models and the validation in multiple independent cohorts as a primary research recommendation. Several AF risk prediction models have been proposed
earlier which were based on single US-cohorts.6,7 However, these models have not been validated
1 Department III of Internal Medicine, Heart Centre of the University of Cologne, Germany 2 Medical Research Council Epidemiology Unit, Institute of Metabolic Science, University of Cambridge, UK 3 Department of Public Health and Primary Care, Institute of Public Health, University of Cambridge, UK
*These authors contributed equally. Corresponding author: Roman Pfister, Department III of Internal Medicine, University of Cologne, Kerpenerstr. 62, 50937 Cologne, Germany. Email: [email protected]
XML Template (2014) [23.7.2014–9:12am] //blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/CPRJ/Vol00000/140099/APPFile/SG-CPRJ140099.3d
(CPR)
[1–8] [PREPRINTER stage]
2 thoroughly or included risk factors which are not routinely available in primary care such as PR interval on ECG and echocardiographic left-ventricular hypertrophy and left atrial enlargement.6,7 Recently the CHARGE-HF consortium published a simple risk model derived from three US cohorts using variables readily available in a primary care setting (age, race, height, weight, blood pressure, smoking, antihypertensive medication, diabetes, myocardial infarction and heart failure) which showed good risk prediction.8 So far it is unclear how the model performs in middle-aged European populations since the validation of the CHARGE-AF model was performed in elderly cohorts with the majority of people aged more than 65 years. Furthermore, from the clinical point of view estimating risk for development of AF might be particularly relevant in people without manifest cardiovascular disease since these individuals usually do not undergo routine monitoring with, for example, electrocardiogram (ECG). Notably, people with prevalent cardiovascular disease were included in the CHARGE-AF derivation cohorts and cardiovascular diseases are important risk variables included in the model. The aim of the present study was to investigate the discriminatory performance of the CHARGE-AF risk model for prediction of development of AF in the European Prospective Investigation into Cancer and Nutrition (EPIC) Norfolk cohort9 with particular focus on the performance in participants without prevalent cardiovascular risk factors and disease.
Methods Participants The EPIC Norfolk is a prospective population study of 25,639 men and women aged between 39 and 79 years, resident in Norfolk, UK. Details of the recruitment process, study design and population characteristics have been published earlier.9 The EPIC Norfolk study was approved by the Norfolk Local Research Ethics Committee, and participants gave signed informed consent at each contact. At the baseline survey between 1993 and 1997 participants completed a detailed health and lifestyle questionnaire. Trained nurses examined individuals at a clinic visit measuring height and weight. Blood pressure was measured with an Accutorr noninvasive blood pressure monitor (Datascope Medical, Huntington, UK) after the participant had been seated for 5 min. We used the mean of two measurements for analysis. Non-fasting blood samples were taken by venepuncture, stored at 4 C overnight and assayed at the department of clinical biochemistry, University of Cambridge, Cambridge, UK. We measured serum total cholesterol, high-density lipoprotein
European Journal of Preventive Cardiology 0(00) cholesterol (HDL), and triglycerides from serum samples with the RA 1000 (Bayer Diagnostics, Basingstoke, UK). Low-density lipoprotein levels were calculated using the Friedewald formula. Prevalent heart failure was defined by self-reported intake of drugs, which has shown high specificity.10
Case ascertainment Since ECG was not performed in EPIC Norfolk participants we defined prevalent AF by self-reported intake of drugs that were used for treatment of AF in clinical practice at the time of the baseline survey (digitalis or vitamin K antagonists). Of patients with AF in primary care 92% were shown to be treated with digitalis in UK and 36–44% were treated with vitamin K antagonists.11,12 Conversely, 87% of patients discharged with digitalis had AF and the reason for vitamin K antagonist treatment in primary care was AF in 28% of patients.13,14 Incident AF cases were ascertained by using hospital record linkage with virtually complete follow up. Participants are linked to National Health Service hospital information systems so that hospital admissions with inpatient treatment anywhere in the UK are reported to EPIC Norfolk through routine annual record linkage. This linkage does not cover hospital outpatient treatment or emergency department introduction without subsequent hospital admission. Incident AF was defined as International Classification of Disease (ICD)-10 hospital discharge code I48, including atrial fibrillation and atrial flutter. The current study is based on follow-up through March 2009. All participants are flagged for death at the UK Office of National Statistics. Observations were censored at the date of death or 31 March 2009.
Statistical analysis We examined differences of risk factors previously evaluated for the derivation of the CHARGE-AF risk model8 comparing participants with and without incident AF by t-test and Chi2 test. Despite an available mean follow-up of 12.5 years we restricted our primary analyses to AF events occurring during the first five years of follow-up since the CHARGE-AF model was developed for prediction of five-year risk of development of AF.8 ß-coefficients were calculated from a multivariate Cox proportional hazard model incorporating all clinical co-variables included in the CHARGE-AF risk model. In the initial CHARGE study an augmented model which initially evaluated ECG and laboratory parameters and finally included left-ventricular hypertrophy and PR interval by ECG in addition to clinical variables showed no
XML Template (2014) [23.7.2014–9:12am] //blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/CPRJ/Vol00000/140099/APPFile/SG-CPRJ140099.3d
(CPR)
[1–8] [PREPRINTER stage]
Pfister et al. improvement in performance.8 For each individual in the EPIC Norfolk cohort the predicted five-year risk of incident AF was therefore calculated following the simple risk model formula given in Alonso et al.8 Subsequently the cohort was split into deciles of predicted risks and the mean risk per decile was compared with the mean observed risk per decile. Discrimination of risk prediction was evaluated using the discrimination slope15 and the overall c-statistic.16 Calibration was assessed with a modified version of the Hosmer– Lemeshow goodness of fit test as proposed by D’Agostino and Nam.17 Chi2 values >20 were considered as a significant lack of calibration. Accordingly p-values 65 years of age, but calibration was substantially weaker in participants 65 years (Table 3). In secondary analyses calculations were repeated excluding participants with a self-report of cardiac arrhythmia (N ¼ 1156), which yielded comparable results (data not shown).
CHARGE-AF risk model in participants free of cardiovascular disease For this analysis people with cardiovascular disease at baseline (N ¼ 1035, 42 events) were excluded and risk predictions with the original CHARGE-AF risk model and the recalibrated model were calculated. Of 22,985 individuals 194 experienced incident AF within five years after study entry. Discrimination was slightly lower compared with the overall cohort (Table 3). The original CHARGE-AF model again overestimated the AF incidence whereas the recalibrated model correctly predicted AF risk (online Supplementary Data Figure S1; Table 3). The observed five-year incidence
Results We excluded 334 participants with baseline use of digitalis or vitamin K antagonists as prevalent AF and 1285 individuals with missing data on one of the variables under study, which left 24,020 participants for analysis. During a mean follow-up of 12.5 years in total 1386 participants experienced incident AF. Of these incident AF cases 236 (0.98% of the total population) occurred within five years after study entry. Table 1 shows baseline characteristics of individuals with and without incident AF within the first five years. Table 2 shows ß-coefficients of individual risk factors derived from the multivariate model. Relative risk associated with individual variables was similar or slightly lower compared with the ß-coefficients of the CHARGE-AF risk model. Of note, race, history of diabetes, and systolic and diastolic blood pressure did not reach statistical significance in the multivariate model in our cohort.
Table 1. Baseline characteristics by incident atrial fibrillation in the first five years after study entry. No incident AF (n ¼ 23,784) Sex (male)
10,656 (44.8%)
Age (years)
58.5 (9.25)
Myocardial infarction (yes)
667 (2.8%)
Incident AF (n ¼ 236) 151 (64.0%) 66.8 (6.79) 27 (11.4%)
p-valuea
(CPR)
[1–8] [PREPRINTER stage]
EURO PEAN SO CIETY O F CARDIOLOGY ®
Original scientific paper
Performance of the CHARGE-AF risk model for incident atrial fibrillation in the EPIC Norfolk cohort
European Journal of Preventive Cardiology 0(00) 1–8 ! The European Society of Cardiology 2014 Reprints and permissions: sagepub.co.uk/journalsPermissions.nav DOI: 10.1177/2047487314544045 ejpc.sagepub.com
Roman Pfister1,*, Johannes Bra¨gelmann1,*, Guido Michels1, Nick J Wareham2, Robert Luben3 and Kay-Tee Khaw3
Abstract Background: Identification of individuals at risk for developing atrial fibrillation (AF) will help to target screening and preventive interventions. We aimed to validate the CHARGE-AF model (including variables age, race, height, weight, blood pressure, smoking, antihypertensive medication, diabetes, myocardial infarction and heart failure) for prediction of five-year incident AF in a representative European population with a wide age range. Methods and results: The CHARGE-AF model was calculated in 24,020 participants of the population-based EPIC Norfolk study with 236 cases of hospitalization with diagnosis of AF within five years. The model showed good discrimination (c-statistic 0.81, 95% confidence interval (CI) 0.75–0.85), but weak calibration (Chi2-statistic 142) with an almost two-fold overestimation of AF incidence. A recalibration to characteristics of the European Prospective Investigation into Cancer and Nutrition (EPIC) Norfolk cohort improved calibration considerably (Chi2-statistic 13.3), with acceptable discrimination in participants both >65 and 65 years of age (c-statistics 0.70, 95% CI 0.61–0.77 and 0.83, 95% CI 0.74–0.88). The recalibrated model also showed good discrimination in participants free of cardiovascular disease (c-statistics 0.80, 95% CI 0.75–0.84). Categories of predicted risk (5%) showed good concordance with observed five-year AF incidence of 0.62%, 3.49% and 8.74% (log rank test p < 0.001), respectively. Conclusion: A recalibration of the CHARGE-AF model is necessary for accurate predictions of five-year risk of AF in the EPIC Norfolk population. The recalibrated model showed good discrimination across a wide age range and in individuals free of cardiovascular disease, and hence is broadly applicable in primary care to identify people at risk for development of AF.
Keywords Atrial fibrillation, score, risk prediction Received 30 April 2014; accepted 28 June 2014
Introduction Atrial fibrillation (AF) is the most frequent sustained arrhythmia in the general population1 and recent projections expect that from 2010 to 2060 the number of adults with AF will more than double.2 Notably, AF is associated with an increased morbidity3,4 and mortality5 and thus places a significant burden in terms of quality of life and health care costs. Increasing emphasis is put on a better understanding of AF epidemiology and development of primary prevention interventions. Recently, an expert panel proposed the development of incident AF risk prediction models and the validation in multiple independent cohorts as a primary research recommendation. Several AF risk prediction models have been proposed
earlier which were based on single US-cohorts.6,7 However, these models have not been validated
1 Department III of Internal Medicine, Heart Centre of the University of Cologne, Germany 2 Medical Research Council Epidemiology Unit, Institute of Metabolic Science, University of Cambridge, UK 3 Department of Public Health and Primary Care, Institute of Public Health, University of Cambridge, UK
*These authors contributed equally. Corresponding author: Roman Pfister, Department III of Internal Medicine, University of Cologne, Kerpenerstr. 62, 50937 Cologne, Germany. Email: [email protected]
XML Template (2014) [23.7.2014–9:12am] //blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/CPRJ/Vol00000/140099/APPFile/SG-CPRJ140099.3d
(CPR)
[1–8] [PREPRINTER stage]
2 thoroughly or included risk factors which are not routinely available in primary care such as PR interval on ECG and echocardiographic left-ventricular hypertrophy and left atrial enlargement.6,7 Recently the CHARGE-HF consortium published a simple risk model derived from three US cohorts using variables readily available in a primary care setting (age, race, height, weight, blood pressure, smoking, antihypertensive medication, diabetes, myocardial infarction and heart failure) which showed good risk prediction.8 So far it is unclear how the model performs in middle-aged European populations since the validation of the CHARGE-AF model was performed in elderly cohorts with the majority of people aged more than 65 years. Furthermore, from the clinical point of view estimating risk for development of AF might be particularly relevant in people without manifest cardiovascular disease since these individuals usually do not undergo routine monitoring with, for example, electrocardiogram (ECG). Notably, people with prevalent cardiovascular disease were included in the CHARGE-AF derivation cohorts and cardiovascular diseases are important risk variables included in the model. The aim of the present study was to investigate the discriminatory performance of the CHARGE-AF risk model for prediction of development of AF in the European Prospective Investigation into Cancer and Nutrition (EPIC) Norfolk cohort9 with particular focus on the performance in participants without prevalent cardiovascular risk factors and disease.
Methods Participants The EPIC Norfolk is a prospective population study of 25,639 men and women aged between 39 and 79 years, resident in Norfolk, UK. Details of the recruitment process, study design and population characteristics have been published earlier.9 The EPIC Norfolk study was approved by the Norfolk Local Research Ethics Committee, and participants gave signed informed consent at each contact. At the baseline survey between 1993 and 1997 participants completed a detailed health and lifestyle questionnaire. Trained nurses examined individuals at a clinic visit measuring height and weight. Blood pressure was measured with an Accutorr noninvasive blood pressure monitor (Datascope Medical, Huntington, UK) after the participant had been seated for 5 min. We used the mean of two measurements for analysis. Non-fasting blood samples were taken by venepuncture, stored at 4 C overnight and assayed at the department of clinical biochemistry, University of Cambridge, Cambridge, UK. We measured serum total cholesterol, high-density lipoprotein
European Journal of Preventive Cardiology 0(00) cholesterol (HDL), and triglycerides from serum samples with the RA 1000 (Bayer Diagnostics, Basingstoke, UK). Low-density lipoprotein levels were calculated using the Friedewald formula. Prevalent heart failure was defined by self-reported intake of drugs, which has shown high specificity.10
Case ascertainment Since ECG was not performed in EPIC Norfolk participants we defined prevalent AF by self-reported intake of drugs that were used for treatment of AF in clinical practice at the time of the baseline survey (digitalis or vitamin K antagonists). Of patients with AF in primary care 92% were shown to be treated with digitalis in UK and 36–44% were treated with vitamin K antagonists.11,12 Conversely, 87% of patients discharged with digitalis had AF and the reason for vitamin K antagonist treatment in primary care was AF in 28% of patients.13,14 Incident AF cases were ascertained by using hospital record linkage with virtually complete follow up. Participants are linked to National Health Service hospital information systems so that hospital admissions with inpatient treatment anywhere in the UK are reported to EPIC Norfolk through routine annual record linkage. This linkage does not cover hospital outpatient treatment or emergency department introduction without subsequent hospital admission. Incident AF was defined as International Classification of Disease (ICD)-10 hospital discharge code I48, including atrial fibrillation and atrial flutter. The current study is based on follow-up through March 2009. All participants are flagged for death at the UK Office of National Statistics. Observations were censored at the date of death or 31 March 2009.
Statistical analysis We examined differences of risk factors previously evaluated for the derivation of the CHARGE-AF risk model8 comparing participants with and without incident AF by t-test and Chi2 test. Despite an available mean follow-up of 12.5 years we restricted our primary analyses to AF events occurring during the first five years of follow-up since the CHARGE-AF model was developed for prediction of five-year risk of development of AF.8 ß-coefficients were calculated from a multivariate Cox proportional hazard model incorporating all clinical co-variables included in the CHARGE-AF risk model. In the initial CHARGE study an augmented model which initially evaluated ECG and laboratory parameters and finally included left-ventricular hypertrophy and PR interval by ECG in addition to clinical variables showed no
XML Template (2014) [23.7.2014–9:12am] //blrnas3/cenpro/ApplicationFiles/Journals/SAGE/3B2/CPRJ/Vol00000/140099/APPFile/SG-CPRJ140099.3d
(CPR)
[1–8] [PREPRINTER stage]
Pfister et al. improvement in performance.8 For each individual in the EPIC Norfolk cohort the predicted five-year risk of incident AF was therefore calculated following the simple risk model formula given in Alonso et al.8 Subsequently the cohort was split into deciles of predicted risks and the mean risk per decile was compared with the mean observed risk per decile. Discrimination of risk prediction was evaluated using the discrimination slope15 and the overall c-statistic.16 Calibration was assessed with a modified version of the Hosmer– Lemeshow goodness of fit test as proposed by D’Agostino and Nam.17 Chi2 values >20 were considered as a significant lack of calibration. Accordingly p-values 65 years of age, but calibration was substantially weaker in participants 65 years (Table 3). In secondary analyses calculations were repeated excluding participants with a self-report of cardiac arrhythmia (N ¼ 1156), which yielded comparable results (data not shown).
CHARGE-AF risk model in participants free of cardiovascular disease For this analysis people with cardiovascular disease at baseline (N ¼ 1035, 42 events) were excluded and risk predictions with the original CHARGE-AF risk model and the recalibrated model were calculated. Of 22,985 individuals 194 experienced incident AF within five years after study entry. Discrimination was slightly lower compared with the overall cohort (Table 3). The original CHARGE-AF model again overestimated the AF incidence whereas the recalibrated model correctly predicted AF risk (online Supplementary Data Figure S1; Table 3). The observed five-year incidence
Results We excluded 334 participants with baseline use of digitalis or vitamin K antagonists as prevalent AF and 1285 individuals with missing data on one of the variables under study, which left 24,020 participants for analysis. During a mean follow-up of 12.5 years in total 1386 participants experienced incident AF. Of these incident AF cases 236 (0.98% of the total population) occurred within five years after study entry. Table 1 shows baseline characteristics of individuals with and without incident AF within the first five years. Table 2 shows ß-coefficients of individual risk factors derived from the multivariate model. Relative risk associated with individual variables was similar or slightly lower compared with the ß-coefficients of the CHARGE-AF risk model. Of note, race, history of diabetes, and systolic and diastolic blood pressure did not reach statistical significance in the multivariate model in our cohort.
Table 1. Baseline characteristics by incident atrial fibrillation in the first five years after study entry. No incident AF (n ¼ 23,784) Sex (male)
10,656 (44.8%)
Age (years)
58.5 (9.25)
Myocardial infarction (yes)
667 (2.8%)
Incident AF (n ¼ 236) 151 (64.0%) 66.8 (6.79) 27 (11.4%)
p-valuea
