Predicting Heart Failure in Patients with Atrial Fibrillation: A Report from the Prospective COOL-AF Registry

Background: This study aimed to determine risk factors and incidence rate and develop a predictive risk model for heart failure for Asian patients with atrial fibrillation (AF). Methods: This is a prospective multicenter registry of patients with non-valvular AF in Thailand conducted between 2014 and 2017. The primary outcome was the occurrence of an HF event. A predictive model was developed using a multivariable Cox-proportional model. The predictive model was assessed using C-index, D-statistics, Calibration plot, Brier test, and survival analysis. Results: There were a total of 3402 patients (average age 67.4 years, 58.2% male) with mean follow-up duration of 25.7 ± 10.6 months. Heart failure occurred in 218 patients during follow-up, representing an incidence rate of 3.03 (2.64–3.46) per 100 person-years. There were ten HF clinical factors in the model. The predictive model developed from these factors had a C-index and D-statistic of 0.756 (95% CI: 0.737–0.775) and 1.503 (95% CI: 1.372–1.634), respectively. The calibration plots showed a good agreement between the predicted and observed model with the calibration slope of 0.838. The internal validation was confirmed using the bootstrap method. The Brier score indicated that the model had a good prediction for HF. Conclusions: We provide a validated clinical HF predictive model for patients with AF, with good prediction and discrimination values.


Introduction
Heart failure (HF) is one of the most common comorbid conditions in patients with atrial fibrillation (AF) [1,2]. Long-standing AF can lead to HF and, conversely, HF can also increase risk of AF [3]. In the Global Anticoagulant Registry in the Field-Atrial Fibrillation (GARFIELD-AF) registry, the incidence rate of HF in patients with newly diagnosed AF was 2.41 per 100 person-years, which is even greater than the incidence rate of ischemic stroke and major bleeding [4]. In Asian patients with AF, the incidence of HF was 3.01-4.25 per 100 person-years [5,6], which is perhaps higher than that reported from non-Asian data [4,7].
AF and HF have a bidirectional pathophysiological link including the development of fibrosis, structural and electrical remodeling, and neurohormonal activation [8]. They have mutual risk factors such as aging, obesity, diabetes, hypertension, and coronary artery disease (CAD). AF can lead to cardiomyopathy and HF especially during rapid ventricular response. Likewise, the increased left ventricular filling pressure can lead to high left atrial pressure and initiate or worsen the AF condition [8]. The rate control of patients

Methods
The Cohort of Antithrombotic Use and Optimal INR Level in Patients with Nonvalvular Atrial Fibrillation in Thailand (COOL-AF Thailand) registry was a prospective nationwide multicenter study of patients with non-valvular AF aged more than 18 years. The exclusion criteria were as follows: (i) rheumatic mitral valve disease; (ii) prosthetic mechanical valve; (iii) inability to have a follow-up visit; (iv) life expectancy less than 3 years; (v) AF from transient reversible cause; (vi) ischemic stroke of less than 3 months; (vii) hematologic disease that increased the risk of bleeding such as myeloproliferative disease; and (viii) refusal to participate.
The study was approved by the Central Research Ethics Committee (CREC) with the Certificate of Approval number COA-CREC 003/2014. Written informed consent was obtained prior to participation.

Study Protocol
After informed consent, the baseline data were collected from the medical records and patient interviews. All the data were written in the case record form and transferred into the web-based system. The case record forms were sent to the data management unit where all the data were verified. Queries were created when any questions arose during the verification process. The follow-up data were collected every 6 months in a similar manner. The site monitoring was performed for every study site to ensure the data quality.

Data Collection
The following baseline data were collected: demographic, weight, height, details of AF, medical history, concomitant diseases, physical examination, medications, laboratory data, ECG and investigational lab data, and components of CHA 2 DS 2 -VASc and HAS-BLED score. For follow-up visits, the data were recorded similar to the baseline visit but also included the clinical outcome data.

Study Outcomes
The main outcome in this study was the occurrence of an HF event. A HF event was defined as a hospital admission or a presentation of the patient for an urgent, unscheduled clinic/office/emergency department visit, with a primary diagnosis of HF, whereby the patient exhibits new or worsening symptoms of HF on presentation, has objective evidence of new or worsening HF, and receives initiation or intensification of treatment specifically for HF [18]. The investigators were required to upload essential documentation to support the diagnosis of HF into the web-based system. All the supporting documents in the web system were sent to the adjudication committee to confirm the diagnosis.

Statistical Analysis
The continuous data were presented as mean and standard deviation (SD) and compared between 2 groups with the student t-test for unpaired data. The categorical data were presented as number and percentages and compared using the chi-square test. The univariable and multivariable analysis was performed to identify the factors that predict the HF outcome using the Cox proportional Hazard model. A univariable analysis was performed using all the baseline variables. The variables with p-values < 0.05 in the univariable analysis were selected for multivariable analysis using backward elimination with p-values < 0.05 as the stopping criteria. Since the data of this study were prospectively collected, the data were relatively completed. The only variable that had missing data is the left ventricular ejection fraction (LVEF) from an echocardiogram, which was missing in 512 cases (15%). We applied imputation for LVEF for the missing data to solve the issue.

Model Development
The risk prediction model for HF was developed from the variables derived from the multivariable analysis. The probability of HF at 3 years for each patient was calculated by using the following equation that is derived from the Cox proportional hazards model: P Heart failure at 3 years = 1 − S 0 (t) exp (Prognostic Index) where S 0 (t) is the average survival probability at time t (i.e., at 3 years), prognostic index is the sum of the predictor product, and the coefficient obtained from the multivariable analysis.

Model Validation
The assessments of the risk of bias and concerns regarding model applicability were performed according to the Prediction model Risk of Bias Assessment Tool (PROBAST) suggestion [19]. Bootstrapping was used to verify the fitted model, based on 100 bootstrap samples. The calibration was tested using the calibration slope based on the observed and predicted hazards of HF. A calibration slope of close to 1 shows a good agreement [20]. The C-statistics were used to measure the model discrimination. The C-statistics vary from 0 to 1 and the value close to 1 indicates a good prediction model [21]. A Receiving Operating Characteristics (ROC) curve was used to compare the C-statistics. The D-statistics assess the discrimination ability of the fitted model into low-risk and high-risk patients [21]. The D-statistic would be 0 or higher; if the D-statistic was 0, it can be concluded that this risk prediction model is not capable of discriminating. The C-and D-statistics were also calculated after the bootstraps as an internal validation. The apparent and optimismcorrected values of the C-and D-statistics were reported. The Cox proportional Hazard model was also used to test the differences in the cumulative event rate between the patients who belong to different risk groups according to the predictive model. The Brier score was evaluated to assess the predictive ability of the model [22]. The Net Reclassification Index (NRI) and Integrated Discrimination Index (IDI) was performed based on the previously proposed methods to determine the influence of the predictive model on the reclassification of the study subjects [23]. Decision-curve analyses (DCA) were used to estimate the clinical usefulness of the HF prediction models by assessing the ability to make better decisions with a model than without [24].
A sensitivity analysis was performed for the assessment of the model for the prediction of HF in those who did not have a history of HF to test the benefit of the model to predict new HF cases. The sensitivity analysis was also performed on the complete dataset after the removal of patients with missing LVEF data. Additional analysis was performed to test the COOL-AF prediction model that was created from the COOL-AF population after the removal of patients with a history of heart failure or LVEF < 50%. The COOL-AF model was compared with other models.
The predictive model of the COOL-AF study was compared with 3 previously reported models from the Framingham Heart Study [16], ORBIT-AF study [15], and a study from Japan [17]. Since 2 previous studies [15,17] reported variables in the multivariable analysis and did not report the regression model, we fitted all the variables in the multivariable analysis into the regression model for comparison purposes. One study [16] did report the regression model but it was the model for the 10-year risk of HF. Therefore, we fitted the reported variables for 3-year HF risk for comparison.
All the statistics were performed using the SPSS statistical software version 18.0 (SPSS, Inc., Chicago, IL, USA) and R version 3.6.3 (www.r-project.org, accessed on 1 March 2020).

Results
We included a total of 3402 patients in this analysis ( Figure 1): mean age 67.4 ± 11.3 years, 1980 (58.2%) male. The average follow-up duration was 25.7 ± 10.6 months or 7912 personyears, during which HF developed during follow-up in 218 patients (6.4%). The incidence rate of HF was 3.03 (2.64-3.46) per 100 person-years. The baseline characteristics of the study population with and without HF outcome are shown in Table 1. There were significant differences between the patients with and without HF in age, proportion of females, and various comorbidities.
after the removal of patients with a history of heart failure or LVEF <50%. The COOL-AF model was compared with other models.
The predictive model of the COOL-AF study was compared with 3 previously reported models from the Framingham Heart Study [16], ORBIT-AF study [15], and a study from Japan [17]. Since 2 previous studies [15,17] reported variables in the multivariable analysis and did not report the regression model, we fitted all the variables in the multivariable analysis into the regression model for comparison purposes. One study [16] did report the regression model but it was the model for the 10-year risk of HF. Therefore, we fitted the reported variables for 3-year HF risk for comparison.
All the statistics were performed using the SPSS statistical software version 18.0 (SPSS, Inc., Chicago, IL, USA) and R version 3.6.3 (www.r-project.org, accessed on 1 March 2020).

Results
We included a total of 3402 patients in this analysis ( Figure 1): mean age 67.4 ± 11.3 years, 1980 (58.2%) male. The average follow-up duration was 25.7 ± 10.6 months or 7912 person-years, during which HF developed during follow-up in 218 patients (6.4%). The incidence rate of HF was 3.03 (2.64-3.46) per 100 person-years. The baseline characteristics of the study population with and without HF outcome are shown in Table 1. There were significant differences between the patients with and without HF in age, proportion of females, and various comorbidities.   . Echocardiography data were available in 834 patients (91.4%). Among the patients with HF who had echocardiographic data, the HF phenotypes at baseline were as follows: heart failure with reduced ejection fraction (HFrEF) 241 (28.9%), heart failure with mildly reduced ejection fraction (HFmrEF) 127 (15.2 %), and heart failure with preserved ejection fraction (HFpEF) 466 (55.9%).

Univariable and Multivariable Analysis
The univariable and multivariable analyses were performed to determine the factors predicting HF (Table 2). On multivariable analysis, independent predictors of HF requiring hospital admission or an urgent, unscheduled clinic/office/emergency department visit were as follows: old age, female, history of HF, history of coronary artery disease (CAD), cardiac implantable electronic devices (CIED), hypertension, diabetes, smoking, renal replacement therapy, and LVEF < 50%.

Model Validation
The C-index was 0.756 (95% CI: 0.723-0.789). The D-statistic was 1.532 (95% CI: 1.303-1.761), which indicated that the model can be used to discriminate the patients with a high risk or low risk of incident HF. When compared to the low-risk group according to the prediction model, the high-risk group had a Hazard Ratio of 6.90 (4.37-10.92).
The patients were categorized into five groups (very low, low, moderate, high, and very high risk) according to their quintile of predicted probability of HF derived from the score.
The agreement between the observed and predicted HF event and percentage of probability for HF is shown in Figure 2. The cumulative HF event rates over time in the patients belong to different risk groups are shown in Figure 3, which demonstrates well-separated Hazard graphs between each risk group. The calibration plots showed good agreement between predicted probability derived from the regression model and actual outcomes among 10 groups of patients with a calibration slope of 0.838 and the intercept of −0.002 ( Figure 4A). 0.379917×Hypertension + 0.541620×Smoking + 1.427287×Renal replaceme 0.569514×LVEF < 50%.

Model Validation
The C-index was 0.756 (95% CI: 0.723-0.789). The D-statistic was 1 1.303-1.761), which indicated that the model can be used to discriminate the a high risk or low risk of incident HF. When compared to the low-risk group the prediction model, the high-risk group had a Hazard Ratio of 6.90 (4.37- The patients were categorized into five groups (very low, low, moder very high risk) according to their quintile of predicted probability of HF der score.
The agreement between the observed and predicted HF event and probability for HF is shown in Figure 2. The cumulative HF event rates ov patients belong to different risk groups are shown in Figure 3, which demo separated Hazard graphs between each risk group. The calibration plots agreement between predicted probability derived from the regression mod outcomes among 10 groups of patients with a calibration slope of 0.838 and of −0.002 ( Figure 4A).    For internal validation, we performed 100 bootstraps. The C-and D-statistics after the bootstraps were 0.756 (0.737-0.775) and 1.503 (1.372-1.634). The Brier score was 0.056, which indicated that the model can predict the HF event very well.
The optimisms for the C-statistics, slope, and intercept of the calibration plot were calculated using the bootstrap method. The optimisms were 0.019, 0.045, and 0.105 for the C-statistics, slope, and intercept, respectively. After correcting the optimism, the Cstatistics, slope, and intercept of the calibration plot were 0.747 (0.745-0.750), 0.836, and −0.100, respectively.
The sensitivity analysis for the assessment of model prediction in patients who did not have a history of HF showed that the C-statistics remain high (0.700 95% CI 0.682-0.718). The R 2 of the calibration plot was 0.933 and the NRIs compared to the other three models were 10.07-29.75%. The sensitivity analysis was also performed on the complete dataset after the removal of the patients with missing LVEF data. The results showed that all the variables in the multivariable analysis model remained in the multivariable model of the complete dataset with the exception of female gender. There was additional analysis of the results of the COOL-AF prediction model with the exclusion of 1147 patients with a history of heart failure or LVEF < 50%. The results demonstrated that the C-statistics remain high (0.712, 95% CI 0.693-0.731), which is superior to the other three models that had C-statistics ranging from 0.630 to 0.690. The R 2 of the calibration plot was 0.803 and the COOL-AF model NRIs compared to the other three models were 4.49-26.90%. Schnabel et al. [16] and D. Imai et al. [17] For internal validation, we performed 100 bootstraps. The C-and D-statistics after the bootstraps were 0.756 (0.737-0.775) and 1.503 (1.372-1.634). The Brier score was 0.056, which indicated that the model can predict the HF event very well.
The optimisms for the C-statistics, slope, and intercept of the calibration plot were calculated using the bootstrap method. The optimisms were 0.019, 0.045, and 0.105 for the C-statistics, slope, and intercept, respectively. After correcting the optimism, the C-statistics, slope, and intercept of the calibration plot were 0.747 (0.745-0.750), 0.836, and −0.100, respectively.
The sensitivity analysis for the assessment of model prediction in patients who did not have a history of HF showed that the C-statistics remain high (0.700 95% CI 0.682-0.718). The R 2 of the calibration plot was 0.933 and the NRIs compared to the other three models were 10.07-29.75%. The sensitivity analysis was also performed on the complete dataset after the removal of the patients with missing LVEF data. The results showed that all the variables in the multivariable analysis model remained in the multivariable model of the complete dataset with the exception of female gender. There was additional analysis

Simplified HF Prediction Score
We developed a simplified HF prediction score using a coefficient of 10 variables in the prediction model. For a simplified HF score, Age > 65 years = 1; Female gender = 1; History of heart failure = 2; History of coronary artery disease = 1; Cardiac implantable electronic device = 1; Diabetes mellitus = 2; Hypertension = 1; Smoking = 1; Renal replacement therapy = 3; and LVEF < 50% = 1. The total score = 14. The patients in this cohort had scores varying from 0 to 11. The C-statistics of the simplified HF prediction score in the prediction of HF at 3 years was 0.751 (95% CI = 0.736-0.766), which is equivalent to the C-statistics of the original model that was 0.757 (0.742-0.771). The incidence rate and HF event rate at 3 years increased as the score increased. The calibration plot of the model using a simplified HF prediction score and observed probability and HF prediction model showed a good agreement (Supplementary Figure S1). Supplementary Table S1 and Supplementary Figure S2A show the incidence rate per 100 person-years and rate of HF at 3 years using a simplified HF prediction score.
We classified the HF risk according to the simplified HF score as low risk (score 0-2), intermediate risk (score 3-5), high risk (score 6-7), and very high risk (score ≥ 8). The incidence rate for HF and the 3 years HF event rate increased from low-risk to very high risk with the 3 years event rate at 2% in the low-risk group and 34.8% in very-high-risk group (Supplementary Figure S2B,C).

Comparisons of COOL-AF Predictive Model with Previous Studies
The predictive model of the COOL-AF study was compared to the models derived from the variables of the predictive models from the three previous studies as mentioned in the statistical section (Section 2.4). All the predictive models are shown in Supplementary Table S2. Supplementary Table S3 shows coefficients of variables in the models of COOL-AF and the other three studies. The calibration plot of the COOL-AF model was closer to the identity line when compared to the calibration plot of the previous studies ( Figure 4). The C-statistics of the COOL-AF model were significantly higher than the other models (p < 0.001) ( Figure 5). The results of the NRI and IDI analysis demonstrated a better reclassification of cases and controls using the COOL-AF model compared to the other three models (Supplementary Figure S3) with the NRI ranging from 31.8% to 59.5%. The decision curve analysis showed a better net number of true positives gained using the COOL-AF model compared to the other three models and no model at different threshold probabilities ( Figure 6). The 3-year HF risk prediction can be applied with mobile application (Supplementary Figure S4).  . Decision curve analysis (DCA) for the assessment of the ability to make better decisions with a model than without. Each graph represents DCA of each model compared to no models (treat all and treat none). COOL-AF model was compared to Pandey et al [15], Schnabel et al [16] and Imai et al [17].

Additional Analysis to Identify Preventive Strategies to Reduce HF Risk
We performed analysis to determine whether better symptom management, cardiovascular risk factors, and comorbidities management, which are the B and C components of the AF better care (ABC) pathway [12], and a good control of blood pressure and heart rate can reduce HF risk. The results showed that the incidence rate of patients with a B compliant, as defined by the European Heart Rhythm Association (EHRA), score ≤ 2 had an incidence rate of HF 2.43 (2.03-2.89) compared to those with B non-compliant [5.59 (4.48-6.88)], p < 0.001. Similarly, the patients with C-compliant, as defined by the appropriate management of cardiovascular risk factors and comorbidities, had a lower incidence rate of HF compared to those with C non-compliant  . Decision curve analysis (DCA) for the assessment of the ability to make better decisions with a model than without. Each graph represents DCA of each model compared to no models (treat all and treat none). COOL-AF model was compared to Pandey et al. [15], Schnabel et al. [16] and Imai et al. [17].

Additional Analysis to Identify Preventive Strategies to Reduce HF Risk
We performed analysis to determine whether better symptom management, cardiovascular risk factors, and comorbidities management, which are the B and C components of the AF better care (ABC) pathway [12], and a good control of blood pressure and heart rate can reduce HF risk. The results showed that the incidence rate of patients with a B compliant, as defined by the European Heart Rhythm Association (EHRA), score ≤ 2 had an incidence rate of HF 2.43 (2.03-2.89) compared to those with B non-compliant [5.59 (4.48-6.88)], p < 0.001. Similarly, the patients with C-compliant, as defined by the appropriate management of cardiovascular risk factors and comorbidities, had a lower incidence rate of HF compared to those with C non-compliant [

Discussion
The results of this prospective multicenter registry of patients with non-valvular AF in Thailand demonstrated that the incidence of HF was 3.03 (2.64-3.46) per 100 person-years. Second, the predictive factors of HF were age >65 years, female gender, history of HF, history of CAD, CIED, diabetes mellitus, hypertension, smoking, renal replacement therapy, and LVEF < 50%. Third, the model can be used to discriminate the patients with a high risk or low risk of incident HF, with good calibration after internal validation. After the data input, the HF risk can be easily calculated using a mobile application (Supplementary Figure S2) and for those who are at high or very high risk, preventive strategies can be initiated to reduce the risk of HF hospitalization.
The incidence rate of HF in patients with AF in our study was 3.03 per 100 person-years, which was higher than the results of the ORBIT-AF study, which is 1.8 per 100 person-years where the majority of patients were non-Asians [15]. In other studies, a HF incidence rate of 1.83 per 100 person-years has been reported from the ACTIVE, RE-LY, and AVERREOS trial populations [25]. The data on the HF phenotypes of our study also showed that, among patients with incident HFrEF, HfmrEF, and HfpEF, the highest rates were evident in HfpEF, which is in agreement with the prior studies showing the close AF and HFpEF [26].
The Asian population has been shown to have more HF burden than non-Asian populations [27]. Another explanation for the different HF incidence rates in the literature is that AF management may have been changed and many guidelines now focus also on the risk factor and comorbidity management in AF patients [13,28]. Previous studies demonstrated disparities in AF management especially with regard to the use of oral anticoagulant (OAC) and the quality of OAC control [29,30].
The COOL-AF HF prediction model, compared to the results of three previous studies, demonstrated that the C-statistics of the COOL-AF model were higher than other models and that the NRI was in the range from 31.8% to 59.5%. The calibration plots showed good agreement between the predicted probability derived from the regression model and actual outcomes among 10 groups of patients with a calibration slope of 0.  [16,17]. In our AF cohort, the important predictors of incident HF were >65 years, female gender, history of HF, history of CAD, CIED, diabetes mellitus, hypertension, smoking, renal replacement therapy, and low LVEF. These predictors in our study are broadly similar to previous reports [15,17,25,31,32]. Our study indicates that female gender is a predictor for HF, which differs from previous studies [33]. The data from a Japanese AF population demonstrated that females had a higher risk of HF than males and the incidence rate of incident HF in both sexes remained high [5]. In our study, the results of the sensitivity analysis after the removal of patients with a history of HF and those with LVEF < 50% showed that the C-statistics of the COOL-AF HF prediction model remained high and greater than the C-statistics from the other three models from previous studies [15][16][17].
The results of our study have some potential clinical applications. The simplified HF prediction score that we proposed on Supplementary Figure S2 can help identify those with a low risk of a 3-year HF event (2%) and a very high risk with an HF event rate of 34.8%. For the HF prediction model, if we applied the model as a mobile application, after the data input, we could calculate the absolute HF risk of each individual patient. For those who are at a high HF risk, physicians can apply some preventive strategies such as maintaining good blood pressure and heart rate control or cardiovascular risk factors and co-morbidity management that might reduce the HF risk.
We also identified some preventive strategies to reduce HF risk from the analysis of our data. These preventive strategies included better symptom management, better management of cardiovascular risk factors and comorbidities, which are the B and C components of the ABC pathway [12], and control of SBP within the target of 120-140 mmHg and heart rate < 100 beats/minute.

Limitations
This study has some limitations. First, the majority of the patients were enrolled from large hospitals. Therefore, the results may not be generalizable to the whole group of AF patients. Second, despite the reminder to all the investigators to enroll consecutive cases, there might be some possibility of selection bias. Third, the definition of heart failure outcome may be different among different studies. In this study, we tried to use the most recommended definition for HF outcome [18]. Fourth, for the comparison of the COOL-AF model with other models, there might be some limitations since the collected variables may be different among studies. Therefore, the confounding factors that depend on the baseline data may have some differences and might impact the results of the comparison despite the attempt to refit the previous models in the COOL-AF population. Fifth, despite internal validation being performed, it would be better if there was external validation.

Conclusions
The incidence rate of HF requiring hospital admission or an urgent, unscheduled clinic/office/emergency department visit in patients with AF was 3.03 per 100 personyears. We also provide a validated clinical HF predictive model for patients with AF, with good prediction and discrimination values.

Supplementary Materials:
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/jcm12041265/s1, Figure S1: Calibration plot of A. Predicted probability from Simplified HF prediction model versus Observed probability and B. Predicted probability from Simplified HF prediction model versus Predicted probability from Complete HF prediction model; Figure S2: A. Heart failure (HF) event rate at 3 years stratified by simplified HF risk score. B and C. Incidence rate of HF and HF event rate at 3 years according to risk groups classified by the simplified risk score. (Simplified HF score, Age > 65 years = 1; Female gender = 1; History of heart failure = 2; History of coronary artery disease = 1; Cardiac implantable electronic device = 1; Diabetes mellitus = 2; Hypertension = 1; Smoking = 1; Renal replacement therapy = 3; and Left ventricular ejection fraction < 50% = 1); Figure S3: Net reclassification index (NRI) and Integrated Discrimination Index (IDI) of COOL-AF predictive model compared to three previous models. A. 2 by 2 table showing comparison of NRI of COOL-AF model compared to the others and category-free NRI and IDI with 95% confidence interval. B. Scattered plot of predictive probability of COOL-AF model on Y-axis compared to previous model; Figure S4: Simulation of mobile application for clinical use of COOL-AF model. A. 3-year heart failure risk based on 10 variables derived from COOL-AF population B. Output of calculation and the risk levels C. Formula used for heart failure risk calculation when left ventricular ejection fraction (LVEF) data are available (Formula 1) and not available (Formula 2) D. Example of a case E. 3-year heart failure risk of this case with LVEF data F. 3-year heart failure risk of this case when LVEF data is not available; Table S1: Incidence rate and risk of heart failure (HF) at 3 years of patients with atrial fibrillation stratified by simplified HF prediction score; Table S2: Predictive model of COOL-AF and predictive models that were derived from 3 previous studies that were refitted in COOL-AF population; Table S3: Coefficients of variables in the models of COOL-AF and the other 3 studies Author Contributions: Conceptualization, R.K.; methodology, R.K.; formal analysis, C.K. and P.S.; investigation, P.C.; data curation, R.K.; writing-original draft preparation, R.K.; writing-review and editing, G.Y.H.L. and P.C.; supervision, R.K.; project administration, R.K.; funding acquisition, R.K. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement:
The dataset that was used to support the results and conclusion of this study are included within the manuscript. The additional data are available from corresponding author upon reasonable request.