Prediction of acute kidney injury after cardiac surgery: model development using a Chinese electronic health record dataset

Acute kidney injury (AKI), a common and potentially life-threatening clinical syndrome, is more and more frequent with increasing cardiac surgical volume in developed and developing countries. A meta-analysis reported an incidence of cardiac surgery-associated acute kidney injury (CSA-AKI) of 26.0–28.5% [1]. AKI not only severely affects in-hospital morbidity and mortality but also long-term prognosis. Patients who survive an episode of AKI after surgery are also at elevated risk of developing major adverse cardiovascular events, advanced chronic kidney disease, and all-cause death [2].

The current consensus definition of AKI depends on the increase of serum creatinine (Scr) and/or the presence of oliguria; however, this can lead to delayed diagnosis and treatment. Consequently, substantial efforts have been made to explore biomarkers or develop clinical prediction models in recent years. Several novel biomarkers were proposed to substitute Scr in the assessment of kidney function, such as NGAL, KIM-1, and DKK3 [3, 4]. However, an individual biomarker is inadequate in predicting AKI; that is to say, the pathophysiology of CSA-AKI is multifactorial and intricate. On the other hand, these biomarkers are costly and difficult to assay and are thus out of consideration by the majority of clinicians [5]. Clinical scoring systems (e.g., Cleveland Clinic score, Simplified Renal Index score, Mehta score, etc.) have been introduced into clinical practice for more than a decade [6‐8], while widespread adoption of these models would be challenging. First, these models are developed following the traditional logistic regression method. Their derivation requires the statistical assumption regarding a linear relationship between covariates and outcomes, with analytical model restricting to selection of a small set of parameters that are known to be clinically relevant. Second, most of these models perform traditional feature selection methods based on a small scale of exposure variables and identify common risk factors (e.g., age, diabetes mellitus, hypertension, cardiac function, surgery type, etc.). These risk factors reflect preoperative conditions and generally present inadequate predictive power in different races or populations [9].

Recent advancements in electronic health record (EHR) systems, data accessibility, and artificial intelligence have raised great interest in developing completely electronic data-driven machine learning (ML) models for predicting specific clinical outcomes. Several studies have used ML to predict inpatient AKI using EHR data. For instance, a continuous deep learning algorithm can predict 55.8% of all inpatient episodes of AKI up to 48 h in advance and over 90% of all AKI patients requiring subsequent renal replacement therapy (RRT) [10]. Therefore, our first objective was herein to develop three tree-based ML models to predict CSA-AKI by incorporating preoperative, intraoperative, and early postoperative data from the EHR of Nanjing First Hospital. In addition, to uncover the “black-box” of ML, SHapley Additive exPlanation (SHAP) values were utilized to explain the ML model and evaluated individual variable prediction [11]. Given the generalizability of the linear model, a nomogram model was finally constructed based on the logistic regression analysis.

In this study, we applied the SWSFS technique to screen for clinical characteristics and developed a series of models to optimize AKI prediction following cardiac surgery. Using SWSFS, we identified 14 important risk factors associated with AKI among the 193 clinical variables, thus boosting efficiency by incorporating preoperative, intraoperative, and early postoperative data from EHR. The performance of the tree ensemble ML algorithms (XGBoost, RF, and DF) was clinically satisfactory, with AUROC ranging 0.857–0.881 in the validation set. In addition, we constructed a nomogram model for AKI which also presented good performance both in terms of discrimination (AUROC 0.824) and calibration (Brier score 0.144). Our study highlights the value of EHR data in the evaluation of AKI. These important perioperative factors might be helpful in providing individualized preventive strategies and delivering proper treatments in the management of AKI after cardiac surgery.

In addition to some well-known risk factors (e.g., age, gender, hypertension, diabetes mellitus, Scr, etc.) that have been identified by previous studies, most variables are novel predictors for AKI risk prediction. Intraoperative factors reflect acute physiological responses during surgery and play pivotal roles in the development of AKI, particularly the unique physiological perturbations of CPB. In this study, we generated developmental trajectories to describe the course of factors over time. Two specific time point measurements of CVP (T3 and T4) were found to be strong predictors, highlighting the effect of intraoperative venous congestion on renal function. Traditionally, CSA-AKI is considered to be caused by renal hypoperfusion due to hypotension, inadequate perfusion flow, or renal ischemia [22, 23]. This concept has been challenged by accumulating evidence that elevated CVP is a more powerful hemodynamic determinant than mean perfusion pressure for the development of postoperative AKI [24]. More recently, Lopez et al. [25] uncovered that higher levels of CVP during cardiac surgery were independently associated with higher odds of AKI. They also demonstrated that venous congestion is more accurate than hypotension in predicting AKI. CVP-induced AKI can therefore be regarded as “congestive kidney failure”. When examining the trajectory pattern of perfusion flow, the separate clusters of perfusion flow at T3 measurement in our trajectory analysis implied that some patients (high-level cluster) might go through cardiac insufficiency, hemodynamic instability, or other unstable intraoperative conditions after aortic declamping and require additional mechanical assistance before weaning off from CPB. The positive effect of ultrafiltration volume on AKI risk may be another latent indicator of the congestive state of the body, as CUF is used for fluid removal to reduce fluid overload [26]. Taken together, these findings demonstrate that not only renal ischemia but also renal congestion plays a vital role in worsening kidney function. Our study suggested that the clinician may need to pay more attention to hemodynamic changes, in particular during the cardiac resuscitation period. Although adequate urine output does not assure normal kidney function as a result of non-pulsatile flow and cold-induced diuresis, the presence of oliguria typically indicates an acute response to renal hypoperfusion. This result is consistent with the study by Tseng et al. [27], who observed that intraoperative urine output was the most influential feature in predicting AKI.

Early postoperative laboratory biomarkers could reflect the acute pathophysiology of kidney injury. In this study, we identified four laboratory biomarkers (intubated PaO₂/FiO₂ ratio, serum potassium, hemoglobin, and LDH) associated with CSA-AKI. These biomarkers reflect the patient’s overall disease severity. For example, hypoxemia (low intubated PaO₂/FiO₂ ratio) is a severe complication after cardiac surgery and has been shown to be highly related to prolonging mechanical ventilation, respiratory complications, and in-hospital death [28]. The loss in glomerular filtration rate reserve usually leads to electrolyte disorder. It was observed that some patients experienced abrupt deterioration in renal function and decreased glomerular filtration capacity in the early postoperative period, resulting in fluid overload and elevated serum potassium levels [29]. LDH is abundant in the kidney, heart, liver, and muscle and is, therefore, most commonly measured to detect tissue damage as well as disease severity of critical patients [30]. Although LDH acts as a nonspecific biomarker for kidney injury, it demonstrates adequate predictive value for AKI risk prediction in several clinical settings [31, 32]. On the other hand, the elevated LDH in the immediate postoperative period may be an indicator of hemolysis from CPB, and CPB-induced hemolysis is associated with the development of AKI [33].

Previous studies have concluded that inflammation, oxidative stress, and endothelial dysfunction are central components of the pathogenesis of AKI. NLR, a promising marker of inflammation, has been identified as a novel predictor of AKI [34]. Moreover, it has emerged as a potential biomarker for lethal outcomes and adverse events in patients undergoing cardiac surgery [35, 36]. Interestingly, lower baseline HDL levels were independently associated with an increased risk of AKI after cardiac surgery. This relationship was also observed in Smith et al.’s study [37]. Systematically, high HDL levels inhibit systemic inflammation and reduce oxidative stress via acting as receptors of prooxidant lipids and associated antioxidant enzymes and therefore play a role in the pathogenesis of AKI [37‐39]. However, the natural effect of HDL may be pharmacologically modified by traditional preoperative lipid-lowering treatment in cardiac patients; novel pharmacologic agents with the potential for improving HDL function are warranted. Collectively, the pathophysiology of CSA-AKI is complex and multifactorial. Our study identified a set of important factors attributed to CSA-AKI such as venous congestion, renal hypoperfusion, inflammatory response, metabolic disorder, and baseline renal function. However, many of these factors are modifiable. Using these factors may therefore provide a basis for early diagnosis, prevention, and treatment strategies in the preoperative, intraoperative, and early postoperative management of AKI patients, such as optimizing hemodynamic status, intensified early postoperative biomarkers monitoring, and individualized blood glucose and lipids management.

Before implementing ML models in clinical practice, their predictive power must be validated in different clinical settings. Previous studies developed ML models using all features as input variables [27, 40]. However, inclusion of numerous features makes the models much more complicated and difficult to validate in additional representative datasets, as most variables are irrelevant to AKI classification. Besides, many ML algorithms exhibit a decrease in predictive power when the number of variables is significantly higher than optimal [41]. For example, as outlined in Additional file 1: Fig. S4, these ML models were further evaluated with all variables as input variables, and no improvement in model discrimination was noted. Our study confirms the significance of feature selection in ML applications. We applied SWSFS technique to determine the optimal and minimum size of features, thus increasing the efficiency and usefulness of the models for further validation.

Our study has several strengths. First, a major barrier to the widespread use of ML models is their correct interpretation, as a true “black-box” can hardly be accepted by clinicians or decision-makers. As an additive model explainable approach, SHAP analysis is seldom employed in ML application. We utilized the SHAP values to demystify the ML, providing a “white-box” AKI prediction model that allowed a quick comprehension of the effect of a single feature on the model’s prediction. This explainable or “white-box” predictive technique may be helpful in ML transportability across hospitals. Second, the three tree-based ensemble learning algorithms demonstrated high calculating efficiency and may be adapted to certain medical working environments. Indeed, XGBoost and RF have the advantages of being trained quickly and providing reliable feature importance estimates and are increasingly emphasized as competitive alternatives to traditional regression methods. In addition, both XGBoost and RF algorithms are bootstrapping method applications, which can improve predictive power when available datasets are small. Notably, for the first time to our knowledge, we used DF to predict AKI after cardiac surgery. As an alternative to the deep learning framework, DF improves the robustness of the traditional deep learning method working on small-scale clinical data and provides an effective solution for binary classification. Third, given the importance of intraoperative factors on AKI development, we applied GBTM method to handle intraoperative longitudinal data in which the dynamic process of variables over time could be clearly observed. GBTM identifies distinctive clusters of individual trajectories within the population, determining the subgroup patients at different risk levels; it provides a critical time point for clinical decision making. By monitoring intraoperative hemodynamic parameters, it is expected to build more accurate dynamic early warning systems that can help clinicians timely identify patients at risk of postoperative AKI. Fourth, the incidence of AKI in our study was generally in line with the report from a meta-analysis [1], indicating that our patient cohort is representative of cardiac patients in general. Finally, the dataset contained only a few missing data because most of the missing values were filled in during a second manual review of EHR. Therefore, the impact of missing data on the prediction models is negligible.

However, several limitations should also be considered. First, the models were developed based on the dataset derived from a single center with patients undergoing on-pump cardiac surgery. Consequently, before the models can be implemented in clinical practice for new patients, their predictive performance would require training and evaluation on other races, nationalities, or additional datasets. Second, postoperative urine output (less than 0.5 mL/kg/h for 6 h) was not used to define AKI due to its unavailability in the majority of patients. However, we were unlikely to miss vital clinical patients as urine output may have been maintained by diuretics. Moreover, due to intensified monitoring in ICU, persistent oliguria is uncommon and transient oliguria may simply imply insufficient volume resuscitation. Third, although SHAP values can explain most traditional ML models, they cannot explain the DF algorithm. In other words, the correct interpretation of DF remains challenging. We are working on developing a more advanced algorithm for DF that could provide variable importance estimation. Fourth, we did not include traditional scoring systems to make model comparisons, because the most extensively used and robust models for AKI are those designed for AKI requiring RRT. Given the high incidence of subclinical AKI and its strong association with adverse outcomes, more advanced models should be developed to predict any-stage AKI after cardiac surgery.

Taken together, our study highlights the potential of tree-based ensemble methods in generating robust AKI prediction tools. By contrast to clinical scoring systems or biomarkers, ML algorithm is a completely data-driven prediction tool. As for clinical practice, a web-based online tool can be developed to facilitate the application of the ML model. By inputting the values of features of a patient, the online tool would estimate the probability of developing AKI. With advances in EHR, it can be easily transferred to the EHR system to calculate AKI risk by automatically reading features.

ML algorithms for clinical data analysis have revolutionized the traditional way of conducting cardiovascular research. As clinicians continue to gather significant amounts of patient data through EHR, more novel associations between specific features and AKI will be identified. Future work is ongoing on the development of more advanced ML algorithms and EHR systems for real-time adjustment of the AKI risk after cardiac surgery, which in return will optimize treatment and enhance prognosis.

In this study, based on the 16 important perioperative predictors, we successfully established three tree-based ML models and a nomogram model to optimize CSA-AKI risk prediction.