Automated Diagnosis of Diseases Based on Classification: Dynamic Determination of the Number of Trees in Random Forests Algorithm

The accurate diagnosis of diseases with high prevalence rate, such as Alzheimer, Parkinson, diabetes, breast cancer, and heart diseases, is one of the most important biomedical problems whose administration is imperative. In this paper, we present a new method for the automated diagnosis of diseases based on the improvement of random forests classification algorithm. More specifically, the dynamic determination of the optimum number of base classifiers composing the random forests is addressed. The proposed method is different from most of the methods reported in the literature, which follow an overproduce-and-choose strategy, where the members of the ensemble are selected from a pool of classifiers, which is known a priori. In our case, the number of classifiers is determined during the growing procedure of the forest. Additionally, the proposed method produces an ensemble not only accurate, but also diverse, ensuring the two important properties that should characterize an ensemble classifier. The method is based on an online fitting procedure and it is evaluated using eight biomedical datasets and five versions of the random forests algorithm (40 cases). The method decided correctly the number of trees in 90% of the test cases.

I. Introduction

An important task of any diagnostic system is the process of attempting to determine and/or identify a possible disease or disorder and the decision reached by this process. For this purpose, machine learning algorithms are widely employed [1], [2]. For these machine learning techniques to be useful in medical diagnostic problems, they must be characterized by high performance, the ability to deal with missing data and with noisy data, the transparency of diagnostic knowledge, and the ability to explain decisions. In this paper, the improvement of the random forests classification algorithm, which meets the aforementioned characteristics, is addressed. This is achieved by determining automatically the only tuning parameter of the algorithm, which is the number of base classifiers that compose the ensemble and affects its performance.