Detection of Mood Disorder using Modulation Spectrum of Facial Action Unit Profiles

Abstract

Facial expression is a direct and natural way for affective expression In the field of affective computing recognition of facial expression is an important and popular research topic In mood disorder diagnosis a high percentage of bipolar disorder (BD) patients are initially misdiagnosed as having unipolar depression (UD) This misdiagnosis carries significant negative consequences for the treatment of the BD patients Therefore it is crucial to establish an accurate distinction between BD and UD in order to make an accurate and early diagnosis leading to improvements in treatment and course of illness The research goal of this thesis is to collect the facial expression of the patients with mood disorder when they were watching video clips for emotion elicitation The features extracted from the elicited facial expressions are used for the detection of BD and UD This thesis focuses on detecting the difference in facial expressions among the BD patients UD patients and healthy people responding to emotional stimuli In this study first the subjects are elicited by emotional video clips and the videos with facial expression of the subjects are recorded The corresponding facial action unit (AU) profiles are obtained using the support vector machines (SVMs) as the facial features The Modulation Spectrum (MS) characterizing the fluctuation of AU profile sequence over a video segment are further extracted Finally a two-layer SVM is constructed for mood disorder detection based on the extracted MSs of the elicited facial expressions in the video segments of the subject In order to evaluate the proposed method the video segments from 24 subjects 8 for BD 8 for UD and the remaining 8 subjects for the control group were collected at CHI-MEI Hospital Tainan Taiwan K-fold (K=8) cross validation method was performed for evaluation and the detection accuracy achieved 68 3% for mood disorder detection Compared with the well-known classifiers such as Gaussian Model and deep neural network (DNN) the experimental results confirmed that our proposed method can achieve the best performance Furthermore the AU profiles and MS features are also beneficial to improve mood disorder detection performance

Date of Award	2015 Aug 19
Original language	English
Supervisor	Chung-Hsien Wu (Supervisor)