Cluster and Rule Based Prediction in Hemodialysis: Adaptivity vs. Robustness

Reliable prediction methods are an essential part of any powerful decision support system in medicine. They are crucial to support medical decision making with respect to the prognosis for patients. This paper reports on the development of cluster and rule based prediction methods to forecast cardiovascular risk of hemodialysis patients. The particular focus was on the prediction of the interventricular septum (IVS) thickness of the heart as an important quantitative indicator to diagnose left ventricular hypertrophy which is characterized by an enlarged septum. The study was based on data from 63 long-term hemodialysis patients of the KfH Dialysis Centre in Jena, Germany. Including the target variable IVS, 42 potentially influential variables, each for the first and fourth treatment year, were taken into account. They cover patient data, anamnesis and diagnosis data, laboratory and medication data. The approach applied involves three major steps: data based clustering, cluster based rule extraction and rule based prediction. Methods used include crisp and fuzzy algorithms and the combination of both. At each step, logical and medical validation was carried out with particular focus on adaptivity of data and methods and robustness of results (data and method tests). The approach applied and validated in this study proved successful for the cluster and rule based prediction of single quantitative variables for individual patients from other variables relevant to the particular medical problem.