Automatic Depression Discrimination Based On Functional Near-infrared Spectroscopy

Since the 21 st century, with the rapid development of economy, the life pressure is becomingheavier and heavier, and some mental illness like depression also become more frequently, what’s more, diagnostic accuracy of depression is very low. After the appearance of functional near- infrared spectroscopy, diagnosis of depression ushered in a new opportunity. After the FNIRS data is get, the doctor can judge whether the subject is depression or not according to their experience, the accuracy of diagnosis is improved.While if we can do some analysis and process about the FNIRS data and train a classifier that can automatically discriminate the depression, it will be significant for the diagnosis of depression.In this paper, the feature of FNIRS data is extracted from time and frequency respectively, and then a classifier is trained and tested by using SVM.The research work of this thesis mainly includes following aspects:(1) In order to reduce the effects of baseline fluctuation and high frequency noise, Band-pass filtering is used to do data preprocessing. Before band-pass filtering, Fourier transform is used to do the spectrum analysis to find which band the FNIRS exists in.(2)Feature extraction of normal person and depression FNIRS data. In this paper, feature is extracted from time and frequency domain respectively. In the frequency domain, wavelet package decomposition is used to extract the feature and the wavelet package energy feature is extracted, this feature can reflect the essence of signal. In the time domain, the general linear model is used to extract the parameter β.(3)C lassifier is trained based on SVM. In this paper, SVM is used to train the classifier, firstly, the feature is formatted, then after some steps like normalization, kenerl function selection and best parameters selection using coss-validation, the classifier is obtained.At last, the classifier is tested based on a testing set. The testing set is inputted to the classifier and the output will be the accuracy. In this paper, the accuracy of depression is up to 85%, a desired purpose is achieved. This provides an objective data support for the diagnosis of depression, the accuracy of depression diagnosis will improve a lo t, and the automatic discrimination of depression is implemented.