Metabolômica integrada ao aprendizado de máquina para o diagnóstico e predição da progressão da doença renal crônica

Abstract

INTRODUCTION: Chronic kidney disease (CKD) is a global public health problem, and currently available therapeutic approaches have limited effectiveness. This highlights the need for new strategies aimed at both early diagnosis and reducing CKD progression, particularly the use of omics approaches, especially metabolomics, integrated with machine learning (ML). This integration can aid in building models for diagnosis, prognosis, and monitoring of CKD progression, as well as overcoming the limitations inherent in conventional methods. OBJECTIVE: To characterize the serum metabolomic profile using high-performance liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) and to develop ML-based models for the diagnosis and prediction of CKD progression. METHODS: The study included individuals without CKD (control group); individuals with CKD (between stages 2 and 4) undergoing conservative treatment (intermediate group); and individuals on hemodialysis (HD group). A total of 148 participants were included: control group (n=55), intermediate group (n=32), and HD group (n=61). Sociodemographic variables, behavioral factors, clinical parameters, anthropometric measurements, and biochemical data were evaluated. Analyses were performed using a high-performance liquid chromatography-mass spectrometer. Statistical analyses were performed using Orange (Version 3.37.0), STATA (version 15.1), and MetaboAnalyst 6.0 software. A p-value < 0.05 was considered statistically significant. After identifying differentially expressed metabolites, machine learning (ML) algorithms were employed to differentiate individuals with chronic kidney disease (CKD) from healthy participants. The algorithms used were Logistic Regression (LG), Support Vector Machine (SVM), Gradient Boosting (GB), Neural Network (NN), and Random Forest (RF). Performance was evaluated through ROC (Receiver Operating Characteristic) curve analysis, area under the curve (AUC) calculations, F1 score, accuracy, precision, and Matthews correlation coefficient (MCC). After training on different algorithms, all diagnostic and CKD progression prediction models underwent external validation. Shapley Additive Explanations (SHAP) were applied to interpret the ML models. RESULTS: A total of 759 metabolites were identified. Differential expression analysis identified 41 distinct metabolites for diagnostic prediction and 29 for progression prediction. These differential metabolites were used to train ML models employing the algorithms. After training, external validation was performed using the previously trained algorithms. The results demonstrated that the SVM and NN models appear to have been the most effective for both diagnosis and progression prediction of CKD. CONCLUSION: Our findings show that the ML models applied to metabolomic data exhibited high performance, especially the SVM and NN models. Therefore, the combination of serum metabolomics and ML proved to be a robust strategy to improve early identification and monitoring of the disease.