Research on Voice Intelligent Monitoring Technology for Early Warning of Recurrence of Depression Disorder

Overview

This study aims to collect the voice output of depression patients and healthy subjects, extract the acoustic and semantic parameters, compare the similarities and differences between the depression group and the healthy control group horizontally, and track the depression patients' changes in the rehabilitation stage to construct a voice-based early warning model of depression recurrence. At the same time, the use of EEG technology, nuclear magnetic resonance and near-infrared brain imaging technology to record and analyze the neural activity characteristics behind the voice variation of depression patients, and build a neural mechanism model. And construct the facial recognition function through the convolutional neural network, extract the facial parameters to enrich the intelligent monitoring and early warning technology. 1. Collect linguistic data of depression patients and healthy people collected in the laboratory, as well as data related to changes in the condition of depression patients in daily life and home care after treatment, and construct comparative data and dynamic observations Large database to analyze its voice mutation characteristics; 2. Using EEG technology, nuclear magnetic resonance, and near-infrared brain imaging to record and analyze the neural activity characteristics behind the voice variation of depression patients, and build a neural mechanism model. 3. Use the convolutional neural network to realize the facial recognition function, and extract the facial parameters to enrich the monitoring indicators. 4. Based on the dynamic observation big data of depression speech mutation, construct the speech feature parameter vector of depression recurrence, and use adaptive personalized intelligent learning algorithm to develop intelligent monitoring and early warning technology. 5. Establish monitoring and diagnostic indicators for recurrence early warning, verify the application of the above-mentioned intelligent monitoring and early warning technology in rehabilitation guidance, and make a comprehensive assessment.

Case control study Clinical symptoms, speech behavior test, EEG data, near-infrared brain imaging data and facial expression data are collected from three groups of subjects (consolidation/maintenance depression patient group, acute episode depression patient group and healthy alternatives), In terms of magnetic resonance data and other aspects, compare the similarities and differences of the data of the three groups of subjects. Follow up research The acute-onset patient group will receive follow-up every two weeks for three months, and then continue in June, September, December, 18 and 24 months. Consolidation/maintenance patients and healthy patients will receive follow-up evaluations in March, June, September, December, 18 and 24 months after enrollment. The content includes clinical symptom assessment, behavioral testing, EEG testing, near infrared testing, etc. The time points of cranial imaging detection are 0, March, December and 24 months. In the laboratory, if the subject agrees to collect language information in the patient's real life environment. Evaluation and inspection items 1. Clinical scale evaluation 2. Speech Behavior Test Using speech production paradigm, semantic feature task, emotional prosody perception paradigm and emotional vocabulary paradigm to comprehensively detect the language perception, comprehension and output of patients with depression to find objective indicators of speech mutation. Voice detection includes short-term energy (Max, Min, Mean), pitch (Max, Min, Mean), voice speed, number of voice interruptions, short-term zero-crossing rate (Max, Min, Mean), first formant, first Two formants, 12-order MFCC (12 coefficients) and other parameters. The experiment is carried out in a well-lit and quiet room. Before the experiment, the subject will explain the experiment process to the subjects. After the experiment, in order to avoid interference, the subject should leave the room. The whole process of the experiment was in the same computer program, and the subjects responded according to the task requirements on the screen. 2.1 Speech and video production paradigm The speech and video output paradigm used in this study includes five tasks: interview, short text reading, recitation, picture description and video viewing. The video viewing task is only used for facial video data collection. The experiment is written and run using E-prime 2.0 software. Participants completed five speech production tasks accordingly, and recorded their voices and facial expressions throughout the experiment to obtain speech and video data. 2.2 Semantic processing tasks The semantic processing task is used to compare the abilities of depressed patients and healthy subjects in speech processing, semantic coding and extraction, and speech fluency, and explore more comprehensively the early warning indicators of the recurrence of depression. 3. Magnetic resonance examination The Siemens 3.0T magnetic resonance scanner is used, which is a 32-channel head coil. Before scanning, use a foam cushion to fix the head to avoid head movement during scanning. The scanning sequence is: first routine scan to obtain the T1-weighted positioning image of the whole brain 3D structure, then use the fast gradient echo imaging sequence to perform sagittal scan of the whole brain for three-dimensional reconstruction and spatial registration, and finally use plane echo ( EPI) sequence for blood oxygen dependent BOLD-fMRI imaging (resting state and task state). 4. Functional near infrared spectroscopy Using Shimadzu lighnirs portable near-infrared brain imager, the sampling rate was set to 0.1s, and the oxygenated hemoglobin concentration (HbO), deoxyhemoglobin concentration (HbR) and total oxygen content (HbT) were used as measurement indicators. Since language processing is mainly in the frontal and temporal lobes, the probe layout covers the lateral fronto-temporal complex area. After the fNIRS test, a 3D locator is used to locate the brain area. The positions of the transmitter and the detector will be referenced to the Brodmann partition, and the fNIRS channel position will be registered with the probability of the MNI space. The subjects in the fNIRS examination need to complete the speech output paradigm and semantic feature tasks. The experiment adopts a block design, the stimulation sequence is written by E-prime software, and the stimulation program programmed by the computer is presented on the computer display screen (approximately 55cm from the subject). Before the start of the experiment, the subjects should wear an optode cap to ensure that the signals of each channel are received normally. At the beginning of the experiment, the screen showed "+", and the subjects only need to watch the screen and try to stay calm as much as possible; then complete the relevant behavioral tests, and the subjects can complete according to the instructions. The entire experiment lasted about 25 minutes. The subjects collected near-infrared brain signals and voice data online while completing the task. 5. EEG examination Use Brainproducts liveamp portable EEG equipment to conduct EEG examination in two stages: resting state and voice task state. In the resting state EEG collection stage, after the subjects put on the EEG collection equipment, sit on a comfortable chair, close their eyes and stay awake, and collect the resting EEG for 5 minutes. In the voice task stage, the subject wears an EEG device and sits in front of the computer screen. According to the instructions displayed on the screen, the voice production paradigm and semantic feature tasks are performed accordingly. The microphone will simultaneously record the subject's voice. The experimental program is written using e-prime2.0. 6. Other physiological tests Physiological indicators (including heart rate, blood pressure, weight, height, etc.), environmental indicators (including weather, temperature, environmental noise, ultraviolet intensity, human comfort, etc.), Cambridge Neuropsychological Test Automatic Battery (CANTAB) The gait test during natural walking, the eye movement during task and facial expression information will be performed once in baseline, March, June, September, December, 18 and 24 respectively. 7. Feasibility and specificity analysis of multiple detection methods This study uses multiple detection methods, including clinical scale evaluation, language testing, and neurophysiological examination. The purpose is to comprehensively and systematically collect subjects' linguistic related information to better develop and develop voice intelligent detection technology. application. Multiple detection methods are feasible. Clinical scales are implemented by trained researchers, language-related paradigms are implemented by computer programs, and neurophysiological testing techniques are operated by experimenters. All detection methods and methods are widely used in clinical research and are feasible. Statistical analysis plan, definition and selection of statistical analysis data set 1. Statistical analysis data set Full analysis set (FAS, Full analysis set): People who have undergone at least one clinical scale evaluation, voice collection, neurophysiological data, and major index evaluation data after enrollment. Cases with missing baseline data for the main evaluation indicators will be excluded. Per Protocol Set (PPS, Per Protocol Set): On the basis of FAS, cases that meet the inclusion and exclusion criteria, have a valid baseline value, have good compliance, and do not violate the clinical research protocol. The curative effect analysis uses both the full analysis set and the conforming plan set, and the FAS result is the main one. 2. Data description Observe whether the data conforms to the normal distribution, if not, modify the statistical method or perform data conversion. Observe whether there are outliers and missing values, and perform statistical and professional analysis to determine the choice and the way of filling the missing values. Descriptive statistics indicate the mean, standard deviation, maximum, minimum, credible interval, frequency, composition ratio, etc. 3. Statistical inference methods Measurement data adopt t test, paired t test, rank sum test, paired rank sum test, median test and other methods. Counting data: Chi-square test, Fisher's exact test, etc.; grade data using Ridit analysis, CMH method. Use machine learning algorithms such as Support Vector Machine (SVM), Dynamic Bayesian Network (DBN) and Deep Neural Networks (DNN) for speech and video data. 4. Statistical expression The report is mainly expressed in tables, with table titles, table notes, and number of cases. In addition to the statistical results of the comparison between the groups, the statistical parameter diagrams and other icons are attached to increase the readability. 5. Statistical software Use SPSS22.0, Matlab software (v2014a) and NIRS-SPM (Statistical Parametric Mapping for Near-infrared Spectroscopy) software for parameter statistics (t-test, analysis of variance, etc.) and non-parametric statistics (rank sum test, chi-square test) Etc.), using two-sided tests, P less than or equal to 0.05 is considered statistically significant.