MACHINE LEARNING EMPOWERED SYSTEM FOR SMART CHRONIC KIDNEY SCREENING

Abstract

CKD is a common and sometimes fatal disease in characteristic of the low resource settings where early diagnosis tools are rare. The following project focuses on clinically feasible ML methods to detect essential features of CKD to inform the development of cost-effective screening tools. Our study will use only clinically obtained data and in order to derive useful and reliable predictive markers of early CKD we will rely on demographic, biochemical and physiologic characteristics observes in routine clinical practice. Here, our strategies include model interpretability, which guarantees comprehensible results, scalability to accommodate future enlargement, and resource constraints in healthcare settings. It also analyses feature reduction approaches that allow identifying the most relevant variables with less reliance on potentially costly diagnostics. The vision is to have a solid, cost-effective and implementable diagnostic tool for early identification of CKD hence enhancing patient prognosis and curtailing the effect of the disease across the world. In this way, our methodology uses various types of different data as clinical data from studies and health surveys data; demographic, biochemical, and physiological data. By using these datasets, we perform exploratory data analysis that investigates the relationship between patient characteristics and CKD stages; and builds accurate ML models to estimate the risk of CKD. Together with RFE and PCA, most relevant predictor of CKD are selected and extracted. These predictors are tested again using cross-validation methods so as to establish the credibility and portability of the model. The ML algorithms developed are the supervised learning algorithms: decision trees, support vector machines, Random forests, and gradients boosting. Further, we examine the role of advanced XAI methodologies to improve the interpretability of the models, so that the healthcare practitioners using it, are aware of the decision-making logistics of the models. Particular attention is paid to developing models that require few computations; this allows for the method’s implementation on low-power devices: smartphones, portable diagnostic devices, etc.