Two prospective cohorts of consecutive patients with LSIE from 4 tertiary university hospitals were used in this study. The first cohort (cohort 1) included all patients consecutively diagnosed with definite LSIE between 2000 and 2017 from 3 hospitals and was used to derive and internally validate the predictive model. The second cohort (cohort 2), used for the external validation, included all patients with a final diagnosis of LSIE between 2012 and 2017 admitted to another hospital. All the centers are tertiary university hospitals with immediate cardiac surgery facilities, and are leaders in treatment and research in IE.

The participating centers have ongoing prospective local databases including all consecutive patients with IE admitted to their institutions. A standardized case report form for each patient was recorded at each site. The protocols conformed to the ethics guidelines of the 1975 Declaration of Helsinki and its subsequent revisions and were approved by the local ethics committees. The proportion of missing data was <10% in all analyzed variables.

We included only patients with definitive LSIE according to the Duke criteria until 2002 and the modified Duke criteria thereafter.10,11

The predictive model was derived and internally validated using data from cohort 1 (n=1002). This population was randomized in a 2:1 proportion for the derivation and internal validation samples. Approximately two thirds of the population were used to derive the model (derivation sample, n=688) and the other third to validate it (internal validation sample, n=314). The predictive model was designed on the basis of the results of a multivariable analysis of in-hospital mortality including all the prognostic variables proposed by the European guidelines. The model was externally validated in cohort 2 (n=133). The study design is presented in figure 1.

Figure 1.

Study design. IE, infective endocarditis.

(0.1MB).

A total of 17 out of the 19 prognostic variables proposed in the European Society of Cardiology (ESC) IE guidelines were included in our analysis. Variables were recorded during hospital admission, but only preoperatively in the case of cardiac surgery. Premature mitral valve closure and echocardiographic signs of elevated diastolic pressures were not included, as these factors were considered as surrogates of heart failure. In addition, the definition of some variables was adapted to achieve higher simplicity and reproducibility in the use of the predictive model. Severe left-sided valve regurgitation and severe prosthetic valve dysfunction were grouped together and valvular vegetation was considered irrespective of their length since there is no clear evidence-based cutoff point to consider a vegetation as large. The prespecified predictors and their definitions are summarized in table 1.

In-hospital mortality was used as the main event and included all-cause mortality during hospital stay. Antibiotic treatment and indications for surgery followed the recommendations of the European guidelines and decisions were taken by multidisciplinary experienced groups on IE. We considered urgent surgery to be surgery performed during the active phase of the disease, before the end of antibiotic treatment.12

Categorical variables are reported as frequency (No.) and percentages and continuous variables as the mean±standard deviation or median and [interquartile range] in cases of nonnormal distribution. Normal distribution of quantitative variables was verified with the Kolmogorov-Smirnov test and visually through Q-Q plot graphics. Qualitative variables were compared with the chi-square test and Fisher exact test. Continuous variables were compared with the Student t test or its equivalent for nonparametric tests, the Mann-Whitney U test, for variables that were nonnormally distributed.

Randomization of cohort 1 was done by individual simple assignment of each episode with a probability of 0.67 for the derivation sample and a probability of 0.33 for the validation sample. We used the C4 Study Design Pack V 1.1 Glaxo Wellcome S.A. program.

Univariable analysis was performed in the derivation sample (cohort 1) to test the linear relation of each variable with the outcome, in-hospital mortality. To derive the predictive model, a logistic regression model with the maximum likelihood method using backward stepwise selection was adjusted, which included the prognostic factors shown in table 1. The ratio variable/event was controlled to avoid overfitting. For the final model, odds ratios (OR) adjusted for each of the variables included were calculated, along with their 95% confidence intervals (95%CI). This model was internally validated in the validation sample (cohort 1) and externally in cohort 2.

Noncollinearity was verified among the variables included in the model. The area under the receiver operating characteristic curve (ROC curve) was used to measure how well the model discriminated between patients with a high and low risk of in-hospital mortality. A value of 0.5 indicates no discrimination and a value equal to 1 indicates perfect discrimination. Calibration was evaluated with the Hosmer-Lemeshow test and with plots comparing predicted and observed mortality for different levels of risk.

P values are bilateral and were considered statistically significant with a P value <.05. Analyses were performed with the use of SPSS software, version 24.0 (IBM), and R software, version 3.4.3 (R Foundation for Statistical Computing).

The description of the main features in cohort 1 and the comparison between the derivation and internal validation samples resulting from its randomization are shown in table 1 of the supplementary data. There were no relevant differences and the distribution of prognostic variables was homogeneous.

In addition, the main features of cohort 1 and cohort 2 were compared (table 2).

Table 3 shows the relationship between the variables proposed by the European guidelines and in-hospital mortality in the derivation sample (n=688). All variables, except ischemic stroke, cerebral hemorrhage, fungi, non-HACEK Gram-negative bacilli (Haemophilus spp, Aggregatibacter spp, Cardiobacterium spp, Eikenella spp, Kingella spp) and severe valve/prosthesis dysfunction, were statistically associated with in-hospital mortality in the univariable analysis.

Then, a multivariate analysis was undertaken (table 3). Independent predictors of in-hospital mortality were age, prosthetic valve IE, comorbidities, heart failure, renal failure, septic shock, Staphylococcus aureus, fungi, periannular complications, ventricular dysfunction, and vegetations. The model showed good discriminatory ability with an area under the ROC curve of 0.855 (95%CI, 0.825-0.885) and good calibration (figure 2A).

Figure 2.

Discriminatory performance and calibration of the model. A: ROC curve and plot comparing predicted and observed in-hospital mortality in the derivation sample. B: ROC curve and plot comparing predicted and observed in-hospital mortality in the internal validation sample. C: ROC curve and plot comparing predicted and observed in-hospital mortality in the external validation sample. HL, Hosmer-Lemeshow; RMSE, root mean square error; ROC, receiver operating characteristic.

(0.35MB).

The formula to predict in-hospital mortality was built by using the logarithms of adjusted OR from the predictive model:

Where z=– 6.288+0.033 x Age+0.602 x Prosthetic valve IE+0.485 x Comorbidity+1.210 x Heart failure+0.800 x Renal failure+1.742 x Septic shock+1.195 x Staphylococcus aureus+1.847 x Fungi+0.690 x Periannular complication+0.747 x Low left ventricular ejection fraction+0.850 x Vegetation.

The model was internally and externally validated with the internal validation sample from cohort 1 (n=314) and from cohort 2 (n=133), respectively. Internal validation showed an area under the ROC curve of 0.823 (95%CI, 0.774-0.873). The model predicted an in-hospital mortality of 30.7% (95%CI, 27.7-33.7) and observed mortality was 29.9% (figure 2B).

External validation showed an area under the ROC curve of 0.753 (95%CI, 0.659-0.847). The model predicted an in-hospital mortality of 26.4% (95%CI, 22.2-30.5) and observed mortality was 29.9% (figure 2C).

The model can be accessed as an informatic application via internet at ENDOVAL score web13 and via google play store (“ENDOVAL score”).

Diagnosis and treatment of IE is a clinical challenge. Early identification of patients with LSIE at high risk is crucial to change the natural course of the disease.8 Previous important research has focused on the prognosis of IE.1,4,5,7,14–16 Although some of these classic studies on IE present a very good overview of the disease, they have important methodological limitations. First, these studies did not differentiate between left- and right-sided IE episodes, despite having very different profiles and prognosis.1,4,14 Furthermore, most studies focused on evaluating a single or a limited number of prognostic factors.3,11,12,17–31 The European guidelines summarize the most important prognostic factors in an attempt to reflect current knowledge and help clinicians in their daily practice; however, the information is not sufficiently accurate and its practical usefulness is limited. We tested the prognostic power of these prespecified variables, as we consider that all of them have clinical importance and have the scientific support of the authors and reviewers of the guidelines.

Our group published a very simple prognostic stratification of patients with LSIE determined at admission and based on the presence of heart failure, Staphylococcus aureus, and periannular complications.15 Our new predictive model is a simple tool to help obtain a quick and accurate estimate of patient prognosis. This should not be regarded as definitive but as a complementary source of prognostic information that, together with other variables, will help clinicians decide whether and when surgery is indicated. It can be inferred from our results that in-hospital mortality risk can be assessed for the same patient at different time points in the course of the disease, but this hypothesis must be confirmed in prospective studies. In addition, the model also may help patients and families to obtain accurate information and a better understanding of the disease and its complications.

Our work has some strengths. This study includes only patients with definite LSIE. The number of episodes is high in a disease that has a low incidence, and information from 4 tertiary hospitals has been included. The information is homogeneous and of high quality. Finally, the study focused on the prognostic factors proposed by the European guidelines, and demonstrates their prognostic power for the first time. This methodology precludes the bias that could exist in our population in the selection of variables for the construction of the predictive model, and favors the generalization of our results.

This work also has some limitations. All centers are tertiary hospitals with cardiac surgery facilities and are leaders in IE management, which restricts the applicability of the model to hospitals with similar characteristics. Cohort 1 included patients between 2000 and 2017, a long period during which different forms of management have been tested, which could have limited the accuracy of the model. The external validation cohort is more recent, which could explain some of the differences between cohorts and could be considered as a methodological shortcoming. Although the good performance of the model in the validation cohort reinforces the clinical usefulness of our work, future external validations, particularly with larger sample sizes and different case-mix populations, would improve the applicability of the predictive model. The definition of variables in the European guidelines is sometimes somewhat simple and, at other times, includes small adaptations that could have limited the prognostic impact of those variables.

Finally, the inclusion of other prognostic variables may improve the predictive performance of the model; however, for the sake of simplicity and general applicability, we tested only variables proposed by the European guidelines. Future investigations will be necessary to validate the results and to explore the effect of including new variables.