CLINICAL RESEARCH
European Heart Journal (2011) 32, 2016–2026 doi:10.1093/eurheartj/ehp085
Endocarditis
Development and validation of a time-dependent risk model for predicting mortality in infective endocarditis Raymond W. Sy1,2, Chirapan Chawantanpipat 2, David R. Richmond 2, and Leonard Kritharides 1* 1 Department of Cardiology, 3rd floor West, Concord Repatriation General Hospital, Sydney South Western Area Health Service, University of Sydney, Hospital Rd, Concord, NSW 2139, Australia; and 2Department of Cardiology, Royal Prince Alfred Hospital, Sydney South Western Area Health Service, University of Sydney, Sydney, NSW, Australia
Aims
Existing risk models in infective endocarditis (IE) have not investigated whether the prognostic value of clinical parameters is time-dependent. We have explored the potential of time-dependent risk stratification to predict outcome in IE. ..................................................................................................................................................................................... Methods We studied 273 patients admitted with IE to two centres (derivation cohort n ¼ 192, validation cohort n ¼ 81). The and results derivation cohort was used to identify independent predictors of 6 months mortality at days 1, 8, and 15 (multivariable Cox regression, P , 0.05). There were six predictors at day 1, five at day 8, and only three at day 15. Whereas heart failure, thrombocytopenia, and severe comorbidity predicted mortality at all three time-points, other predictors were time-dependent (age, tachycardia, renal impairment at day 1; severe embolic events, renal impairment at day 8). These predictors were incorporated into a time-dependent model. The model was validated in an independent cohort with concordance indices of 0.79 (95% CI 0.68–0.91) at day 1, 0.79 (95% CI 0.65–0.93) at day 8, and 0.84 (95% CI 0.73–0.95) at day 15. Six months mortality was 2.4% in patients deemed as low-risk at all timepoints, compared with 78.2% in patients classified as high-risk at any evaluation. ..................................................................................................................................................................................... Conclusion Prognostic factors in IE are time-dependent. Time-dependent risk stratification accurately predicts outcome in IE.
----------------------------------------------------------------------------------------------------------------------------------------------------------Keywords
Endocarditis † Risk factors † Mortality † Prognosis
Introduction Infective endocarditis (IE) is a disease with wide variations in its clinical course and prognosis.1 Current management recommendations vary significantly according to the perceived patient risk. Early hospital discharge and outpatient monitoring is recommended for stable patients at low risk of mortality, whereas intensive strategies including surgery may be considered for unstable patients at the highest risk of mortality.2 – 4 Accurate and timely risk assessment of patients is mandatory for such decisions to be made effectively. Several parameters including heart failure are associated with increased mortality in IE.5 – 11 However, published studies differ with respect to the reported prognostic value of individual
parameters and this inconsistency may relate to the timing of evaluation. For example, one study demonstrated that increased levels of C-reactive protein after 1 week of treatment were associated with adverse outcomes whereas increased levels at baseline or after 2 weeks were not.10 The possibility that the prognostic value of clinical parameters may be time-dependent has not been systematically explored, either in the context of individual parameters or as components of an integrated risk model.6 The aims of the present study were to identify clinical predictors of mortality in IE, to determine whether these predictors vary over time, and to develop and validate a risk model for IE that allows patient prognosis to be accurately determined at different time-points during early hospitalization.
* Corresponding author. Tel: þ612 9767 6296, Fax: þ612 9767 6994, Email:
[email protected] Published on behalf of the European Society of Cardiology. All rights reserved. & The Author 2009. For permissions please email:
[email protected].
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Received 27 May 2008; revised 5 January 2009; accepted 18 February 2009; online publish-ahead-of-print 28 March 2009
2017
Time-dependent risk model to predict outcome in IE
Methods Patients and setting We retrospectively reviewed the medical records of consecutive patients with IE admitted between January 1996 and January 2006 to the Royal Prince Alfred Hospital (derivation cohort) and Concord Repatriation General Hospital (validation cohort). Both hospitals are university teaching hospitals in Sydney, Australia. Cases were included if they satisfied the modified Duke criteria for a definite or possible diagnosis of IE.12,13 For patients with recurrent admissions for IE (n ¼ 17), only the first admission was included for analysis. Data from the derivation cohort were used to identify important prognostic factors and develop the risk model. Data from the independent validation cohort were used to assess the predictive accuracy of the risk model. All patients were over the age of 18 and the study was approved by the Ethics Review Committees of both hospitals.
We collected data at three pre-specified time-points during a patient’s hospitalization: day 1 (baseline), day 8, and day 15. Baseline was defined as the triage time in the emergency department. Days 1, 8, and 15 were chosen as clinically important landmarks for abstracting the data because existing guidelines suggest that the first fortnight of hospitalization is a critical phase of illness during which complications occur frequently and important management decisions are often required.2,3 At these time-points, we systematically collected data on a range of variables that included all parameters found to be predictive of outcome in previous studies.5 – 11 Definitions of the study variables are provided in Supplementary material online, Section S1. Clinical variables included demographic information and overall comorbid status as assessed by the Charlson comorbidity scale.14 Examination findings at days 1, 8, and 15 were also recorded, including haemodynamic parameters, temperature, neurological signs, and the presence of heart failure as defined by the modified Framingham criteria.15 Clinical outcomes such as embolic events, intensive care admissions and surgery were also recorded. Laboratory variables were recorded at days 1, 8, and 15. These included haemoglobin, white cell count, platelet count, creatinine, albumin, and C-reactive protein. Electrocardiographic and echocardiographic data were also collected. This included details of the infected valve, size and mobility of vegetations, left ventricular function, severity of valvular regurgitation, and the presence of intracardiac abscesses or valvular dehiscence. Data available from transoesophageal examinations (n ¼ 124) were used preferentially. Microbiological data included the organism isolated in culture, antibiotic sensitivities, and serology results. An a priori decision was made to assess surgery as a potential confounder rather than as a predictor variable for risk modelling because its inclusion would invalidate the model for predicting patients who may benefit from surgery.16
Outcome and follow-up The study outcome was all-cause mortality during the 6 months period following admission. Survival status was determined from medical records and confirmed with data from the State Government of New South Wales Registry of Births and Deaths. Patients were censored according to their status at the time of last contact if 6 months outcome data were unavailable (3.3% of cases).
Using data from the derivation cohort, univariate associations with mortality were assessed using Cox regression. Separate analyses were performed at days 1, 8, and 15. For dynamic variables such as examination findings and laboratory parameters, the most proximate data collected within 24 h of each time-point were entered into the respective analyses. For static variables such as demographic information and comorbid conditions, the same data were entered into all analyses. Where data was unavailable, variables were imputed with the most frequent category for categorical variables and the median value for continuous variables.17 Variables were confirmed to fulfil the assumption of proportionality using a partial residual method.18 The linearity assumption was tested for all continuous variables by fitting higher-order terms. The primary risk model of the study incorporated continuous variables on their original scale and was developed using multivariable Cox regression analyses at days 1, 8, and 15. All variables that had been previously identified as predictors in the literature5 – 11 were considered as potential candidate variables for modelling. The number of candidate variables was reduced by the use of the Charlson scale14 as a global marker of comorbidity and the modified Framingham criteria15 to define heart failure; and, the exclusion of variables which exhibited high biological variability19 or had a low prevalence (i.e. present in ,5% of the derivation cohort). Furthermore, if significant collinearity was found between two variables (correlation coefficient 0.60), the clinically more important variable was entered into the model.20 Clinically plausible interactions were also included for analysis. The heuristic shrinkage estimator of van Houwelingen and le Cessie21 was calculated for the full model. A backward stepwise method was subsequently used and variables were removed on the basis of x2 Wald statistics. The resultant risk model comprised independent variables with P , 0.05. Based on the independent variables identified in the multivariable analysis, continuous variables were then dichotomized using clinically relevant cut-offs to derive an alternative simplified model suitable for clinical use. The potential loss of predictive value in the simplified model was assessed by comparing its discrimination with that of the primary model. The impact of potential confounders such as surgery, the site of valvular infection, and transoesophageal echocardiography were evaluated by repeating multivariable Cox regression analyses with these covariates as dummy variables. In particular, surgery was assessed as a time-dependent covariate.22 Independent variables were assigned point scores according to a linear transformation of their b-regression coefficient in the risk model.6 An overall risk score was calculated for each patient by adding together the points corresponding to their risk factors. Risk scores were then defined for low-, intermediate-, and high-risk groups on the basis of observed mortality gradients in the derivation cohort (low-risk ,15%, intermediate-risk 15– 50%, high-risk .50% 6 months mortality). Model calibration was determined by comparing the predicted and observed mortality rates for each risk group.23,24 Model discrimination was determined by calculating the concordance index for predicting 6 months mortality.18 Kaplan– Meier survival curves were generated to illustrate the partitioning of risk over time and differences in survival between risk groups were assessed by the log-rank test. The performance of the model was internally validated using 200 bootstrapped samples of the derivation cohort,18 and then externally validated in an independent cohort. The simplified time-dependent risk model developed in this study was then compared with the one published risk model for IE described
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Data collection and study variables
Statistical analyses
2018 by Hasbun et al.6 This existing risk model evaluated prognosis at only a single time-point and its predictive value has not been previously tested at different time-points during hospitalization. The existing model was also restricted to patients with complicated left-sided native endocarditis. To ensure consistency, both models were tested in a subset of patients in our combined cohort, who satisfied the inclusion criteria used by Hasbun et al. Extended statistical methods are provided in Supplementary material online, Section S2. All analyses were performed using SPSS 13.0 for Windows (SPSS Inc., Chicago, IL, USA).
Results Cohort identification
Characteristics of the derivation cohort Patients in the derivation cohort were young, predominantly male, and recent intravenous drug use was frequently reported (Table 1). The mitral valve was the most common site of infection and Staphylococcus aureus was the most common organism. Five patients died between days 1 and 8. Twelve patients died and a further three patients were discharged between days 8 and 15. Overall, 46 patients (24%) died within 6 months of hospital admission.
Univariate predictors of 6 months mortality in derivation cohort Several risk factors were consistent univariate predictors of mortality (P , 0.05) at days 1, 8, and 15: older age, comorbidity (Charlson score), heart failure, altered mental state, methicillinresistant Staphylococcus aureus, anaemia, thrombocytopenia, renal
impairment, hypo-albuminaemia, and multi-valvular infection (Table 2). In addition, tachycardia, absence of sinus rhythm, and echocardiographic left ventricular dysfunction were significant univariate predictors at days 1 and 8, but not at day 15. In contrast, hypotension, severe embolic events, and intracardiac abscesses became significant univariate predictors by days 8 and 15. Leucocytosis and elevated C-reactive protein were significant univariate predictors at day 8 only. Of note, prosthetic valve infection, vegetation size, and mobility, and severity of valvular regurgitation were not significant univariate predictors in the derivation cohort at any timepoint.
Multivariable risk modelling Independent of the univariate analyses, 11 clinical variables and 2 interaction terms were included for multivariable analyses on the basis of pre-specified clinical and statistical criteria [see Supplementary material online, Section S2(c)]. The estimated shrinkage for the full model fit was 0.84 for day 1, 0.83 for day 8, and 0.84 for day 15. Following backward stepwise selection, the final model contained six independent predictors at day 1, five at day 8, and only three at day 15 (Table 3). Comorbidity, heart failure, and thrombocytopenia independently predicted mortality at all three timepoints. Other predictors were older age, tachycardia, and renal impairment at day 1; and severe embolic events and renal impairment at day 8. The bootstrap corrected estimate of the concordance index for the continuous risk model was estimated as 0.82 at day 1, 0.87 at day 8, and 0.77 at day 15. Based on the independent predictors already identified, a simplified clinical risk model was developed with continuous variables dichotomized according to accepted clinical cut-offs (Table 4). Independent predictors were assigned points based on their ß-regression coefficient. The bootstrap corrected estimate of the concordance index for the simplified risk model was 0.80 at day 1, 0.80 at day 8, and 0.77 at day 15.
Figure 1 Cohort identification. Cases were excluded if they did not satisfy the modified Duke criteria for a definite or possible diagnosis of infective endocarditis or represented recurrent admissions.
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
From 1996 to 2006, there were 344 consecutive admissions with a primary diagnosis of IE to the two independent hospitals. Of these, 273 patients met the inclusion criteria: 192 patients in the derivation cohort and 81 patients in the validation cohort (Figure 1).
R.W. Sy et al.
2019
Time-dependent risk model to predict outcome in IE
Table 1 Patient characteristics and clinical outcomes Derivation cohort (n 5 192)
Validation cohort (n 5 81)
P-valuea
47.1 (18.7) 132 (69)
62.6 (17.6) 54 (67)
,0.001 0.78
96 (50) 96 (50)
69 (85) 12 (15)
,0.001
Diabetes
20 (10)
15 (19)
0.12
Chronic renal impairment Dialysis-dependent
17 (9) 12 (6)
12 (15) 8 (10)
0.21 0.45
7 (4)
5 (6)
0.52
165 (86)
57 (70)
,0.001
27 (14)
24 (30)
Predisposing cardiac disease Intravenous drug use
71 (37) 45 (23)
42 (52) 7 (9)
0.06 0.002
Central venous catheter
12 (6)
10 (12)
0.15
169 (88) 23 (12)
72 (89) 9 (11)
1.00
Variable
............................................................................................................................................................................... Demographic Age—mean (SD), years Male sex—n (%)
............................................................................................................................................................................... Admission source—n (%) Direct admission Referred from another hospital
............................................................................................................................................................................... Comorbid conditions—n (%)b
Immunosuppression
3
............................................................................................................................................................................... Duke status—n (%) Definite Possible
............................................................................................................................................................................... Valve implicated—n (%) Mitral
79 (41)
44 (54)
0.05
Aortic Aortic and mitral
63 (33) 8 (4)
29 (33) 2 (3)
1.00 0.73
Tricuspid
35 (18)
4 (5)
0.004
Otherc Prosthetic
7 (4) 37 (19)
2 (2) 17 (21)
1.00 0.74
Staphylococcus aureus
89 (46)
28 (35)
0.08
Methicillin resistant Staphylococcus aureus Viridans Streptococci
10 (5) 31 (16)
5 (6) 22 (27)
0.77 0.04
4 (2)
7 (9)
0.02
13 (7) 6 (3)
10 (12) 4 (5)
0.49 0.15
............................................................................................................................................................................... Organism isolated—n (%)
Other Streptococci Enterococci Coagulase negative Staphylococci Otherd
16 (8)
3 (4)
,0.001
Culture negative
33 (17)
7 (9)
0.09
............................................................................................................................................................................... Clinical outcomes Length of stay—mean (SD), days
a
37.4 (26.7)
38.6 (25.4)
0.73
Any embolic event—n (%) Stroke—n (%)b
66 (34) 28 (15)
34 (42) 15 (19)
0.42 0.47
Severe embolic event—n (%)b
33 (17)
23 (28)
0.05
Intensive care admission—n (%) Valvular surgery—n (%)
50 (26) 57 (30)
18 (22) 9 (11)
0.54 0.001
In-hospital mortality—n (%)
39 (20)
24 (30)
0.12
Six-month mortality—n (%)
46 (24)
26 (32)
0.18
Comparisons between cohorts were determined by the x2 test, Fisher’s exact test, or the independent samples t-test as appropriate. As defined in Supplementary material online, Section S1. c Included pacing and defibrillation leads, atrial wall, atrial baffle, pulmonary artery, and aortic root. d Included fungi, Gram-negative bacilli including the HACEK group, Streptococcal-like organisms, and Gram-positive bacilli. b
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Charlson comorbidity scoreb 0– 2
2020
R.W. Sy et al.
Table 2 Univariate association of variables and 6 months mortality for the derivation cohort Variablea
Day 1 analysis (n 5 192)
Day 8 analysis (n 5 187)b
Day 15 analysis (n 5 172)b
Hazard ratio (95% CI)
Hazard ratio (95% CI)
Hazard ratio (95% CI)
.................................. P-value
.................................. P-value
.................................. P-value
............................................................................................................................................................................... Demographic and comorbidity Age (per increase of 10 years) Male
1.20 (1.02– 1.40) 1.03 (0.55– 1.93)
0.02 0.93
1.18 (1.01–1.37) 1.09 (0.55–2.13)
0.04 0.81
1.30 (1.07–1.58) 1.13 (0.50–2.56)
0.01 0.77
Diabetes
2.12 (0.98– 4.51)
0.06
1.74 (0.73–4.14)
0.21
2.39 (0.90–6.34)
0.08
Charlson comorbidity score (per unit increase)
1.25 (1.12– 1.40)
,0.001
1.26 (1.12–1.42)
,0.001
1.31 (1.15–1.50)
,0.001
............................................................................................................................................................................... Clinical findingsc Mean blood pressure (per decrease of 10 mmHg)
1.03 (0.83– 1.28)
0.80
1.36 (1.02–1.83)
0.04
1.14 (1.01–1.28)
0.04
Pulse rate (per increase of 10 b.p.m.)
1.28 (1.08– 1.53)
0.004
1.19 (1.07–1.31)
0.01
1.16 (0.80–1.69)
0.42
Temperature (per increase of 18C) Congestive heart failure
0.91 (0.67– 1.24) 3.26 (1.74– 6.11)
0.56 ,0.001
1.31 (0.77–2.22) 4.71 (2.43–9.11)
0.76 ,0.001
1.32 (0.62–2.83) 7.14 (3.19–16.00)
0.48 ,0.001
3.11 (1.59– 6.11)
0.001
2.59 (1.22–5.51)
0.01
2.91 (1.14–7.37)
0.03
Non-sinus rhythm on ECG
2.69 (1.37– 5.29)
0.004
3.84 (1.93–7.63)
,0.001
2.52 (0.84–7.51)
0.10
1.60 (0.89– 2.86)
0.12
1.50 (0.80–2.70)
0.22
1.44 (0.68–3.02)
0.34
3.35 (1.42– 7.91)
0.006
3.02 (1.30–7.01)
0.01
3.90 (1.35–11.24)
0.01
1.20 (1.04– 1.37) 1.26 (1.00– 1.58)
0.02 0.05
1.21 (1.01–1.44) 1.44 (1.24–1.67)
0.04 ,0.001
1.17 (1.01–1.38) 1.07 (0.93–1.22)
0.04 0.35
............................................................................................................................................................................... Microbiology Staphylococcus aureus Methicillin resistant Staphylococcus aureus
............................................................................................................................................................................... Blood resultsc Haemoglobin (per decrease of 10 g/L) White cell count (per increase of 5 109/L) Platelet count (per decrease of 20 109/L)
1.13 (1.04– 1.22)
0.002
1.20 (1.11–1.29)
,0.001
1.15 (1.06–1.24)
,0.001
Serum creatinine (per increase of 20 mmol/L) Serum albumin (per decrease of 5 g/L)
1.06 (1.02– 1.10) 1.30 (1.05– 1.61)
0.04 0.02
1.08 (1.04–1.13) 1.29 (1.01–1.65)
,0.001 0.04
1.20 (1.11–1.29) 1.62 (1.19–2.22)
,0.001 0.002
C-reactive protein (per increase of 20 mg/L)
1.13 (1.00– 1.27)
0.05
1.25 (1.07–1.46)
0.009
1.06 (0.69–1.63)
0.81
Vegetation size 10 mm Vegetation mobility score 3
1.02 (0.57– 1.82) 1.16 (0.65– 2.08)
0.95 0.61
0.95 (0.51–1.76) 0.98 (0.53–1.83)
0.87 0.96
0.99 (0.47–2.09) 0.94 (0.44–2.03)
0.99 0.87
,0.001
............................................................................................................................................................................... d
Echocardiography
Left ventricular ejection fraction ,40%
3.76 (1.97– 7.15)
,0.001
3.63 (1.81–7.24)
2.15 (0.82–5.66)
0.12
Moderate-to-severe regurgitation Aortic and mitral infection
0.79 (0.44– 1.40) 5.37 (2.26– 12.74)
0.42 ,0.001
0.81 (0.44–1.49) 5.29 (2.06–13.55)
0.49 0.001
0.95 (0.45–2.00) 7.46 (2.57–21.68)
0.89 ,0.001
Prosthetic valve infection
1.39 (0.70– 2.73)
0.35
1.63 (0.82–3.25)
0.17
1.53 (0.65–3.59)
0.33
1.87 (0.79– 4.41) 2.68 (0.96– 7.49)
0.15 0.06
2.68 (1.39–5.17) 3.49 (1.71–7.12)
0.003 0.001
2.43 (1.07–5.51) 2.63 (1.00–6.93)
0.03 0.05
............................................................................................................................................................................... Clinical complications Severe embolic event Intracardiac abscess a
Variables as defined in Supplementary material online, Section S1. Hazard ratios for continuous variables are calculated for incremental changes relative to the mean values for the cohort. Data were available in 90% of patients for all variables with the following exceptions: C-reactive protein (64% availability) and electrocardiography (82% availability) at day 1; C-reactive protein (66% availability), clinical heart failure status (77% availability), and electrocardiography (87%) at day 8; clinical heart failure status (83% availability), haemodynamic status (70% availability), electrocardiography (57% availability), serum albumin (61% availability), and C-reactive protein (33% availability) at day 15. When data were missing, variables were imputed with the most frequent category for categorical variables and the median value for continuous variables. b Analyses at days 8 and 15 were performed on surviving inpatients (taking into account the death of five patients by day 8 and a further 12 patients by day 15, as well as the discharge of three patients between days 8 and 15). c Time-specific analyses were performed for these variables using first available data within 24 h of the reference timepoint. d Transoesophageal data were used for vegetation characteristics; in the absence of a visible vegetation, size was coded as ,10 mm and mobility as ,3 for the purpose of analysis; mobility score as defined by Sanfilippo et al.8
Prognostic stratification using risk model Patients in the derivation cohort were stratified into three prognostic groups on the basis of observed mortality gradients (low-risk ,15%, intermediate-risk 15–50%, high-risk .50%
6 months mortality) and these groups were confirmed to have significantly different survival over time (Figure 2, left panels). Observed and predicted mortality rates were estimated for each risk group using bootstrap resampling, and the apparent
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Altered mental state
2021
Time-dependent risk model to predict outcome in IE
Table 3 Multivariable risk model using continuous variables
External validation of risk model
1.23 (1.01–1.48) 2.57 (1.20–5.51)
0.035 0.015
Platelet count (per decrease of 20 109/L)
1.15 (1.06–1.24)
,0.001
Creatinine (per increase of 20 mmol/L)
1.08 (1.00–1.17)
0.037
Severe embolic event
3.59 (1.63–7.93)
0.002
Serial evaluation of patient prognosis
Risk factora
Hazard ratio (95% CI)b
P-value
................................................................................ Day 1 Age (per increase of 10 years) Charlson score (per unit increase)
1.34 (1.10–1.63) 1.23 (1.06–1.40)
0.007 0.004
Pulse rate (per increase of 10 b.p.m.)
1.36 (1.14–1.63)
0.001
Congestive heart failure
2.52 (1.26–4.98)
0.008
Platelet count (per decrease of 20 109/L)
1.13 (1.04–1.22)
0.012
Creatinine (per increase of 20 mmol/L)
1.11 (1.02–1.20)
0.010
................................................................................ Day 8
................................................................................ Day 15 Charlson score (per unit increase) Congestive heart failure
1.22 (1.03–1.44) 8.99 (3.61–22.37)
Platelet count (per decrease of 20 109/L)
1.13 (1.04–1.22)
0.020 ,0.001 0.005
a
Missing data were imputed for analysis (see Table 2 for details). Hazard ratios are calculated for incremental changes relative to the mean values for the cohort. b
calibration was satisfactory across all timepoints (Figure 3, left panels).
The impact of surgery and other potential confounders In all, 57 patients (30%) in the derivation cohort underwent surgery at a median time of 18 days (IQR 9–37) after admission. This included 30 patients who underwent surgery during the first fortnight of admission. In this specific subgroup, the majority of patients (n ¼ 18, 60%) did not change risk groups between days 1 and 15 despite interval surgery. In the remaining patients, seven crossed to a lower risk group, while five crossed to a higher risk group by day 15. And, while early surgery corrected intracardiac abscesses, heart failure resolved acutely in only two patients as a result of surgery. Moreover, repeating the multivariable analysis with the inclusion of surgery as a time-dependent covariate resulted in a comparable final risk model (see Supplementary material online, Table S1a). Likewise, neither the inclusion of the site of infection (left- vs. right-sided IE) nor the availability of transoesophageal data as dummy covariates during model development changed the composition of the final risk model (see Supplementary material online, Table S1b–c).
Considering the two cohorts as a whole, patient prognosis was highly dynamic with considerable crossover between risk groups during serial evaluation at days 1, 8, and 15 (see Supplementary material online, Figure S1). Compared with their initial risk assessment at day 1, 90 patients (34%) crossed over to a different risk group at day 8. Similarly, 84 patients (34%) changed risk groups between days 8 and 15. Two clinically important subgroups were identified as a result of serial evaluation at days 1, 8, and 15. First, patients who were classified as low-risk at all three timepoints (n ¼ 84) had a relatively favourable prognosis with an overall 6 months mortality of 2.4%. In contrast, patients who were assessed as high-risk at any timepoint during the first fortnight (n ¼ 55) had an unfavourable prognosis with an overall 6 months mortality of 78.2%. The remaining patients had a variable prognosis and their overall 6 months mortality was 20.3%.
Comparison with an existing risk model Both the time-dependent risk model developed in this study and an existing risk model6 were applied to a subset of patients from our combined cohort who satisfied the criteria for complicated leftsided native IE (n ¼ 137). First, the existing risk model was used to stratify patients into four risk groups at days 1, 8, and 15 (Table 5). This risk model provided excellent discrimination of patients in the highest risk group and the overall linear trend for mortality across the four prognostic groups was statistically significant. However, if the highest risk group was excluded, the differences in the observed mortality rates between prognostic groups 1, 2, and 3 were clinically modest and not statistically significant (P ¼ 0.75 at day 1, P ¼ 0.49 at day 8, and P ¼ 0.72 at day 15, x2 test for linear trend). The concordance index ranged from 0.72 to 0.74, suggesting moderate overall discrimination.
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Charlson score (per unit increase) Congestive heart failure
Differences in baseline characteristics between patients in the derivation and validation cohorts are summarized in Table 1. Patients in the validation cohort were on average 15 years older, were more commonly direct admissions from the community, and had more comorbid conditions. Recent intravenous drug use and tricuspid valve infections were less frequent. The validation cohort also underwent surgery less frequently but the 6 months mortality rates were similar between the two cohorts. The derived risk model was used to stratify patients in the validation cohort into risk groups at days 1, 8, and 15. We observed that the risk groups had significantly different survival over time (Figure 2, right panels), and there was satisfactory calibration between observed and predicted mortality rates for each risk group (Figure 3, right panels). The overall concordance index of the continuous risk model for predicting mortality in the validation cohort was 0.81 (95% CI 0.70– 0.93) for day 1, 0.80 (95% CI 0.63– 0.97) for day 8, and 0.83 (95% CI 0.71–0.95) for day 15. The concordance index of the simplified risk model was comparable, being 0.79 (95% CI 0.68– 0.91) for day 1, 0.79 (95% CI 0.65– 0.93) for day 8, and 0.84 (95% CI 0.73 –0.95) for day 15.
2022
R.W. Sy et al.
Table 4 Simplified dichotomous risk model and associated risk score assignment Risk factora
Hazard ratio (95% CI)
P-value
b regression coefficient
Points assignedb
............................................................................................................................................................................... Day 1 risk model Age 65
2.14 (1.09–4.23)
0.029
0.76
2
Charlson score 3 Pulse rate 120 b.p.m.
2.43 (1.25–4.71) 2.85 (1.48–5.49)
0.009 0.002
0.89 1.05
2 3
Congestive heart failure
2.47 (1.34–4.56)
0.004
0.91
2
Platelet count ,150 109/L Creatinine 120 mmol/L
2.13 (1.12–4.05) 3.71 (2.01–6.84)
0.021 ,0.001
0.76 1.31
2 4
Charlson score 3
2.26 (1.10–4.66)
0.027
0.82
2
Congestive heart failure Platelet count ,150 109/L
2.79 (1.41–5.52) 2.45 (1.24–4.83)
0.003 0.010
1.03 0.90
3 2
............................................................................................................................................................................... Day 8 risk model
2.70 (1.43–5.09)
0.002
0.99
2
2.61 (1.33–5.12)
0.005
0.96
2
............................................................................................................................................................................... Day 15 risk model Charlson score 3
2.98 (1.21–7.36)
Congestive heart failure Platelet count ,150 109/L
6.19 (2.72–14.07) 4.14 (1.57–10.96)
0.018
1.09
2
,0.001 0.004
1.82 1.42
3 3
a
Continuous variables were dichotomized using accepted clinical cut-offs. Missing data were imputed for analysis (see Table 2 for details). Assignment of points to risk factors was based on a linear transformation of the b regression coefficient. The coefficient was divided by the lowest b value, multiplied by a constant of 2, and rounded to the nearest integer. b
In comparison, the time-dependent risk models developed in this study performed equally well at identifying high-risk patients (Table 5). However, it was better able to differentiate between patients at low-risk and those at intermediate-risk with the difference in observed mortality rates between these two groups being clinically meaningful (two-to-five-fold difference in mortality) and statistically significant (P ¼ 0.02 at day 1; P , 0.01 at day 8; and P , 0.01 at day 15; Fisher’s exact test). The concordance index of the time-dependent risk model was also higher at all timepoints.
Discussion Infective endocarditis is a complex disease with wide variations in its clinical course and prognosis. This study establishes for the first time that the key prognostic factors of IE change over time, and that they can be combined into a simple risk model that accurately estimates the risk of mortality at different timepoints during early hospitalization. Since the model accounts for potential changes in patient prognosis during treatment, it may improve the prognostic classification of patients with this life-threatening disease. The observation that the prognostic value of individual parameters in IE changes over time is compatible with the natural history of IE during antibiotic treatment. For example, embolic events dramatically diminish in frequency after 1-to-2 weeks of treatment25 and therefore their association with mortality may be expected to be time-dependent. Previous studies in IE have also demonstrated that parameters such as heart failure,26 platelet count,27 and C-reactive protein10 are subject to time-dependent
changes in their occurrence and prognostic value, enhancing the clinical plausibility of our observation that prognostic parameters in IE change over time. In addition, the use of a time-dependent model to serially evaluate overall patient prognosis may be preferable to evaluation at a single time-point as we observed considerable crossover in risk group assignment during the first 2 weeks of hospitalization. Indeed, five patients in our cohort classified as low-risk at day 1 were subsequently reclassified as high-risk, and 23 patients initially classified as intermediate-risk also became high-risk later in the admission. Moreover, patients assessed as high-risk at any timepoint had a poor outlook with a 6 months mortality of 78.2%. In contrast, patients consistently classified as low-risk on serial evaluation had a 6 months mortality of only 2.4%. The time-dependent risk model developed in this study compared favourably with an existing risk model designed to evaluate prognosis at a single timepoint.6 Both models used 6 months mortality as the primary outcome, a separate cohort for validation, and comorbid disease and heart failure were important components of both models. The time-dependent model was at least equivalent in its identification of high-risk patients and may have an advantage in discriminating between low-risk and intermediate-risk patients. Our results should encourage the further development and application of time-dependent risk models for IE because they may better reflect the dynamic nature of the disease and changes in prognosis that occur during treatment in hospital. There are several potential applications of the approach developed in this study. First, time-dependent evaluation may be
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Creatinine 120 mmol/L Severe embolic event
Time-dependent risk model to predict outcome in IE
2023
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Figure 2 Kaplan– Meier survival curves for risk groups assigned using time-dependent risk model. P-values are shown for the log-rank test comparing survival between risk groups. The number of patients at risk of mortality is shown below the x-axis. useful in the design of future trials of interventions such as surgery by more accurately identifying high-risk patients, thereby possibly reducing the sample size of the study. Secondly, since the present risk model incorporates readily available parameters, it should be widely applicable for clinical use in community hospital settings as well as tertiary referral centres. However, caution should be exercised when applying populationderived risk models to individual patient care. In particular, variables
not represented in our final risk model may be significant in other populations. For example, our study should not be interpreted as diminishing the significance of established adverse features such as diabetes5 and perivalvular infection1 in individual patients since these conditions were under-represented in our cohort. Similarly, although transoesophageal echocardiography was not an important confounder in this study, it was performed in only 45% of our patients and higher utilization rates may have
2024
R.W. Sy et al.
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Figure 3 Comparison of observed and predicted 6 months mortality rates across risk groups. Column graphs comparing the observed (light columns) and the predicted (dark columns) 6 months mortality rates (mean with SEM) for each risk group. Number of patients assigned to each risk group and corresponding range of scores are shown below the x-axis. Mortality rates for the derivation cohort were estimated using the bootstrap method. Agreement between observed and predicted mortality rates determined using the goodness-of-fit test.23 P-values .0.05 suggested satisfactory calibration between observed and predicted mortalities for the overall risk model at all time-points. increased the observed prognostic value of echocardiographic parameters.8,9 In addition, although a post hoc analysis demonstrated that surgery was not an influential confounder on risk modelling, the influence of surgery on outcomes remains a contentious issue.22 This study had several limitations. First, bias inherent to the retrospective study design was reduced by our use of strict definitions for case selection and predictor variables, the confirmation of outcomes using a government registry, and the fact that the level of missing data for most variables was low. The study was also limited by the small number of events relative to the number of
variables fitted18,20 and the potential for overestimation of the model’s accuracy in the derivation cohort. Moreover, although the model was externally validated in an independent cohort and mortality was reliably predicted at all timepoints, it would be ideal to confirm the model in an independent and larger cohort, such as the International Collaboration on Endocarditis database.28 Thirdly, simplification of our primary risk model using dichotomized variables could in theory reduce prognostic accuracy, but the simplified model actually demonstrated similar accuracy during external validation, supporting its potential utility in clinical practice. Finally, although days 8 and 15 were pre-specified as
2025
Time-dependent risk model to predict outcome in IE
Table 5 Comparison of previous risk model and current time-dependent risk model Prognostic groups
Day 1a
Day 8a
Day 15a
............................................................................................................................................................................... Existing risk model: Hasbun et al.6 Group 1
Number of patients 6 months mortality, n (%)
29 5 (17)
25 4 (16)
25 3 (12)
Group 2
Number of patients 6 months mortality, n (%)
34 6 (18)
29 5 (17)
30 6 (20)
Group 3
Number of patients 6 months mortality, n (%)
45 9 (20)
35 8 (23)
37 6 (16)
Group 4
Number of patients 6 months mortality, n (%)
29 21 (72)
21 15 (71)
15 11 (73)
,0.001 0.73 (0.64– 0.83)
,0.001 0.74 (0.62–0.85)
,0.001 0.72 (0.59–0.84)
P-value for linear trendb Concordance index (95% confidence interval)
...............................................................................................................................................................................
Time-dependent risk model Number of patients 6 months mortality, n (%)
71 10 (14)
74 10 (14)
72 6 (8)
Inter-risk
Number of patients 6 months mortality, n (%)
50 16 (32)
16 8 (50)
15 7 (47)
High-risk
Number of patients 6 months mortality, n (%)
P-value for linear trendb
16 15 (94) ,0.001
20 14 (70) ,0.001
20 13 (65) ,0.001
Concordance index (95% confidence interval)
0.82 (0.74– 0.90)
0.77 (0.66–0.87)
0.82 (0.72–0.92)
a At day 1, 137 patients in the combined cohort fulfilled the inclusion criteria used by Hasbun et al.6 At subsequent timepoints, only a proportion of these patients could be assessed because risk stratification required a complete set of variables for each patient (n ¼ 110 at day 8, n ¼ 107 at day 15). With regards to missing data (see Table 2 for details), patients with any missing data were excluded from this particular analysis. b 2 x test of linear trend comparing 6 months mortality rates across risk groups.
clinically relevant timepoints for reassessing prognosis, this does not preclude the potential importance of other timepoints. In conclusion, we have used a contemporary cohort to derive and validate a simple time-dependent risk model for IE. The risk model incorporates readily available clinical parameters and takes into account the dynamic nature of the disease. It accurately predicts the risk of mortality in patients before and during the early phase of hospitalization, and may improve the clinical management of patients with IE.
Supplementary material Supplementary Material is available at European Heart Journal online.
Acknowledgements We acknowledge Dr Federica Barzi and Ms Rachel O’Connell for their assistance with the statistical analyses.
Funding University of Sydney Postgraduate Award to R.W.S.
Conflict of interest: none declared.
References 1. Baddour LM, Wilson RW, Bayer AS, Fowler VG, Bolger A, Levison M, Ferrieri P, Gerber M, Tani L, Gewitz M, Tong D, Steckelberg J, Baltimore R, Shulman S, Burns J, Falace D, Newburger J, Pallasch T, Takahashi M, Taubert K. Infective endocarditis: diagnosis, antimicrobial therapy, and management of complications. Circulation 2005;111:e394–e433. 2. Bashore T, Cabell CH, Fowler VG. Update on infective endocarditis. Curr Probl Cardiol 2006;31:274 –352. 3. Andrews MM, von Reyn CF. Patient selection criteria and management guidelines for outpatient parenteral antibiotic therapy for native valve infective endocarditis. Clin Infect Dis 2001;33:203 –209. 4. Horstkotte D, Follath F, Gutschik E, Lengyel M, Oto A, Pavie A, Soler-Soler J, Thiene G, von Graevenitz A. Guidelines on prevention, diagnosis and treatment of infective endocarditis. Eur Heart J 2004;25:267–276. 5. Chu VH, Cabell CH, Benjamin DK Jr, Kuniholm EF, Fowler VG Jr, Engemann J, Sexton DJ, Corey GR, Wang A. Early predictors of in-hospital death in infective endocarditis. Circulation 2004;109:1745 –1749. 6. Hasbun R, Vikram HR, Barakat LA, Buenconsejo J, Quagliarello VJ. Complicated left-sided native valve endocarditis in adults: risk classification for mortality. JAMA 2003;289:1933 –1940. 7. Wallace SM, Walton BI, Kharbanda RK, Hardy R, Wilson AP, Swanton RH. Mortality from infective endocarditis: clinical predictors of outcome. Heart 2002;88: 53 – 60. 8. Sanfilippo A, Picard M, Newell J, Rosas E, Davidoff R, Thomas J, Weyman A. Echocardiographic assessment of patients with infectious endocarditis: prediction of risk for complications. J Am Coll Cardiol 1991;18:1191 –1199. 9. Thuny F, Di Salvo G, Belliard O, Avierinos JF, Pergola V, Rosenberg V, Casalta JP, Gouvernet J, Derumeaux G, Iarussi D, Ambrosi P, Calabro R, Riberi A, Collart F, Metras D, Lepidi H, Raoult D, Harle JR, Weiller PJ, Cohen A, Habib G. Risk of embolism and death in infective endocarditis: prognostic value of echocardiography: a prospective multicenter study. Circulation 2005;112:69 –75. 10. Verhagen D, Hermanides J, Korevaar J, Bossuyt P, van den Brink R, Speelman P, van der Meer JT. Prognostic value of serial C-reactive protein measurements in left-sided native valve endocarditis. Arch Intern Med 2008;168:302–307.
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
Low-risk
2026
19. Ricos C, Alvarez V, Cava F, Garcia-Lario JV, Hernandez A, Jimenez CV, Minchinela J, Perich C, Simon M. Current databases on biologic variation: pros, cons and progress. Scand J Clin Lab Invest 1999;59:491 –500. 20. Concato J, Feinstein A, Holford T. The risk of determining risk with multivariable models. Ann Intern Med 1993;118:201 –210. 21. van Houwelingen J, le Cessie S. Predictive value of statistical models. Stat Med 1990;8:1303 –1325. 22. Tleyjeh IM, Baddour LM. The potential impact of survivor treatment selection bias on the perceived efficacy of valve surgery in the treatment of infective endocarditis. Clin Infect Dis 2007;44:1392 –1393. 23. May S, Hosmer D. A simplified method for calculating an overall goodness-of-fit test for the Cox proportional hazards model. Lifetime Data Analysis 1998;4:109 –120. 24. Marschner I, Colquhoun D, Simes R, Glasziou P, Harris P, Singh B, Friedlander D, White H, Thompson P, Tonkin A. Long-term risk stratification for survivors of acute coronary syndromes. J Am Coll Cardiol 2001;38:56–63. 25. Steckelberg JM, Murphy J, Ballard D, Bailey K, Tajik A, Taliercio C, Giuliani E, Wilson W. Emboli in infective endocarditis: the prognostic value of echocardiography. Ann Intern Med 1991;114:635 –640. 26. Mills J, Utley J, Abbott J. Heart failure in infective endocarditis: predispoing factors, course, and treatment. Chest 1974;66:151 –157. 27. Sy R, Chawantanpipat C, Richmond D, Kritharides L. Thrombocytopenia and mortality in infective endocarditis. J Am Coll Cardiol 2008;51:1824 –1825. 28. Cabell C, Abrutyn E. Progress towards a global understanding of infective endocarditis. Early lessons from the International Collaboration on Endocarditis investigation. Infect Dis Clin North Am 2002;16:255 –272.
People’s corner
doi:10.1093/eurheartj/ehr277
.............................................................................................................................................................................
Alec Vahanian new International Associate editor for EHJ Alec Vahanian was appointed International Associate editor for the EHJ effective July 2011 and brings with him considerable experience. His previous position at the EHJ was International Editorial Board member. As head of cardiology at Bichat Hospital, Paris, and professor of cardiology at Paris VII University, France, Alec is no stranger to the EHJ & ESC. He has been a reviewer for the European Heart Journal for many years as well as for the Lancet, NEJM, Circulation, JACC, and Am J Cardiol, and is an Associate Editor of EuroIntervention, and a member of the Editorial committee of eight other journals. He is Fellow of the European Society of Cardiology, and has been chairman of the working group on valvular heart disease, chairman of the Euroheart survey on valvular heart disease and of the first ESC Task force on valvular heart disease. Additionally, he was chairman of the ESC Committee for Practice Guidelines and a nominated member of the ESC Board (2006–2010), and is currently chairman of the update on the guidelines for valvular heart disease for ESC/EACTS. He is also a Fellow of the Royal College of Physicians of Edinburgh, and an associate member of the French Academy of Medicine. His earlier positions were: president of the French Federation of Cardiology and member of the Board of Directors of PCR. As a cardiologist, his special interests are interventional cardiology, coronary artery disease, and mainly valvular heart disease—in particular, evaluation, surgical management, and transcatheter valve interventions on both the mitral and aortic valves. His group performed the first percutaneous mitral commissurotomy in Europe and are actively involved in transcatheter aortic valve implantation and transcatheter mitral valve repair. Alec studied medicine in Paris VI University graduating in 1972, is 61 years old, and married with four children (and three grandchildren!). He has authored 297 papers in peer-reviewed journals and 43 chapters in books. He states ‘I am pleased and honoured to be invited to participate even more actively in the EHJ which has undergone very significant development over the recent years. I would be happy to contribute to further disseminate the best evidence in the field of my expertise’. Published on behalf of the European Society of Cardiology. All rights reserved. & The Author 2011. For permissions please email:
[email protected].
Downloaded from http://eurheartj.oxfordjournals.org/ by guest on February 20, 2013
11. Hill EE, Herijgers P, Claus P, Vanderschueren S, Herregods MC, Peetermans WE. Infective endocarditis: changing epidemiology and predictors of 6-month mortality: a prospective cohort study. Eur Heart J 2007;28:196 –203. 12. Durack DT, Lukes AS, Bright D. New criteria for diagnosis of infective endocarditis: utilization of specific echocardiographic findings. Am J Med 1996;96: 200 –209. 13. Li JS, Sexton DJ, Mick N, Nettles R, Fowler VG, Ryan T, Bashore T, Corey GR. Proposed modifications to the Duke criteria for the diagnosis of infective endocarditis. Clin Infect Dis 2000;30:633 –638. 14. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis 1987;40:373 –383. 15. McKee P, Castelli W, McNamara P, Kannel W. The natural history of congestive heart failure: the Framingham Study. N Engl J Med 1971;285:1441 –1446. 16. Granowitz EV, Longworth DL. Risk stratification and bedside prognostication in infective endocarditis. JAMA 2003;289:1991 –1993. 17. Wang A, Pappas P, Anstrom KJ, Abrutyn E, Fowler VG Jr, Hoen B, Miro JM, Corey GR, Olaison L, Stafford JA, Mestres CA, Cabell CH, Investigators ICoE. The use and effect of surgical therapy for prosthetic valve infective endocarditis: a propensity analysis of a multicenter, international cohort. Am Heart J 2005;150: 1086 –1091. 18. Harrell F, Lee K, Mark D. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996;15:361 –387.
R.W. Sy et al.
