Cenários de aplicação de técnicas de aprendizado de máquina para a predição de estágios da doença renal crônica em uma base de dados do sistema público de saúde do Brasil

Abstract

Chronic kidney disease (CKD) represents a significant public health concern in Brazil and globally. Characterized by the progressive and irreversible loss of kidney function, CKD has an estimated average prevalence of 9.5% among the global adult population. In Brazil, millions have been diagnosed with the disease, with the total reaching an estimated 17 million in 2017. Early interventions can slow disease progression and reduce the need for renal replacement therapies. Early detection of CKD across its six clinical stages is critical for appropriate clinical management and is the most effective approach to reducing treatment-associated costs. This thesis aimed to develop application scenarios for algorithms and machine learning (ML) techniques to predict CKD stages using a Brazilian public health database comprising over seven thousand records containing personal, socioeconomic, clinical, and laboratory information from patients. The proposal explored various combinations of predictor variables to serve as a basis for ML methods capable of predicting disease progression, thereby reducing reliance on traditional markers such as serum creatinine, commonly used in related literature. Based on these variable combinations, connected to clinical exams and personal data, this work also sought to identify the smallest viable set of predictor variables for the six stages of CKD. In developing the scenarios, three different dataset configurations were explored to derive distinct perspectives on content representation. Various methods were applied to infer missing data, along with techniques aimed at balancing the data, selecting relevant variables, and partitioning the data for classification purposes. Additionally, supervised ML algorithms with diverse theoretical foundations and objectives were employed to facilitate a comparative analysis of the results. The outcomes varied across the proposed scenarios. In the first scenario, classifications were performed using the random forest (RF) algorithm, with all approaches achieving high accuracy, except for the dataset excluding serum creatinine as a predictor variable. In the second scenario, which included 25 variables but excluded creatinine, the extreme gradient boosting (XGBoost) algorithm demonstrated high accuracy comparable to values reported in the literature, despite the latter’s inclusion of creatinine. In the third scenario, dataset imbalance was addressed using different methods, and classification was performed based on only three predictor variables. Although the overall results did not meet expectations, some findings were promising for detecting early CKD stages. In the fourth scenario, missing data inference was handled using the copula-based approach, but results were unsatisfactory. Lastly, the fifth scenario was the most comprehensive in terms of data organization, processing, selection, and classification; however, the new approaches did not lead to significant improvements in results. In conclusion, some of the developed scenarios successfully met the objectives outlined in this thesis, especially as they retained the use of traditional disease markers. The promising results may have potential applications in daily clinical practice and could assist in the early diagnosis of chronic kidney disease.