Machine learning algorithms for the prediction of adverse prognosis in patients undergoing peritoneal dialysis

Abstract

Background: An appropriate prediction model for adverse prognosis before peritoneal dialysis (PD) is lacking. Thus, we retrospectively analysed patients who underwent PD to construct a predictive model for adverse prognoses using machine learning (ML).

Methods: A retrospective analysis was conducted on 873 patients who underwent PD from August 2007 to December 2020. A total of 824 patients who met the inclusion criteria were included in the analysis. Five commonly used ML algorithms were used for the initial model training. By using the area under the curve (AUC) and accuracy (ACC), we ranked the indicators with the highest impact and displayed them using the values of Shapley additive explanation (SHAP) version 0.41.0. The top 20 indicators were selected to build a compact model that is conducive to clinical application. All model-building steps were implemented in Python 3.8.3.

Results: At the end of follow-up, 353 patients withdrew from PD (converted to haemodialysis or died), and 471 patients continued receiving PD. In the complete model, the categorical boosting classifier (CatBoost) model exhibited the strongest performance (AUC = 0.80, 95% confidence interval [CI] = 0.76-0.83; ACC: 0.78, 95% CI = 0.72-0.83) and was selected for subsequent analysis. We reconstructed a compression model by extracting 20 key features ranked by the SHAP values, and the CatBoost model still showed the strongest performance (AUC = 0.79, ACC = 0.74).

Conclusions: The CatBoost model, which was built using the intelligent analysis technology of ML, demonstrated the best predictive performance. Therefore, our developed prediction model has potential value in patient screening before PD and hierarchical management after PD.

Keywords: Machine learning; Peritoneal dialysis; Prediction model; Prognosis.

Fig. 1

The ROC curves of the models. A The complete model ROC curves of five algorithms. The CatBoost algorithm had the highest AUC of 0.80. B The compact model ROC curve of the optimal algorithm. The algorithm with the best performance in the complete model was adjusted, and the top 20 variables with the strongest correlation were selected to create a compact model with an AUC of 0.79. ROC, receiver operating characteristic; AUC, area under the curve

Fig. 2

The SHAP values of the Catboost model. A The variables with the strongest correlation in the prediction model were ranked, and the top 20 were obtained. B The SHAP value of these variables. SHAP, Shapley additive explanation

Fig. 3

Two examples of model interpretation. A A patient who was predicted to be unfit for PD failed after a short period of PD. B A patient predicted to be suitable for PD succeeded for over two years and continued PD for five years