Cálculo de estadiamento, preditores de risco por meio da comparação de algoritmos de machine learning como estratégia de rastreio à doença renal crônica

Abstract

One of the biggest challenges of Modern Medicine is working with the quantity and availability of data, as the decision for a certain treatment or confirmation of diagnosis is supported by various methodological approaches. Among the various existing pathologies that are subject to and relevant to an approach through the use of generative Artificial Intelligence (AI) are Chronic Kidney Diseases (CKD), with multiple causes in the organ or in the Urinary System combined with multiple risk factors previously existing in the patient. Faster and more accurate testing and evaluation of results are extremely necessary. That said, the objective of this research was to apply three supervised Machine Learning (ML) techniques - Decision Tree (DT), Multilayer Perceptron (MLP), Kolmogorov-Arnold Network (KAN) and compare their performances (accuracy, precision, recall, F1 score and ROC-AUC) of predisposing variables and correlated with CKD staging, using the Explainable Artificial Intelligence (XAI) metric. Using 24 variables available in The UCI Machine Learning Repository (2015); with 400 white men, through exploratory and transversal analysis. As a result, it was obtained that the MLP model outperformed the others, achieving the highest average precision (87.2%) and ROC-AUC score (0.98), indicating strong predictive capacity for this effectively studied population, however KAN provided a balance between strong performance and moderate explainability, showing promise for further development as an interpretable model for CKD prediction. This research highlights the importance of considering accuracy and interpretability when selecting models for healthcare applications and positions KAN as a potential XAI model for predicting CKD stages. Future research could focus on improving the interpretability of KAN and exploring its use in other medical datasets.