Construction and validation of a Klotho-based machine learning model for predicting all-cause mortality in chronic kidney disease

Abstract

Objective To develop and validate a machine learning (ML) model based on serum Klotho protein that can accurately predict all-cause mortality in chronic kidney disease (CKD) patients. Methods A retrospective cohort trial was conducted on all the non-dialysis adult patients diagnosed with CKD stages 1~5 in our department from February 7, 2012 to October 18, 2019. They were assigned into a training set and an internal validation set in a ratio of 7∶3. A total of 47 clinical features, including serum Klotho protein level, were used as variables to inform these models. Based on the training set, univariate Cox regression model was employed to screen out the possible risk factors for all-cause mortality, and Lasso-Cox regression model was further applied for the screening. Then multivariate Cox stepwise regression analysis was conducted to develop a nomogram risk prediction model for all-cause mortality, and the model performance was evaluated through internal validation. Results There were totally 400 patients enrolled in this trial, and 280 of them (including 52 dead and 228 survival) were assigned into the training set and other 120 (including 21 dead and 99 survival)into the validation set. For the constructed 5-year all-cause mortality risk prediction model, the area under the curve (AUC) value was 0.760 (95%CI: 0.676~0.844) in the training set and 0.788 (95%CI: 0.679~0.897) in the validation set, and the overall C-index was 0.755 (95%CI: 0.685~0.826) and 0.720 (95%CI: 0.614~0.826), respectively in the 2 sets. Univariate Cox regression analysis showed that age, history of cardiovascular disease(CVD), cystatin C(Cys-C), alkaline phosphatase (ALP), albumin, eosinophil (EOS) count, hemoglobin (Hb), complement C3, calcium, C-reactive protein (CRP), TNF-α and serum Klotho protein may be predictive factors for all-cause mortality (P<0.05). Multivariate Cox stepwise regression analysis finally screened age, albumin, complement C3 and serum Klotho protein as independent predictors (P<0.05). Based on these 4 predictors, a risk prediction model for all-cause mortality was constructed and validated. Conclusion A Klotho-based risk ML model for predicting all-cause mortality in CKD patients is successfully developed and validated. Advanced age is a risk factor, and higher albumin, complement C3 and serum Klotho protein levels are protective factors for all-cause mortality in CKD patients.

Keywords

• chronic kidney disease
• all-cause mortality
• prediction model
• machine learning