The prediction of in-hospital mortality in chronic kidney disease patients with coronary artery disease using machine learning models

In the past few decades, chronic kidney disease (CKD) has become increasingly prevalent among various countries and regions around the world, increasing the enormous financial burden of many countries [1]. A major cause of death among patients with chronic kidney disease is cardiovascular disease [2], and CKD patients with coronary artery disease (CAD) have a poorer prognosis than CKD patients without CAD [3, 4]. Moreover, the risk factors of patients with CKD combined with CAD are much different from those with only CAD [5]. Some studies demonstrated that atherosclerosis is the leading cause of death in advanced CKD patients with CAD, especially end-stage renal disease (ESRD) patients [6]. In addition, the pathogenesis of CKD patients with CAD has not been clearly elucidated [7]. Thus, the present indicators and prediction models perform poorly in predicting clinical outcomes for CKD patients with CAD.

Machine learning (ML) is a cutting-edge technology with the rapid development of artificial intelligence [8]. Compared to the traditional statistical method, ML has better clinical predictive accuracy and performance with faster processing speed [9]. With the development of the online public standard database, such as the Medical Information Mart for Intensive Care IV (MIMIC-IV), ML has increasingly penetrated the medical analysis field [10]. However, a few ML algorithms focused on the mortality prediction of CKD patients with CAD.

The purpose of our study is to (1) construct novel predictive models based on the various machine learning algorithm for in-hospital mortality of patients with CAD and CKD in intensive care units (ICU); (2) select an ML model with the best predictive performance and clinical value; and (3) validate these ML models via external set from the Telehealth Intensive Care Unit Collaborative Research Database (eICU-CRD) database.

Patients with CKD and CAD became more and more popular in recent decades. And mortality in patients suffering from these two conditions is twice as compared to patients with CAD alone [4]. Despite the increased incidence and incredibly lethal, their patients were excluded from most clinical trials due to the disease complexity and treatment conflicts. To date, factors associated with the prognosis in CKD patients with CAD were not clear and current risk stratification tools could not be applied to these patients. With the development of artificial intelligence, accurate prediction of these complex conditions could be achieved using machine learning methods.

MIMIC-IV and eICU-CRD were large-scale and high-quality databases performed in many crucial pieces of research in recent years. In this retrospective study, CKD patients with CAD admitted to ICU were extracted from MIMIC-IV to develop predictive models for in-hospital mortality via various ML algorithms. The GBDT model outperformed the predictive performance of seven other ML algorithms, including LR, RF, Decision Tree, KNN, SVM, NN, and XGBoost, according to the features selected by the Boruta algorithm. Next, the SHAP method was conducted to explain GBDT visually, ensuring clinical interpretability and facilitating the utilization of the prediction model. The performance and clinical application value of GBDT were also validated by an external set from the eICU-CRD database. This is the first prediction method especially for CKD patients with CAD to evaluate the in-hospital mortality with precise efficiency in two large cohorts, which means good generalization to extend to clinical practice.

Depending on the visualization technique SHAP, our study identified several crucial variables related to the in-hospital mortality of patients with CKD and CAD in the ICU. This study identified a factor strongly associated with the in-hospital mortality observed in our study which was serum phosphate. Previous studies have shown that elevated serum inorganic phosphorous (P) is tightly associated with cardiac death in CKD patients [15]. A national study illustrated that hyperphosphatemia could lead to a predisposition to metastatic calcification and the development and progression of secondary hyperparathyroidism, which may contribute to the abundant morbidity and mortality of patients with ESRD [16]. Another research with a 2-year follow-up also identified strong relationships between hyperphosphatemia and cardiac causes of death in hemodialysis patients [17]. Moreover, a cross-sectional study showed elevated serum levels of P were significantly related to calcified coronary atherosclerotic plaque detected by cardiac computed tomography, even in patients with normal kidney function [18]. The previous studies exhibited the significance of P in prognosis in CKD patients. In our study, we focused on CKD patients with CAD and showed that serum P was a strong predictor of in-hospital mortality. Therefore, phosphate is a promising therapeutic target to improve the clinical outcome in CKD patients with CAD. Both dietary and pharmacological therapeutic strategies should be used to reduce of serum phosphate levels to prevent hyperphosphatemia in CKD patients with CAD.

Whether Coronary Artery Bypass Grafting (CABG) or PCI is the better approach for revascularization of CAD in CKD patients was still controversial. Several observational studies reported CABG was associated with lower mortality than PCI in CKD patients [19‐21]. But the Coronary REvascularization Demonstrating Outcome Study in Kyoto PCI/CABG Registry Cohort-2 study showed the risk of all-cause death was similar between PCI and CABG in ESRD patients requiring dialysis [22], which was consistence with the result of ISCHEMIA-CKD research [3]. Another meta-analysis also pointed out that patients with stage 3–5 CKD who underwent either approach to revascularization did not experience significant differences in mortality. However, CABG significantly reduced the myocardial infarction risks and required fewer additional revascularization procedures [23]. Different results in these studies might be attributed to different study participants, some focused on advanced CKD patients, while others focused on ESRD patients. Our study included patients with all staged CKD, ML visible results showed that both PCI and CABG were beneficial to the prognosis of CKD patients with CAD, and CABG was a more critical feature than PCI to the in-hospital mortality in those patients in ICU.

A growing number of machine learning applications in cardiovascular medicine have been made possible by the development of artificial intelligence [24, 25]. Using machine learning, it has been possible to predict death risk among CAD patients more accurately than before. Motwani et al. constructed a boosted ensemble algorithm combining clinical and coronary computed tomographic angiography (CCTA) to predict 5-year all-cause mortality with higher AUC (0.79) than clinical or CCTA metrics alone [26]. Silva et al. established a prognostic model using health conditions, including age and maximal exercise capacity, to precisely predict the mortality of CAD patients via the survival tree (ST) algorithm (C-index 0.729) [27]. In addition, Pezel and colleagues developed multiple fractional polynomial algorithm ML models, including 31,752 consecutive patients, to predict 10-year death [28]. This ML model also has a higher prognostic value than traditional clinical or Cardiac Magnetic Resonance scores (AUC 0.76). However, the mechanism of CKD combined with CAD is more complex and harder to explain than the mechanism of CAD alone [4]. For example, statin lipid-lowering therapy is still contradictory in improving the prognosis of patients with ESRD and CAD [29]. Predictions based on the traditional model cannot be made with reasonable accuracy and comprehensiveness for patients suffering from such complex diseases [5, 30]. For this reason, machine learning is of great significance.

The GBDT algorithm, also known as the multiple additive regression trees, has more accurate predictive ability and sophisticated algorithms than the LR, decision tree, and random forest algorithms [31]. It has many nonlinear transformations and solid, expressive ability, and does not require complex feature engineering and transformation [32]. The XGBoost model, a modified GBDT algorithm, could cope efficiently and flexibly with missing data and combines weak predictors to produce accurate predictions [33]. The no free lunch theorem (NFL) illustrates that the expected performance of each learning algorithm is the same if all possible problems are considered, which means there is no single, universal best machine learning algorithm for every situation [34]. Among eight ML models, the GBDT model performed the best clinical predictive value in in-hospital mortality risks in this kind of patient.

In conclusion, machine learning algorithms can be reliable tools for accurately predicting the in-hospital mortality risk for CKD patients with CAD in the ICU. GBDT technology had the best predictive performance, which may provide optimal resource allocation and reduce in-hospital mortality by tailoring precise management and implementing early interventions.