A Detection Method for Depression Based on Individual Difference Evaluation Using Domain Shift Metrics

Abstract

Machine learning using physiological indices has attracted attention for emotion estimation and mental disorder diagnosis. On the other hand, physiological indices have individual differences due to factors such as age and gender. Therefore, it is difficult to improve the accuracy of machine learning using physiological indices. In this study, we investigate a method to evaluate individual differences using domain shift, which is used in domain adaptation, with the aim of improving machine learning accuracy. Domain shift is a measure of the degree of discrepancy of simultaneous probability distributions among datasets, and we believe that it can be used to evaluate individual differences by applying it to data for each individual. We applied this method to a dataset for depression detection, and found a significant difference between depressed and healthy subjects in clustering using individual differences.