The study of methods for correcting class imbalances in medical and psychological data for the development of a random forest algorithm

Abstract

Despite being a powerful mechanism for building relevant models, the algorithm can produce incorrect results and therefore needs improvement. There are many such methods available, but not all of them are necessary in our case, namely in monitoring the development of psychological disorders among people with hypo- and hyperthyroidism. In this experiment, the next step is to choose an approach to eliminate class imbalance among patient medical data, such as undersampling, oversampling, SMOTE, RUSBoost, balanced random forest (BRF), and ADASYN. Based on the data, a linear additive convolution was constructed for decision making according to the criteria of time, accuracy, precision, recall, and f1. Its indicators show that in our case, RUSBoost should be chosen as a way to combat minority classes so that the algorithm produces more accurate results.