Specıfyıng the Boundarıes of Gray Zone in Dıagnostıc Tests Wıth Informatıon Crıterıa
Özet
The decision-making process in medicine is a crucial subject due to the classification of subjects as healthy or diseased. Mostly, it concludes with binary outcomes, such as whether the person has a condition or not. The various information from subjects is taken, such as the complaints, family history, symptoms, or laboratory tests (known as diagnostic tests) to rule-in or rule-out the disease. Due to their advantages such as being cost-effective or rapid, even if some diagnostic tests cannot perfectly discriminate subjects they are commonly used in clinics. One method for assessing quantitative diagnostic tests to diagnose subjects is to specify an optimal cut-off point. Yet, this may cause issues on quantitative diagnostic tests with a single cut-off value as the distributions of diseased and healthy subjects overlap. Forcing the subjects in this overlapped area one of the classes causes the false negative or false positive rates. To deal with this issue, there are some approaches called a gray zone or middle inconclusive area in which subjects are classified diseased, non-diseased, and neither diseased nor non-diseased. In this thesis, we aim to propose a new solution to find the boundaries of the gray zone based on the information theory approach. We intend to compare and evaluate the performance of this proposed solution against existing methods (“grey zone” and “uncertain interval” approaches). The proposed algorithm was based on joint entropy. In the simulation scenarios, we considered effect size, sample size, the homogeneity of variance and prevalence of the disease. To compare the results of the proposed methods with existing algorithms, the length of the gray zone was examined under the condition of fixed area under the receiver operating curve in out of the gray zone. In simulations, the suggested approach mostly produced the lowest gray zone length with equal variances. In some simulation scenarios, it outperformed for unequal variances. However, it has the benefit that the suggested algorithm has no previous knowledge.
