Feature extraction from speech signals using empirical mode decomposition for depression detection: A comparative study with machine learning models

Depression is a prevalent mental disorder that affects quality of life, and its early detection through voice analysis could improve diagnosis. This study investigates the effectiveness of Intrinsic Mode Functions in differentiating depression through voice signals. We used data from the Distress Analysis Interview Corpus. Empirical Mode Decomposition was applied to extract IMFs, and statistical characteristics and similarities were analyzed using a Gaussian kernel between depression and healthy groups according to sex. The results revealed significant differences in the mean of the first IMFs in men, but not in women, while there were no differences in other statistics. Gaussian kernel analysis showed variations in the probability density function in the first Intrinsic Mode Functions (up to IMF 7), with differences according to sex. Six Machine Learning models were experimentally tested, trained, and adjusted. The best accuracy results in women were obtained with Gradient Boosting (94.1%), while in men they were obtained with Gradient Boosting (88.0%). Intrinsic Mode Functions proved to be useful for detecting depression, suggesting their potential in developing non-invasive tools for early detection of this. • Depression patterns differ by sex in IMF distribution and density similarity. • Voice waves show more distinct statistical patterns in lower-order IMFs. • Balanced data significantly reduces false negatives in model predictions based in IMFs. • IMF selection improves model accuracy, especially for SVM in females and GB in males. • The parameterization of the Gaussian kernel and the choice of representative statistics are key in the classification of distributions, as the data vary according to sex.