Classification methodology for mitigating the effects of unreliable class assignments in high dimensional biomedical data

点击翻译按钮获取中文摘要

Classifying biomedical data involves finding a mapping (relationship) from patterns (e.g., data relating to some type of tissue) to a set of classes (e.g., disease states). Patterns are represented by features (e.g., concentrations of biological compounds) and class labels are assigned using a reference test (e.g., a medical expert's analysis of tissue being "normal" or "abnormal"). This process often suffers from three significant challenges: the number of features in a pattern is high; the number of patterns is low; and the reference test may be unreliable. The first two challenges, known collectively as the "curse of dimensionality", cause an inability to find robust, general solutions. This is often addressed by reducing, in some fashion, the number of features; however, a direct correspondence back to the original features is necessary for medical experts to make informed judgments about the mapping's predictive power. While reference tests may be well-established benchmarks, they are seldom perfectly accurate and sometimes improperly applied. Nevertheless, any strategy that compensates for this imprecision must ensure that the mapping is correctly validated against the benchmark.