Detecting Manic State Of Bipolar Disorder Based On SVM-GMM Using Spontaneous Speech

Objectives: Bipolar disorder(BD)is characterized by a mood fluctuation,ranging from mania to depression.Current BD diagnosis and evaluation mainly depend on interview and self-report,thus could lead to subjective biases and time-consuming.Researches have confirmed that patients' speech characteristics correlate with mental disorders and speech signal provides a feasible biomarker to assist diagnosis,especially the speech detection research of Major depressive disorder(MDD).However,the literatures on BD speech detection are relatively rare.The present primary work investigated manic state detection accuracy by selecting speech features and utilizing Support Vector Machine(SVM)and Gaussian Mixture Model(GMM)with spontaneous speech.This work will lay a basic scientific foundation for explore of BD Speech Recognition System,and help doctors and patients for better diagnosis and mood state monitoring in different situations.Materials and methods: 21 hospitalized BD patients(14females,average age 34.5 ± 15.3)were recruited after admission.Spontaneous speech was collected through preloaded smartphone.Each BD patient's speech data was recorded in manic and euthymic sate,and each mood state speech were collected twice times in two days.First,speech features(pitch,formants,mel-frequency cepstrum coeffic ients(MFCC),linear prediction cepstral coefficient(LPCC),gammatone frequency cepstral coefficients(GFCC)etc.)were extracted after pre-processing from speech signals.Results: Speech features were selected using the method of between-c lass variance and within-class variance.Finally,the pitch?formants?LPCC?MFCC?GFCC were selected and made of features set.The results shows that the manic state differentiating ability of LPCC and GFCC was higher than pitch?the first three formants.The above single speech feature's manic state detecting accuracy rate of one randomly selected BD patient is lower than that of mixed speech features,whether SVM or GMM was utilized;Manic state was then detected by SVM and GMM.SVM accuracy of state detection for single patient was up to 88.56 ± 5.26%.GMM detection accuracy(72.27 ± 6.90%)for multiple patients was higher than SVM(60.87 ± 18.90%).In order to facilitate medical staff or patients' families of the patient's emotional state real-time monitoring,we cooperate with School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University to explore a bipolar disorder prototype speech recognition system based on Android mobile phone and realize the interface of the operating system.This speech recognition system and operating system interface is concise,easy to use,to achieve the goal of the man-machine dialogue.Conclusion: The spontaneous speech of BD patients communicating freely with professionals which was collected through smart phone,is applicable to the detection of manic state.Single speech feature's manic state detecting accuracy rate of is lower than that of mixed speech features.SVM provided an appropriate tool for detecting manic state for single patient,while GMM worked better for multiple patients' manic state detection.This work confirmed that speech could be an objective biomarker for assistant diagnosis and mood state monitoring in BD.This system can realize real-time monitoring of patient's condition,and has the prospect of research transformation and commercial promotion.