Implementing decision tree-based algorithms in medical diagnostic decision support systems

Abstract

As a branch of healthcare, medical diagnosis can be defined as finding the disease based on the signs and symptoms of the patient. To this end, the required information is gathered from different sources like physical examination, medical history and general information of the patient. Development of smart classification models for medical diagnosis is of great interest amongst the researchers. This is mainly owing to the fact that the machine learning and data mining algorithms are capable of detecting the hidden trends between features of a database. Hence, classifying the medical datasets using smart techniques paves the way to design more efficient medical diagnostic decision support systems. Several databases have been provided in the literature to investigate different aspects of diseases. As an alternative to the available diagnosis tools/methods, this research involves machine learning algorithms called Classification and Regression Tree (CART), Random Forest (RF) and Extremely Randomized Trees or Extra Trees (ET) for the development of classification models that can be implemented in computer-aided diagnosis systems. As a decision tree (DT), CART is fast to create, and it applies to both the quantitative and qualitative data. For classification problems, RF and ET employ a number of weak learners like CART to develop models for classification tasks. We employed Wisconsin Breast Cancer Database (WBCD), Z-Alizadeh Sani dataset for coronary artery disease (CAD) and the databanks gathered in Ghaem Hospital’s dermatology clinic for the response of patients having common and/or plantar warts to the cryotherapy and/or immunotherapy methods. To classify the breast cancer type based on the WBCD, the RF and ET methods were employed. It was found that the developed RF and ET models forecast the WBCD type with 100% accuracy in all cases. To choose the proper treatment approach for warts as well as the CAD diagnosis, the CART methodology was employed. The findings of the error analysis revealed that the proposed CART models for the applications of interest attain the highest precision and no literature model can rival it. The outcome of this study supports the idea that methods like CART, RF and ET not only improve the diagnosis precision, but also reduce the time and expense needed to reach a diagnosis. However, since these strategies are highly sensitive to the quality and quantity of the introduced data, more extensive databases with a greater number of independent parameters might be required for further practical implications of the developed models.