Validation of EuroSCORE II in a modern cohort of patients undergoing cardiac surgery

Abstract

OBJECTIVE

We aimed to validate the new EuroSCORE II risk model in a contemporary cardiac surgery practice in the United Kingdom (UK).

METHODS

The original logistic EuroSCORE was compared to EuroSCORE II with regard to accuracy of predicting in-hospital mortality. Analysis was performed on isolated coronary artery bypass grafts (CABG; n = 2913), aortic valve replacement (AVR; n = 814), mitral valve surgery (MVR; n = 340), combined AVR and CABG (n = 517), aortic (n = 350) and miscellaneous procedures (n = 642), and the above cases combined (n = 5576).

RESULTS

In a single-institution experience, EuroSCORE II is a reasonable risk model for in-hospital mortality from isolated CABG (C-statistic 0.79, Hosmer-Lemeshow P = 0.052) and aortic procedures (C-statistic 0.81, Hosmer-Lemeshow P = 0.43), and excellent for mitral valve surgery (C-statistic 0.87, Hosmer-Lemeshow P = 0.6). EuroSCORE II is better than the original EuroSCORE, using contemporaneous data for combined AVR and CABG operations (C-statistic 0.74, Hosmer-Lemeshow P = 0.38). However, EuroSCORE II failed to improve on the original EuroSCORE model for isolated AVR (C-statistic 0.69, Hosmer-Lemeshow P = 0.07) and miscellaneous procedures (C-statistic 0.70, Hosmer-Lemeshow P = 0.99). EuroSCORE II has better calibration than the original EuroSCORE or the Society of Cardiothoracic Surgeons of Great Britain and Ireland (SCTS) modified EuroSCORE for cumulative sum survival (CUSUM) curves.

CONCLUSIONS

EuroSCORE II improves on the original logistic EuroSCORE, though mainly for combined AVR and CABG cases. Concerns still exist, however, over its use for isolated AVR procedures, aortic surgery and miscellaneous procedures. There is still room for improvement in risk modelling.

Risk modelling, EuroSCORE

INTRODUCTION

EuroSCORE, both additive and logistic, has become the standard by which risk assessment with regard to death is measured in the scientific literature [1, 2]. The recent update of the EuroSCORE assessment system, from version I to II, has been published based on 23 000 patients undergoing cardiac surgery in over 150 hospitals in 43 countries over a 12-week period (May to July 2010) [3]. The accuracy of risk prediction of a mortality model depends on the receiver operating characteristics and the predictive power of the model.

The original logistic EuroSCORE, despite its known limitations, is actually a reasonable in-hospital risk-predictive model [4–6]. A number of publications from various countries have, however, identified issues with the accuracy of the EuroSCORE system, hence the work on EuroSCORE II [7–9].

Our aim was to validate the EuroSCORE II risk model against the original logistic EuroSCORE in a contemporary group of patients, undergoing isolated coronary artery bypass surgery, aortic and mitral valve replacement, and combined aortic valve replacement and bypass surgery [3].

METHODS

Local institutional review board approval was granted for this study.

Database

A cardiac surgery database (n = 18 706) was utilized, validated by the hospital data analysis department and accredited by the Society of Cardiothoracic Surgeons of Great Britain and Ireland (SCTS). Due to the need for the estimation of the pre-operative glomerular filtration rate (GFR), only data from the last five years were included as, prior to this, the records of pre-operative creatinine levels were not in electronic format.

Patients

Patients who had undergone CABG, AVR, MVR, combined AVR and CABG, aortic and miscellaneous procedures (left ventricular aneurysmectomy, acquired VSD, atrial myxoma, pulmonary embolectomy, cardiac trauma, pericardiectomy, atrial septal closure (ASD), etc.) between January 2006 and March 2010 were included, totalling 5576 subjects. Our institution had contributed data to the EuroSCORE II project, but from May to July 2010 only, so no patients were incorporated in both studies. In-hospital mortality was defined as death at the institution where surgery was performed.

Benchmarking

We benchmarked our in-hospital mortality figures against the United Kingdom (UK) national results (http://www.scts.org).

Methodology

The characteristics of the patients in this study are shown in Table 1. EuroSCORE [1], SCTS-modified EuroSCORE [10] and EuroSCORE II [3] models were compared. Sensitivity and specificity was assessed by the use of the C-statistic, and the Hosmer–Lemeshow test was utilized to assess the predictive power of the risk models (Table 2). Analysis was performed across the whole group and by individual procedure.

Table 1:

Patient characteristics

Variable	All patients (n = 5576)	Isolated CABG (n = 2913)	Isolated AVR (n = 814)	Isolated MVR (n = 340)	AVR + CABG (n = 517)	Aortic (n = 350)	Miscellaneous (n = 642)
Preoperative
Age at operation (years)	69.3 (61.6, 75.2)	67.9 (60.7, 74.1)	70.7 (62.6, 78.0)	66.7 (57.2, 73.8)	73.8 (69.1, 78.3)	63.4 (50.0, 71.8)	69.2 (58.8, 75.4)
Body mass index (kg/m2)	28.1 (25.2, 31.5)	28.3 (25.6, 31.6)	27.9 (24.9, 31.9)	26.2 (23.3, 29.1)	28.0 (25.0, 31.6)	27.4 (24.6, 30.83)	26.5 (23.4, 30.2)
Female gender	26.1 (1195)	18.2 (529)	44.0 (358)	45.0 (153)	30.0 (155)	30.5 (107)	45.8 (294)
Angina class IV	12.6 (579)	18.4 (536)	1.0 (8)	0.0 (0)	6.8 (35)	1.1 (4)	4.2 (27)
Previous Q-wave myocardial	39.9 (1829)	55.5 (1616)	5.8 (47)	3.5 (12)	29.8 (154)	6.8 (24)	21.7 (139)
Q-wave MI within previous 30 days	18.8 (860)	26.9 (784)	1.0 (8)	0.0 (0)	13.2 (68)	0.9 (3)	7.3 (47)
NYHA class ≥ III	35.5 (1629)	24.9 (724)	54.3 (442)	60.6 (206)	49.7 (257)	27.1 (95)	60.1 (386)
Current smoker	12.4 (567)	14.3 (416)	8.5 (69)	8.5 (29)	10.3 (53)	11.4 (40)	14.2 (91)
Diabetes	22.1 (1014)	25.1 (730)	15.9 (129)	10.6 (36)	23.0 (119)	8.0 (28)	15.6 (100)
Hypercholesterolaemia	84.3 (3866)	94.5 (2753)	64.4 (524)	42.9 (146)	85.7 (443)	49.9 (175)	54.7 (351)
Hypertension	62.1 (2847)	66.5 (1937)	54.3 (442)	35.3 (120)	67.3 (348)	51.3 (180)	49.7 (319)
Cerebrovascular disease	10.0 (459)	9.2 (267)	12.2 (99)	10.0 (34)	11.4 (59)	7.7 (27)	11.5 (74)
Respiratory disease	29.5 (1351)	25.0 (729)	37.1 (302)	37.0 (126)	37.5 (194)	28.5 (100)	43.0 (276)
Peripheral vascular disease	15.6 (714)	17.8 (519)	10.2 (83)	6.5 (22)	17.4 (90)	10.0 (35)	10.4 (67)
Renal dysfunction	9.8 (449)	8.9 (260)	10.2 (83)	8.8 (30)	14.7 (76)	9.1 (32)	15.3 (98)
Left main stem disease	18.6 (851)	26.9 (783)	0.3 (2)	0.0 (0)	12.8 (66)	1.4 (5)	5.0 (32)
Triple-vessel disease	45.6 (2502)	78.5 (2288)	2.8 (23)	1.2 (4)	36.2 (187)	5.1 (18)	15.9 (102)
Urgent operation	28.3 (1295)	35.6 (1036)	14.6 (119)	9.1 (31)	21.1 (109)	23.7 (83)	26.6 (171)
Left internal mammary artery used	–	93.5 (2722)	–	–	64.2 (330)	58.8 (40)	65.0 (147)
Number of distal anastomosis	–	3 (3, 4)	–	–	2 (1, 3)	2 (1, 2.5)	2 (1, 3)
Logistic EuroSCORE	4.8 (2.3, 10.7)	3.5 (1.9, 7.7)	7.1 (3.7, 14.4)	6.8 (3.3, 13.2)	10.6 (5.2, 20.4)	15.6 (8.4, 26.7)	10.7 (5.2, 23.5)
Logistic EuroSCORE - SCTS-modified	2.2 (1.1, 4.8)	1.7 (0.9, 3.6)	2.8 (1.5, 5.4)	3.2 (1.7, 6.5)	5.2 (2.4, 10.0)	11.0 (5.6, 19.4)	6.7 (3.3, 14.5)
Logistic EuroSCORE II	2.0 (1.2, 4.0)	1.7 (1.1, 3.3)	2.1 (1.2, 3.9)	2.1 (1.3, 4.0)	4.5 (2.7, 9.0)	5.6 (3.1, 11.1)	5.6 (2.6, 10.1)
Intraoperative
Off pump procedure	23.9 (1096)	35.9 (1047)		–	–	2.9 (10)	11.7 (75)
CPB time	104 (85, 129)	101 (83, 125)	92 (74, 113)	104 (91, 125)	144 (116, 178)	250.5 (184, 340.5)	147 (106, 191)
AXC time	72 (58, 89)	65 (53, 79)	70 (56, 85)	77 (66, 91)	107 (86, 134)	174 (133, 222)	106 (75, 143)
Postoperative characteristics
Myocardial Infarction	0.4 (18)	0.6 (17)	0.1 (1)	0.0 (0)	0.0 (0)	0.6 (2)	0.3 (2)
Stroke	1.5 (70)	1.0 (30)	1.2 (10)	2.7 (9)	4.1 (21)	3.4 (12)	3.7 (24)
Acute renal failure	7.0 (319)	5.9 (172)	7.5 (61)	7.9 (27)	11.4 (59)	6.6 (23)	14.5 (93)
Re-operation for bleeding	4.6 (211)	3.6 (105)	5.8 (47)	3.5 (12)	9.1 (47)	8.6 (30)	7.6 (49)
Surgical wound infection	1.7 (77)	1.9 (54)	0.7 (6)	0.9 (3)	2.7 (14)	0.9 (3)	1.1 (7)
Inotrope delivery	33.0 (1512)	28.7 (835)	34.3 (279)	31.5 (107)	56.3 (291)	57.6 (202)	56.1 (360)
Harvest site infection	1.1 (49)	1.2 (35)	–	–	2.7 (14)	0.9 (3)	1.1 (7)
Atrial fibrillation	27.7 (1268)	25.9 (753)	27.3 (222)	28.5 (97)	37.9 (196)	24.4 (85)	22.6 (144)
Intubated >48 h	3.8 (174)	3.0 (87)	3.2 (26)	4.4 (15)	8.9 (46)	8.6 (30)	10.0 (64)
In-hospital mortality	2.2 (101)	1.9 (54)	2.3 (19)	1.5 (5)	4.5 (23)	6.8 (24)	10.3 (66)

Categorical variables shown as % (n); continuous variables shown as median (25^th percentile, 75^th percentile).

Table 2:

Mortality prediction statistics

Variable	All patients (n = 5576)	Isolated CABG (n = 2913)	Isolated AVR (n = 814)	Isolated MVR (n = 340)	AVR + CABG (n = 517)	Aortic (n = 350)	Miscellaneous (n = 642)
Logistic EuroSCORE (1999)
C-statistic (95% CI)	0.77 (0.73–0.80)	0.77 (0.70–0.83)	0.67 (0.56–0.78)	0.88 (0.73–0.99)	0.67 (0.56–0.78)	0.77 (0.67, 0.86)	0.68 (0.61, 0.76)
Hosmer-Lemeshow P-value	0.002	0.41	0.63	0.28	0.38	0.40	0.82
EuroSCORE SCTS update (2008)
C-statistic (95% CI)	0.78 (0.75–0.82)	0.78 (0.71–0.84)	0.67 (0.57–0.77)	0.90 (0.76–0.99)	0.67 (0.55–0.78)	0.78 (0.69, 0.86)	0.70 (0.63, 0.77)
Hosmer-Lemeshow P-value	<0.001	0.09	0.50	0.59	0.48	0.16	0.45
Logistic EuroSCORE II (2012)
C-statistic (95% CI)	0.79 (0.77–0.83)	0.79 (0.73–0.85)	0.69 (0.57–0.79)	0.87 (0.72–0.99)	0.74 (0.65–0.83)	0.81 (0.73, 0.89)	0.70 (0.63, 0.77)
Hosmer-Lemeshow P-value	<0.001	0.052	0.07	0.60	0.38	0.43	0.99

CUSUM curves were constructed for a visual analysis of the models.

Statistical software

All statistical analysis was performed with SAS for Windows, version 9.2 (SAS Institute, Cary NC, USA) and MedCalc for Windows, version 12.2.1 (MedCalc Software, Mariakerke, Belgium).

RESULTS

Benchmarking of our institutional mortality rates, compared to the UK as a whole, did not reveal any differences (part of the continuous UK cardiac surgery quality assessment program by the society of cardiothoracic surgeons, www.scts.org). The logistic EuroSCORE ranged from 0.8 to 87 (mean 6.2, 95% CI: 6.1 to 6.3). Institutional mortality was 2.5% for all isolated CABG, AVR, MVR and combined AVR and CABG cases. The decile breakdown of the Hosmer-Lemeshow statistic for observed and predicted mortality for the three models and all procedures, as analysed below, is documented in the data supplement.

All patients

The EuroSCORE II model resulted in a clinically insignificant improvement in C-statistic value: 0.79 vs 0.77, respectively, when compared to the original EuroSCORE (Table 2 and Fig. 1A). The Hosmer-Lemeshow test was, however, significant for the original and EuroSCORE II: P = 0.001 and P < 0.001, respectively, indicating poor predictive power. However, the CUSUM curves obtained were better calibrated with EuroSCORE II than the original EuroSCORE and the SCTS-modified EuroSCORE (Fig. 1B).

Figure 1:

All cardiac surgery: (A) C-statistic comparison; (B) CUSUM curves.

Isolated coronary artery bypass surgery

The EuroSCORE II model resulted in a clinically insignificant improvement in C-statistic values: 0.77 vs 0.76, respectively (Table 2 and Fig. 2A), P = 0.13. The Hosmer-Lemeshow test for EuroSCORE II was not significant (P = 0.052), indicating an acceptable predictive power: however the original EuroSCORE was better calibrated with a Hosmer-Lemeshow statistic of 0.41. The CUSUM curve, on the other hand, was better calibrated (Fig. 2B).

Figure 2:

Isolated CABG: (A) C-statistic comparison; (B) CUSUM curves.

Aortic valve replacement

The EuroSCORE II model resulted in no improvement in C-statistic values: 0.69 vs 0.67, respectively, compared to the original EuroSCORE (Table 2 and Fig. 3A), P = 0.68. A C-statistic of less than 0.7 is usually thought to be too inaccurate to utilize as a predictive model [11]. The Hosmer-Lemeshow test for EuroSCORE II was not significant (P = 0.07) indicating acceptable predictive power, which was reflected in the CUSUM curves, which were better calibrated than the original EuroSCORE (Fig. 2c).

Figure 3:

AVR: (A) C-statistic comparison; (B) CUSUM curves.

Mitral valve surgery

The EuroSCORE II model resulted in no clinical improvement in C–statistic values: 0.87 vs 0.90, respectively, compared to the original EuroSCORE (Table 2 and Fig. 4A), P = 0.85. The Hosmer-Lemeshow test was not significant (P = 0.6) indicating adequate predictive power and an improvement over the original EuroSCORE: Hosmer-Lemeshow statistic 0.28. The CUSUM curve was also better calibrated with EuroSCORE II than with the original EuroSCORE and the SCTS-modified EuroSCORE (Fig. 4B).

Figure 4:

MVR: (A) C-statistic comparison; (B) CUSUM curves.

Combined aortic valve and coronary artery bypass surgery

The EuroSCORE II model resulted in a significant improvement, P = 0.01, in the C-statistic: 0.74 vs 0.67, respectively, compared to the regional EuroSCORE (Table 2 and Fig. 5A). The Hosmer-Lemeshow test was not significant (P = 0.38), indicating adequate predictive power, which was similar to the original EuroSCORE. The C-statistic, being above 0.7, now indicates that the EuroSCORE II model is potentially valid for this subgroup of patients, unlike the original model. The CUSUM curve was also better calibrated than the original EuroSCORE and the SCTS-modified EuroSCORE (Fig. 5B).

Figure 5:

AVR and CABG: (A) C-statistic comparison; (B) CUSUM curves.

Aortic surgery

The EuroSCORE II model improved the C-statistic from 0.77 to 0.81, respectively, as compared to the original EuroSCORE (data supplement). The Hosmer-Lemeshow test was not significant (P = 0.43), indicating adequate predictive power and an improvement over the original EuroSCORE: Hosmer-Lemeshow statistic 0.40 (Table 2). The CUSUM curve was also better calibrated, compared to the original EuroSCORE and the SCTS-modified EuroSCORE (data supplement).

Miscellaneous procedures

The EuroSCORE II model resulted in a marginally improved C-statistic, 0.70 vs 0.68, respectively, compared to the original EuroSCORE (data supplement). The Hosmer-Lemeshow test was not significant (P = 0.99), indicating adequate predictive power, but was not an improvement over the original EuroSCORE: Hosmer-Lemeshow statistic 0.82 (Table 2). The CUSUM curve was also better calibrated than the original EuroSCORE and the SCTS-modified EuroSCORE (data supplement).

DISCUSSION

EuroSCORE II offers benefit over the original logistic EuroSCORE with regard to risk prediction in combined AVR and CABG in a large regional unit in the UK. However, concerns still exist over its use for isolated AVR procedures. The original EuroSCORE may actually outperform EuroSCORE II for isolated CABG patients. A highly significant Hosmer-Lemeshow statistic for the whole study group indicates that the EuroSCORE (either I or II) is not well calibrated and should be used with caution.

The strength of our analysis is that our series is recent, consecutive, large and that our database has been independently validated. In addition, as our institution has results that are not significantly different from those contained in the Great Britain and Ireland Sixth National Adult Cardiac Surgical Database Report, 2008, we feel our data is robust and applicable to other units in the UK [12].

With over 43 countries and 150 institutions being involved in the creation of EuroSCORE II, natural variations in practice and outcomes due to local demography undoubtedly exist [3]. The finding that a risk model, generated using international data, does not accurately predict a regional institutional performance is not entirely surprising. In the UK, the risk from cardiac surgery for all cases varies, between institutions, from 1.6% to 5.1%.

Predictive models can be quantified by the use of C-statistic, Hosmer-Lemeshow test and CUSUM plots. The C-statistic is the area under a receiver operating characteristic (ROC), or ROC curve, and represents the probability that predicting the outcome is better than chance alone. It can be used to compare the ‘goodness of fit’ of logistic regression models. A value of 0.5 indicates that the model is no better than chance at predicting the outcome, while a value of 1.0 indicates that the model perfectly identifies the outcome. Models are typically considered reasonable when the C-statistic is higher than 0.7 and strong when C exceeds 0.8 [13]. The Hosmer–Lemeshow assesses goodness of fit for logistic regression models, and whether or not the observed event rates match expectations in deciles of fitted risk values of the model population. Models that have similar expected- and observed event rates in the subgroups are called ‘well calibrated’ and have a non-significant Hosmer-Lemeshow statistic. CUSUM curves are used as a quality control method to detect deviations from benchmark values. With regard to risk model assessment, they are useful to visually compare observed vs expected mortality over time.

The C-statistic and CUSUM curves can be misleading if the data is skewed with regard to risk profile. The Hosmer-Lemeshow test assesses model performance based on risk, so is able to detect deciles of the risk that model predictive power may be suboptimal, despite a good C-statistic and CUSUM curve.

When the whole study group is analysed, EuroSCORE II has a C-statistic that is good, but the Hosmer-Lemeshow statistic is very significant, indicating a very poor fit. This could have negative consequences for surgeons who operate on patients with certain risk profiles. This is virtually identical to results with the original EuroSCORE and could be predicted, based on the results for the models with regard to isolated CABG and AVR. The original EuroSCORE II manuscript reports a Hosmer-Lemeshow statistic of 0.0505, indicting a fit that is only just acceptable [14].

Isolated CABG is the commonest operation performed in adult cardiac surgery. A C-statistic of 0.79 for EuroSCORE II compares to the 0.81 of the STS risk model [15]. However the Hosmer-Lemeshow statistic demonstrates that the original EuroSCORE has a better fit than EuroSCORE II, raising concerns over its replacement.

A C-statistic below 0.7 for isolated AVR is troubling [11]. The STS risk model has a C-statistic of 0.76; however this is for an American population and may not be directly applicable to a modern European population [16]. In combination with a Hosmer-Lemeshow statistic that is close to significance, EuroSCORE II is actually inferior to the original EuroSCORE model, where the Hosmer-Lemeshow statistic was highly insignificant.

EuroSCORE II improves slightly on the original EuroSCORE model for isolated MVR. EuroSCORE II does not differentiate between mitral repair and replacement: the merits of this stratification have already been published [16]. The STS isolated valve model distinguishes repair from replacement, as mortality is nearly four times higher in those having a replacement [16]. However, the lack of consensus in preoperatively delineating who should have a repair and who a replacement, potentially allows gaming to occur.

The original EuroSCORE model was poor in cases of combined AVR and CABG. EuroSCORE II improves on this and now matches the STS model [17].

No widely accepted model exists for risk prediction in aortic surgery. EuroSCORE II has improved risk prediction with regard to aortic surgery but we remain concerned that an aortic model that does not include the actual aortic procedure (ascending aortic replacement vs thoracoabdominal replacement) as a risk factor is liable to error in predicting mortality.

A C-statistic of 0.7 is worrying for miscellaneous procedures and probably represents the heterogeneous nature of these procedures. We would recommend caution in its use, despite a non-significant Hosmer-Lemeshow statistic.

The STS development of separate models for isolated CABG, isolated valve and combined valve and grafts may explain the higher C-statistic associated with their models, and the potential weakness of the ‘one model fits all’ concept.

The inability of the EuroSCORE risk model—original and II—to be able to predict in-hospital mortality with 100% accuracy indicates the limitations of any modelling process but creates a number of potential avenues for improvement. Demographic, institutional and individual variations in practice may contribute to the divergence of observed vs expected [18]. To date the emphasis has been on the individual surgeon variation, which is known to be significant. Risk factors that are known to affect outcomes are not included in the model itself, for example coronary artery and ascending aortic disease profile. Concomitant medical conditions that are uncommon but affect mobility are now included in EuroSCORE II under the factor ‘poor mobility’. However, medical conditions that do not affect mobility are not included and, due to the large number of conditions that have variable manifestations and significance, their exclusion will continue unless ‘other significant medical conditions’ becomes a risk factor. However, this may result in the scientific validity of any model being destroyed secondary to gaming.

The updating from EuroSCORE to EuroSCORE II implies an inherent belief that logistic modelling is the best technique for mortality prediction in cardiac surgery. The use of neuronal networks [19], Bayesian techniques [20, 21] and structured query language matching [22] may, however, offer a better modelling technique, due to the non-linear interaction of the known risk factors. Logistic EuroSCORE assumes that a risk factor has the same weight, regardless of the other risk factors or operative intervention. For instance, our institutional database indicates that dialysis is a risk factor for in-hospital death, with an odds ratio of 7.4 for CABG and 1.8 for AVR with CABG, which may partially explain the improved performance of the separate STS models.

We have used the SCTS-modified EuroSCORE purely for reference, as this is the model utilized in the UK. The C-statistic, Hosmer-Lemeshow and the CUSUM curves represent different methods of model assessment. The demonstration of the Hosmer-Lemeshow test being significant—but of the CUSUM curves being better calibrated for all patients—indicates that the model could be further improved in certain risk deciles.

Despite being a very large institution, we did not have sufficient patients to evaluate the numerous other combinations and permutations of possible cardiac and aortic surgery interventions. We did not include these cases, so as to avoid drawing conclusions from a statistically underpowered analysis.

The time period of accrual of data for the development of EuroSCORE II may have introduced an error in their modelling, as it has been previously demonstrated that there is a seasonal variation in mortality due to cardiac surgery [23]. Accounting for this seasonal variation may help to improve the accuracy of future developments in risk modelling.

CONCLUSIONS

EuroSCORE II improves on the original EuroSCORE model, with regard to in-hospital mortality in a large regional institutional practice, for combined aortic valve replacement and CABG. The Hosmer–Lemeshow test for isolated CABG, however, is not as good as the original EuroSCORE model, raising concerns over its potential use. Room for model improvement still exists, particularly with regard to isolated aortic valve replacement surgery, aortic surgery and miscellaneous procedures. Further independent validation is needed, prior to the universal acceptance of EuroSCORE II.

LIMITATIONS

EuroSCORE II has introduced a new variable—poor mobility—as a risk factor. This is not included on our database and is not part of the national dataset in cardiac surgery in the UK; all patients were analysed as not scoring for this risk factor. The number of patients in our institution with poor mobility—which is difficult to define exactly—that would affect our analysis is very low. A large number of patients did not have pulmonary artery pressures measured preoperatively. For purposes of analysis, they were assumed to be normal and so did not score. Analysis of a single institution's results has limitations and may not represent national and international practice and outcomes, and thus may represent an important source of potential bias.