Identifying risk of progression for patients with Chronic Kidney Disease using clustering models

Abstract

Chronic Kidney Disease ("CKD") and its comorbidities, diabetes, hypertension and cardiovascular disease ("CVD"), are frequently measured by routine procedures and lab tests, creating a large amount of historical data about this patient population. In this paper, we conducted a retrospective study based on Electronic Health Records ("EHR") data, in order to identify patterns in the development of CKD. In particular, we used a clustering approach to quantifiably identify diabetic patients who are at risk of progressing to advanced stages of CKD. Using values from routine measurements and lab tests such as systolic blood pressure, diastolic blood pressure, body mass index ("BMI"), Hemoglobin A1c ("HbA1c"), triglycerides and high density lipid cholesterol ("HDL cholesterol"), patients were classified into four clusters with a distinct separation between the cluster with the best value for each lab test and a cluster with the worst value for each lab test. We used lab values from each subsequent visit to calculate a progression score using the distance to the best and worst clusters, which indicated whether a patient's health was improving or deteriorating. We believe that this approach holds promise for future tools, as it is able to provide an ordered list of patients who are at greater risk of deterioration and should benefit from intervention by healthcare providers. The conclusions made in this paper are aimed at enabling timely monitoring and earlier intervention for patients that are associated with higher possibility of CKD progression.

Cluster Analysis of Mixed and Missing Chronic Kidney Disease Data in KwaZulu-Natal Province, South Africa

“…In addition, though it was stated that Euclidean distances were used in heatmap and hierarchical clustering, it is not clear how this was achieved, given the mixed nature of their dataset. A similar claim was made in Lenart et al [60] who also performed cluster analysis on a mixed and missing longitudinal CKD dataset with 10 variables. Of 10,014 observations, missingness was handled by complete case analysis which resulted in the analysis being performed on 2,696 observations.…”

Prediction of kidney disease using machine learning algorithms

“…Because it is able to produce an ordered list of individuals who are at a higher risk of deterioration and who might benefit from intervention by healthcare practitioners, we feel that this method has the potential to hold promise for the development of future tools. The findings presented in this study are intended to facilitate faster monitoring and earlier management for individuals who have a greater risk of developing chronic kidney disease (CKD) as a result of their medical history [9].…”