Çok Kategorili Hastalık Durumlarında Tanısal Modele Yeni Bir Belirteç Eklenmesinin Tanı Performansındaki Değişime Etkisinin İncelenmesi

Abstract

Accurate assessment of the risks in the field of health, which are derived from diagnostic prediction models is important in identifying patients and healthy people. In some cases, classifications may consist of more than two categories. By using the ROC analysis method, which is the most commonly method in two-category classifications, some methods have been developed for multicategory classification problems. These are HUM (Hypervolume Under the ROC Manifold) and CCP (Correct Classification Probability). However, these methods are considered to be unsuccessful in measuring improvement in the performance of the model that occurs when a new biomarker is added to an existing model. Therefore, when a new biomarker is added to the model, two performance measures have been developed to show the performance of the new classification result. These are NRI (Net Reclassification Improvement) and IDI (Integrated Discrimination Improvement) methods. In this study, we aim to investigate the relationship between HUM, CCP measures, which are the performance measures used in multi-category classification models, and NRI, IDI which measure the effect of markers affecting model performance. Considering the type of dependent variable (ordinal and nominal), the effect of the order of adding the markers to the model performance was investigated. Investigations were made with real data sets and a simulation study. In the simulation study, we simulated data based on the correlation structure with low, medium and strong magnitude in both positive and negative directions for relationships between ordinal dependent variable and the markers and within the markers. According to the simulation results of negative correlation structures, it was seen that there was a positive relationship between HUM, CCP and NRI, IDI. As the NRI and IDI methods are affected by the order of the marker added to the model as opposed to the HUM and CCP methods, it is concluded that the correct classification performance can be improved with fewer markers added in the correct order.