UCLA

Magnetic Resonance Spectroscopy (MRS) is an in-vivo, non-invasive technique to measure biochemical metabolite concentrations in the brain. Here, we propose novel tools for processing and analysis of MRS data. We apply these tools to classify disease status in major depressive disorder (MDD). Our tools enable the manipulation and formatting of MRS data, and allow for the application of several machine-learning classifiers used to predict disease status. We test and derive several representations of MRS data including the conventional metabolite concentration levels, full MRS spectral function data points, elastic functional distances from the mean, and principal component analysis (PCA) components of the MRS spectra. We apply a number of different machine learning algorithms including logistic regression, random forest classification, and support vector machines (SVMs). The highest performance of all these classifiers resulted from the use of random forest classification on full spectral data points with a precision of 0.698 and a recall of 0.695. This predictive approach to disease status classification can be used to allow possible depression cases to be caught early on as well as for confirmation of diagnosis.