Indeed, while the majority of the young patients were considered suitable for surgery, only the healthiest elderly were selected to undergo surgical treatment. Such a protective effect Pazopanib of aging could also be explained by the presence of young patients in the cohort, in whom the prevalence of post-operative degradation in renal function was very high. This interpretation was supported by the fact that when comparing middle-aged patients (40 to 75 y) to the elderly (>75 y), the effect of age was no longer significant (OR 0.80, 95% CI 0.55, 1.18, P = 0.26).Although AKI has been associated with a higher risk of mortality in patients with IE, the timing and different contributing factors of AKI have not been clearly explored so far.

In a Spanish multicenter observational study of 705 patients with left-sided IE, using multivariate analysis, G��lvez-Acebal et al. reported AKI to be associated with mortality [34]. In our series, we found that post-operative AKI was clearly associated with in-hospital mortality, whereas the association between pre-operative AKI (excluding patients receiving pre-operative RRT) and mortality was not significant. This suggests the importance of developing and evaluating perioperative strategies to prevent the occurrence of post-operative AKI. Interestingly, we observed a sharp difference in mortality between patients reaching stage 3 AKI and patients with stage 1 or 2, suggesting that all forms of AKI should not be considered equal in their severity in this setting.

Although we only observed a statistical association between AKI progression and mortality, we cannot exclude a lack of power in our cohort to show such an association with pre-operative AKI.The statistical analysis used some innovative tools such as SuperLearner [35] for prediction and TMLE [16,17] for estimation. The idea behind super learning is to optimize the prediction performances, accepting the fact that we do not know anything about the true shape of the underlying data distribution, so that every kind of parametric regression model would be biased. SuperLearner allows us to use a large library of candidate regression algorithms, parametric of data-adaptive, to honestly evaluate their prediction performance, and to build a new, tailored algorithm that is a combination of the best candidates. As expected from theory, we found that SuperLearner outperformed each candidate algorithm included in its library.

From such results, we expect SuperLearner to do the best Anacetrapib possible job to estimate the overall probability distribution of the outcome in our dataset. However, when looking at risk factors, we do not really care about the whole probability distribution of the outcome. In fact, we do care about a far less dimensional object, which is the distribution of the outcome given the level of a given potential risk factor.