A Klotho-Based Machine Learning Model for Prediction of both Kidney and Cardiovascular Outcomes in Chronic Kidney Disease

Abstract

Introduction: This study aimed to develop and validate machine learning (ML) models based on serum Klotho for predicting end-stage kidney disease (ESKD) and cardiovascular disease (CVD) in patients with chronic kidney disease (CKD).

Methods: Five different ML models were trained to predict the risk of ESKD and CVD at three different time points (3, 5, and 8 years) using a cohort of 400 non-dialysis CKD patients. The dataset was divided into a training set (70%) and an internal validation set (30%). These models were informed by data comprising 47 clinical features, including serum Klotho. The best-performing model was selected and used to identify risk factors for each outcome. Model performance was assessed using various metrics.

Results: The findings showed that the least absolute shrinkage and selection operator regression model had the highest accuracy (C-index = 0.71) in predicting ESKD. The features mainly included in this model were estimated glomerular filtration rate, 24-h urinary microalbumin, serum albumin, phosphate, parathyroid hormone, and serum Klotho, which achieved the highest area under the curve (AUC) of 0.930 (95% CI: 0.897-0.962). In addition, for the CVD risk prediction, the random survival forest model with the highest accuracy (C-index = 0.66) was selected and achieved the highest AUC of 0.782 (95% CI: 0.633-0.930). The features mainly included in this model were age, history of primary hypertension, calcium, tumor necrosis factor-alpha, and serum Klotho.

Conclusion: We successfully developed and validated Klotho-based ML risk prediction models for CVD and ESKD in CKD patients with good performance, indicating their high clinical utility.

Keywords: Cardiovascular disease; Chronic kidney disease; End-stage kidney disease; Machine learning; Prediction model.

Fig. 1.

Heatmap of correlation between variables and the Mantel test. The heatmap in the top right-hand corner shows the pairwise correlations between the 47 variables, with a color gradient denoting the Pearson correlation coefficient. The magnitude and direction of the correlation are reflected by the size (larger is stronger) and color (red is negative and blue is positive) of the circles, respectively. The Mantel test in the bottom left shows the correlation between the two endpoints (ESKD, CVD) and the other variables. The statistical significance or lack of significance of the correlation between the variables and the endpoints is reflected by the color of the network connecting line (gray represents no significance [Mantel p ≥ 0.05] and the rest of the colors are significant [Mantel p < 0.05]). The magnitude of the correlation between the variables and the endpoints is reflected by the width of the network connecting lines (the thicker the stronger). When p < 0.05, the greater the Mantel r, the greater the effect of the variables on ESKD or CVD.

Fig. 2.

a Box-plot of the predictive accuracy of the five models for ESKD. b Risk factor screening for ESKD. The two dotted lines in the figure refer to the values of lambda.min and lambda.lse, respectively. c Risk prediction nomogram for ESKD. This figure is able to predict 3-, 5-, and 8-year renal survival in patients with CKD when combined scores for each risk variable. Each quantitative variable has a specific value that corresponds to a specific point. Each point extends a vertical line upward to the intersection of the “Points,” which corresponds to the value of the corresponding score and the total score of all variables can be found on the “Sum of all points” with corresponding coordinates. A vertical line is drawn from the “Sum of all points” coordinate and the value corresponding to its intersection with the “probability of 3-, 5-, and 8-year survival” coordinate is the probability of kidney survival.

Fig. 3.

Predictive performance of the ESKD model. a ROC curve in the training set. b Calibration curve in the training set. c DCA curve in the training set. d ROC curve in the validation set. e Calibration curve in the validation set. f DCA curve in the validation set.

Fig. 4.

a Box-plot of the predictive accuracy of 5 ML models for CVD. b Parameter testing for CVD. It illustrates the trend of the out-of-bag (OOB) error rate with the increasing number of trees. The error rate stabilized at approximately ntree = 600, indicating that further increases in the number of trees did not significantly improve model performance. c Risk factor screening for CVD. It provides insights into the importance of the variables considered. d Risk prediction nomogram for CVD. Risk scores are calculated in the same way as above.

Fig. 5.

Predictive performance of the CVD model. a ROC curve in the training set. b Calibration curve in the training set. c DCA curve in the training set. d ROC curve in the validation set. e Calibration curve in the validation set. f DCA curve in the validation set.