An Evaluation of Four Predictive Scores for New Onset Atrial Fibrillation After Cardiac Surgery in a Tertiary Cardiac Centre

ABSTRACT

Background

Postoperative atrial fibrillation (POAF) is a common complication after cardiac surgery, leading to increased mortality, morbidity, and length of stay. Various predictive indices have been validated to identify patients at risk of developing POAF. This study aimed to compare the discriminative ability of four bedside indices that predict POAF in cardiac surgical patients.

Method

A total of 2465 consecutive patients who underwent cardiac surgery at an Australian tertiary hospital between 29 December 2015, and 31 December 2022, were retrospectively analysed. Exclusion criteria included pre‐existing atrial fibrillation and transcatheter interventions. The area under the receiver‐operating characteristic curve (AUC‐ROC) analysis and Hosmer–Lemeshow calibration were performed to compare discriminative abilities. A logistic regression model was constructed to evaluate independent risk factors for developing POAF.

Results

Older patients (aged > 69) were more likely to develop POAF compared to younger age groups (p < 0.001). The Atrial Fibrillation Risk Index, CHA₂DS₂‐VASC, HATCH and POAF scores had weak discrimination and demonstrated poor predictive ability in this cohort (AUC‐ROC 0.49, 0.49, 0.50, 0.52, respectively).

Conclusion

Although previously validated in various cohorts, the four bedside predictive indices demonstrated poor discriminative ability and limited generalizability to our tertiary‐hospital cohort.

Our paper shows that the four bedside predictive indices for POAF are not generalisable to our single‐centre tertiary cohort. Care should be taken in the utilisation of these scores in the clinical setting, as they may carry limited value in predicting POAF in patients undergoing cardiac surgery.

1. Introduction

Postoperative atrial fibrillation (POAF) is the most common arrhythmia after cardiac surgery, occurring in up to 50% of patients [1]. Despite efforts in perioperative management of atrial fibrillation (AF), the incidence continues to rise with the ageing population [2]. While POAF is usually thought to be benign and self‐limiting, its risks include higher risk of mortality, renal insufficiency, systemic embolisation and increased length of stay [3, 4, 5]. Predictive tools have been developed to better stratify patients who could benefit from longer postoperative surveillance or targeted prophylaxis. Prophylaxis with antiarrhythmic agents such as amiodarone and sotalol for patients after cardiac surgery have previously been proposed [6, 7, 8]. However, routine prophylaxis could expose up to 80% of patients to adverse effects of antiarrhythmic medications for no clear benefit–risk advantage and result in increased drug costs [6, 7, 8]. Many of these tools have important limitations, hindering widespread use in clinical practice. For example, previous studies have used postoperative factors such as prolonged ventilation as a predictor, limiting use of the tool to the postoperative period [9, 10, 11].

Many existing models were developed in isolated CABG cohorts, limiting applicability to other cohorts such as valve surgery, after which POAF most commonly occurs [11, 12, 13]. Of the predictive models, the preferred tools should be easily calculated at the bedside, externally validated, and of acceptable discriminative ability. The CHA₂DS₂‐VASc score, HATCH score, POAF score, and atrial fibrillation risk index are validated indices, demonstrating moderate predictive ability [1]. The well‐known CHA₂DS₂‐VASc Score is routinely used to calculate thromboembolism risk for patients with atrial fibrillation and guide clinical decision‐making for anticoagulation [14]. It has also been shown to be strongly predictive of new‐onset AF after valve surgery [14]. The HATCH Score was initially developed to predict progression of paroxysmal to persistent AF after 1 year, but has also been validated for prediction of AF after cardiac surgery [1, 15, 16]. The POAF Score was the only score specifically developed to assess the risk of POAF after cardiac surgery, including valve surgery, using a large multicentre European cohort of 17 262 patients [17]. The atrial fibrillation risk index uses four clinical variables that have been validated in both CABG and valve surgery cohorts [18, 19].

Our aim is to firstly identify predictors of POAF in a single tertiary cardiac centre and secondly to compare the clinical utility of four predictive indices in patients who underwent cardiac surgery to help identify those most at risk of developing POAF.

2. Methods

This single‐institutional observational cohort study was approved by the Gold Coast Hospital and Health Service Human Research and Ethics Committee (Reference number: HREC/2022/QGC/90325). Data were collected retrospectively from the Gold Coast University Hospital Cardiothoracic Surgery Database. All consecutive patients who underwent cardiac surgery between 29 December 2015, and 31 December 2022, were included. Patients with preoperative AF of any subtype were excluded. Transcatheter valve interventions were also excluded.

The Gold Coast University Hospital Cardiothoracic Surgery Database corresponds to the Australian and New Zealand Society of Cardiac and Thoracic Surgeons (ANZSCTS) Cardiac Surgery Database, to maintain consistency of data definitions [20]. New‐onset arrhythmia was defined as a new onset of AF that required treatment, which specifically excludes recurrence of AF present preoperatively. This was captured on telemetry or ECG monitoring in the postoperative period and confirmed by a Consultant Cardiothoracic Surgeon or Cardiologist. The period of telemetry or ECG monitoring was variable, as it was a clinical decision determined by the Cardiothoracic Surgeon, with common practice of keeping telemetry for up to 4 days postoperatively, or longer if clinically indicated.

Four predictive indices were compared, including CHA₂DS₂‐VASc Score [14], HATCH Score [16], POAF Score [17], and the Atrial Fibrillation Risk Index [19]. The estimated glomerular filtration rate (eGFR) was calculated using the CKD‐EPI Creatinine Equation [21].

The CHA₂DS₂‐VASc Score was calculated by assigning points based on the following criteria: one point for congestive heart failure or left ventricular dysfunction, one point for hypertension, two points for age greater than or equal to 75 years, one point for diabetes, two points for a history of stroke, transient ischemic attack (TIA), or thromboembolism, 1 point for vascular disease, one point for ages between 65 and 74, and one point for female sex. The maximum score possible is nine points [22].

The POAF score assigns one point for patients aged 60–69, two points for those aged 70–79, and three points for individuals older than 80. Additionally, a single point was allocated for each of the following conditions: chronic obstructive pulmonary disease, eGFR less than 15 mL/min, emergency surgery, perioperative intra‐aortic balloon pump use, left ventricular ejection fraction less than 30%, and valve surgery [17].

The atrial fibrillation risk index was calculated by the following:

• One point if the individual is male and aged over 60 or female and over 66 years.

• One point if the individual is male and weighs over 76 kg or female and weighs > 64 kg.

• One point if the individual is male and more than 176 cm tall or female and over 168 cm tall.

• One point if the individual has peripheral vascular disease [19].

The HATCH score, one point is assigned for hypertension, one point for patients over 75 years, two points for those with a history of stroke or TIA, one point for chronic obstructive pulmonary disease, and two points for heart failure.

3. Statistical Analysis

Data analysis was conducted with Stata 16.1 (College Station, Texas, USA). Demographic variables were summarised with descriptive statistics. Categorical data was evaluated using the chi‐squared test, while continuous data was assessed with the Student's t test. The four scores were calculated for each participant. The AUC‐ROC assessed the sensitivity, specificity and the diagnostic performance of each of the four indices for prediction of POAF. A logistic regression model was constructed, with goodness‐of‐fit assessed by the Hosmer–Lemeshow test, as performed in previous validation studies [1, 23]. A univariate analysis was first performed for each variable independently to assess their significance in predicting the outcome. Variables found to have a p < 0.1 were included in the refined model, as previously described in the Hosmer–Lemeshow methodology [23]. Previously excluded covariates are then reintroduced individually to assess their influence on model fit, employing the Akaike information criterion (AIC) to enhance model reliability. The AIC measures model quality and estimates the model's prediction error and likelihood, given the data [24]. The linearity of continuous variables was assessed, and the variables were included in the model, with log transformation as required. A p‐value of less than 0.05 was considered statistically significant.

4. Results

Baseline demographic and operative details are presented in Table 1. Of the 2465 patients included in the analysis, 625 (25%) developed POAF. Among the variables examined in the univariate analysis, those who developed POAF were older (p < 0.001), had longer cardiopulmonary bypass duration (p = 0.03) and a significantly longer length of hospital stay (p = 0.02). Smoking status, body mass index (BMI), hypertension, peripheral vascular disease, left ventricular ejection fraction, and urgency of surgery were found to be similar between those who developed POAF and those who did not develop POAF. In patients who developed POAF, the median hospital length of stay was significantly longer than those without POAF (12 vs. 10 days, p = 0.02). The median postoperative length of stay was also significantly longer in patients who developed POAF (p < 0.001), likely representing an association with a more complex postoperative course rather than being a predictive factor.

TABLE 1.

Baseline demographics of study cohort from an Australian tertiary cardiac centre with univariate p values.

Demographic/clinical	With POAF (n = 625)	Without POAF (n = 1840)	p
Age (mean), n (95% CI)	69 (68–70)	65 (64–66)	< 0.001
Age category, n (%)
≤ 69 years	297 (48)	1099 (60)	< 0.001
70–79 years	256 (41)	599 (33)
≥ 80 years	72 (11)	142 (8)
Indigenous, n (%)	44 (2)	15 (2)	0.9
Male, n (%)	493 (79)	1402 (76)	0.17
Smoker, n (%)
Current smoker	80 (23)	301 (28)	0.1
Current or former smoker	355 (57)	1064 (58)	0.8
Weight, kg (95% CI)	84 (83–85)	85 (84–86)	0.4
Height, cm (95% CI)	172 (171–173)	171 (171–172)	0.2
BMI, kg/m2 (95% CI)	28.4 (28–29)	28.8 (29–29)	0.2
Diabetes, n (%)	115 (18)	402 (22)	0.07
Hypertension, n (%)	414 (66)	1176 (64)	0.3
Peripheral vascular disease, n (%)	40 (6)	127 (7)	0.7
COPD, n (%)	79 (13)	225 (12)	0.8
Cerebrovascular disease, n (%)	50 (8)	176 (10)	0.2
Left ventricular ejection fraction, n (%)
Normal > 60%	248 (44)	644 (38)	0.08
Mild 46%–60%	247 (43)	823 (48)
Moderate 30%–45%	60 (11)	205 (12)
Severe < 30%	14 (2)	38 (2)
Congestive heart failure, n (%)	85 (14)	227 (12)	0.4
Previous intra‐aortic balloon pump, n (%)	5 (1)	12 (1)	0.2
Prior open cardiac surgery, n (%)	13 (2)	96 (5)	0.002
Urgent surgery, n (%)	202 (32)	656 (36)	0.5
Type of surgery, n (%)			0.047
Total valve surgery	286 (28)	711 (71)
Isolated valve surgery	216 (28)	551 (72)
Aortic valve surgery	210 (30)	485 (69)
Mitral valve surgery	81 (26)	229 (73)
Total CABG	380 (24)	1293 (75)
Isolated CABG	310 (23)	1043 (77)
Combined CABG and valve surgery	70 (30)	160 (69)
Other cardiac surgery	29 (25)	86 (74)
Cardiopulmonary bypass time, mins (range)	113 (109–118)	108 (105–110)	0.03
Median total hospital length of stay, median (IQR)	12 (8, 20)	10 (7, 17)	0.02
Median postoperative length of stay, median (IQR)	9 (7, 14)	8 (6, 11)	< 0.001

Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmonary disease.

From the logistic regression model, when all the other variables are kept stable, for every 1‐year increase in age, patients are 1.02 (SE = 0.01) times more likely to have an incidence of AF (p < 0.001). Prior surgery was also found to have a significant influence on the development of POAF (OR 0.29, p = 0.002). From the logistic regression model, this suggests that those who are having redo surgery actually have a lower odds of developing POAF. The model is significant, with the log‐likelihood chi‐square statistic of 40.88 (p = 0.003) (Table 2).

TABLE 2.

Final logistic regression multivariate analysis for predictors of postoperative atrial fibrillation in cardiac surgical patients.

Logistic regression model	OR (SE)	p	95% CI
Age at procedure	1.02 (0.006)	0.000	1.01–1.04
Diabetes	0.66 (0.15)	0.062	0.43–1.02
BMI	0.98 (0.01)	0.070	0.95–1.00
Prior cardiac surgery	0.29 (0.10)	0.002	0.14–0.63
Urgent surgery	0.70 (0.14)	0.074	0.48–1.03
Overall model/baseline odds	0.19 (0.11)	0.003	0.06–0.58

Abbreviations: 95% CI, 95% confidence interval; AIC, 1120.68; OR, odds ratio; SE, standard error.

There were no statistically significant differences between those who had POAF compared to those who did not have POAF in the CHA₂DS₂‐VASc score (chi‐squared 7.1, df = 7, p = 0.4), atrial fibrillation risk index (chi‐squared 6.1, df = 4, p = 0.2) or HATCH score (chi‐squared 9.2, df = 7, p = 0.524). Patients who developed POAF tended to have higher POAF scores and those who did not develop POAF tended to have lower POAF scores (chi‐squared 15.3, df = 5, p = 0.009), which is also reflected in a comparatively higher AUC‐ROC of 0.571. Overall, the four scoring systems performed similarly in terms of their predictive ability for POAF in the AUC‐ROC analysis (Figure 1).

FIGURE 1.

AUC‐ROC curves comparing the discriminative abilities of the CHA₂DS₂‐VASC, POAF, HATCH scores, and the atrial fibrillation index scores.

5. Discussion

Our study is one of the first in the literature to compare the discriminative ability of four validated bedside indices that predict POAF in cardiac surgical patients. Previously validated indices have demonstrated inconsistent predictive values, amongst important clinical limitations. We found that the atrial fibrillation risk index, CHA₂DS₂‐VASc score, HATCH score, and POAF score all had weak discrimination and demonstrated poor predictive ability in our single‐centre cohort.

The multivariate analysis showed the patients who are older in age (p < 0.001) were more likely to develop POAF (Table 2). Age as a prognostic indicator has been well recognised in the literature [1, 17]. However, previous cardiac surgery has seldom been included in other studies investigating predictive indices for POAF. Of the larger studies in this area, Mariscalco et al. [17] did not find previous cardiac surgery a significant variable in the univariate or multivariate analyses. Surprisingly, the logistic regression suggested that redo surgery was actually a protective factor, which we postulate is due to selection bias, as in the era of an increasing number of transcatheter interventions being performed, patients undergoing redo surgery tend to be either better surgical candidates with fewer comorbidities or those who have no transcatheter options.

In our cohort, 25% of patients developed POAF, which is consistent with the widely reported incidence ranging between 15% and 40% [25, 26]. Our findings are consistent with a previous validation study, which demonstrated that the Atrial fibrillation risk index, HATCH score and POAF score failed external validation in a cohort of 1416 patients, suggesting that these three scores have limited value in identifying patients at risk of POAF [18]. Importantly, the study was limited by its short 2‐week follow up period as well as a significantly higher incidence of POAF than reported in the literature. Conversely, a single‐centre validation study including 3113 patients found that the POAF, HATCH and CHA₂DS₂‐VASc scores showed good calibration and discrimination, of which the CHA₂DS₂‐VASc Score demonstrated the best discriminative ability [1]. Due to the risk of measurement bias, these results should be interpreted with caution as not all patients had the same duration of monitoring or follow‐up, with more severely ill patients having longer monitoring time. While the CHA₂DS₂‐VASc score has been externally validated for the prediction of POAF after cardiac surgery, it demonstrated similar performance to the other three indices, when applied to our single‐centre Australian cohort [1, 26].

The atrial fibrillation risk index was retrospectively developed and validated in a CABG‐only cohort and demonstrated the worst discrimination of the three scores, in keeping with the results of this study (AUC‐ROC 0.508) [19]. The HATCH score was previously evaluated by two groups, both of which showed good discriminative ability (AUC‐ROC 0.70¹ and 0.77 [16]). This was significantly different from our findings (AUC‐ROC 0.53). This may be accounted for by differences in baseline demographics, as the proportion of patients undergoing valve surgery is almost double compared to Burgos et al. (2018), while Emren et al. (2016) studied a CABG‐only cohort [1, 16]. Variability in the literature between data definitions and cohort exclusion also limits the generalisability of these results. Of the four predictive indices compared in this study, POAF Score showed the greatest discriminative ability (AUC‐ROC 0.57), although lower than the derivation cohort (AUC‐ROC 0.65) [17]. A potential contributor to the difference in findings is the definition of POAF. The definition of POAF used in the POAF Score derivation study was ECG or telemetry evidence of AF, while the definition used in this study is new POAF requiring treatment, which is a more clinically relevant outcome.

More recently, a single‐centre study developed a nomogram in a cohort of 422 valve surgery patients to predict the risk of POAF [27]. Another large‐scale study constructed a prognostic model as part of the PARADISE project to identify patients at higher risk of POAF and adverse outcomes who could benefit from longer‐term monitoring, risk factor management, and thromboprophylaxis [28]. These studies remain to be externally validated. Given the variability in sample size, baseline demographics, inclusion criteria, and definitions across the literature, future meta‐analyses or larger scale, prospective studies with more consistent methodology would help to refine the utility of these predictive indices in the clinical setting.

The main limitations of this study are its retrospective and single‐centre nature, which may predispose to selection bias. The period of telemetry monitoring was not standardised and variably documented, and often depended on both clinical indication and resource management, which can account for some disparity between reported outcomes in the literature. The database, while offering consistent data definitions across national cardiac centres across the public and private sector, has the limitations of any database or registry, such as multiple data managers, need for quality control and coding bias. However, the consecutive sampling of the cohort spanning 7 years helps to minimise selection bias and enhances generalisability of the results.

6. Conclusion

Although previously validated in various cohorts, the four bedside predictive indices are not generalisable to our single‐centre tertiary‐centre cohort.

Conflicts of Interest

The authors declare no conflicts of interest.

Chang S., van der Merwe M., Zaka A., Gupta A., He C., and Stroebel A., “An Evaluation of Four Predictive Scores for New Onset Atrial Fibrillation After Cardiac Surgery in a Tertiary Cardiac Centre,” ANZ Journal of Surgery 95, no. 7‐8 (2025): 1394–1399, 10.1111/ans.70229.