Neuroimaging Subjective Labeling Dichotomization and Class Imbalance Alleviation

Abstract

Ground truth labels are expected to be certain, and their existence is essentially a vital component of supervised learning problems. In certain cases, however, they can prove to be obstacles. They can lead to two possible issues: class imbalances due to skewed label distributions, and unreliability due to the uncertainty of raters underlying rationale. In cases where the labels are continuous, they need to be dichotomized for a classification task. Dichotomization is often decided statistically or based on the subject matter. However, the subjectivity of participants and its impact is neglected. In this paper, we investigate the effect of thresholding on an EEG emotional self-assessment. We propose a modification in the prediction pipeline to minimize subjectivity, improving model outcomes as well as alleviating the effect of label imbalance.