Using a large cohort from 14 cardiac centers, we developed and validated a nomogram model to predict AKI and evaluate subsequent adverse outcomes for patients undergoing cardiac surgery in China. The nomogram demonstrated adequate discrimination and calibration in both cohorts and showed a better performance when compared with other five scoring systems. Additionally, we stratified patients into low-risk, moderate-risk, and high-risk groups according to the score generated from the nomogram. Higher score group was associated with higher risks of death from all causes and MAKEs during 7-year follow-up. These findings suggested that the nomogram could serve as a risk-calculated tool for enhancing the risk stratification of CSA-AKI and its relevant clinical outcomes in Chinese cardiac patients.
Among 54 exposure variables, we identified 12 important risk factors associated with CSA-AKI. Variable selection procedure is one of the most important processes when constructing a prediction model. Logistic regression is commonly used. Traditional approach that included a set of variables to generate a model tends to lead to overfitting [21]. The feature selection strategies we applied in this study, which is a combination of machine learning and logistic regression method, provide insights on handing clinical data. LASSO is an alternative and effective option in handling high-dimensionality data. It penalizes magnitude of regression coefficients and excludes variables with a zero coefficient. A particular advantage of this technique is that it avoids both overfitting and overestimation during model derivation [18]. After this selection process, the strong predictors were identified in the final model. Several predictors identified in this study are consistent with previous researches, including age, diabetes mellitus, hypertension, critical preoperative state, Scr, surgery type, CPB time, and intraoperative RBC transfusion. Renal disease, infective endocarditis and PMV are new risk factors that have not been incorporated into those scoring systems. This indicated that, despite differing in races, AKI may have some common risk factors. However, these risk factors have different coefficients between races, and thus contributing different weights in the models.
We compared the performance of the five conventional models and found none of them demonstrated adequate power for predicting AKI in Chinese cardiac patients, both in terms of discrimination and calibration. Several factors may affect their wide application. First, the AKI definitions were various in the previous studies. In 2012, diagnostic criteria of AKI were revised significantly by the KDIGO organization. Therefore, the models established before 2012 may be imprecise and should be cautiously applied to current clinical practice [22]. Second, the Mehta score, Cleveland Clinic score, and SRI score were developed principally for AKI requiring renal replacement therapy (RRT-AKI), which is rare and not less severe stages of AKI, with incidences of 1.4-2.2% in their reports, obviously lower than the any-stage AKI rate (27.6%) in the present study. Ranucci et al. [23] validated the three models in a single-center study. They found that the three models showed excellent predictive ability for RRT-AKI, but were not well-performed for predicting non-RRT-AKI. Similarly, Che et al. [24] found that the Cleveland Clinic score and SRI score had poor classification (AUCs ranged from 0.516 to 0.673), and cannot be applied effectively in Chinese AKI patients. Third, most of previous prediction models only attached importance to preoperative variables. Given that the occurrence of AKI is a dynamic process and is particularly affected by procedure-related factors, we suggest that all preoperative, intraoperative, and early postoperative parameters should be screened for assessment during model derivation. Taken together, these data indicated that the study end point, race, and sample source are still pivotal factors and different models may be more suitable to apply to their specific populations.
Although AKI may be reversible, some of the patients develop mild or even transient AKI that could lead to CKD or adverse clinical outcomes [25, 26]. Particularly in patients with pre-existing renal disease (e.g., in older patients, high levels of Scr), AKI substantially accelerates the severity of kidney dysfunction and its progression to end-stage renal disease or adverse events [27, 28]. However, the pathophysiology and precise mechanism of AKI-to-CKD transition are complex and remain not fully understood. Notably, the nomogram model did not only predict CSA-AKI, higher score group was also associated with higher rates of mid-term death and MAKEs. These findings highlight the nomogram as a useful tool assisting in the risk stratification of AKI and mid-term outcomes. When considering the clinical implications, we suggest that this model may be useful in enriching patient cohorts for clinical trials or establishing benchmarks of cardiac surgical care. Using the nomogram may help in choosing preventive strategies in the perioperative management of patients. These strategies might include individualized blood pressure control, change of the surgical procedures (e.g., change from on-pump to off-pump surgery), reduced CPB time, and intensified hemodynamic monitoring and airway management (e.g., early extubation).
Our study has several strengths. We included 14 hospitals, and more than 11000 Chinese cardiac patients. Compared with previous study of AKI, defined as RRT-AKI, a definition of any-stage AKI was used in this study, thus extending the risk model to milder AKI patients. Besides, our study revealed that AKI is strongly associated with poor clinical prognosis; an AKI prediction model could also be useful for evaluating subsequent adverse outcomes of CSA-AKI. However, several limitations of this study should also be acknowledged. First, although the data were prospectively collected with adequate quality, the ascertainment or selection bias from a retrospective in design could not be fully avoided. Second, the model was developed based on routine variables extracted from the electronic health records. Therefore, the risk model can only apply the variables that have been collected. Third, although we developed and validated the model with a multicenter data resource, the model was not externally validated in other races or regions. It would be of caution to apply the model to other developing nations. External validation of the model in larger scale of Chinese patient cohorts is also needed.