Aberrant Temporal-spatial Patterns to Sad expressions in Major Depressive Disorders via Hidden Markov Model

Background The pathological mechanisms of Major depressive disorders (MDD) is associated with over-expressing of negative emotions, and the overall temporal-spatial patterns underlying over-representation in depression still remained to be revealed to date. We hypothesized that the aberrant spatio-temporal attributes of the process of sad expressions relate to MDD and help to detect depression severity. Methods We enrolled a total of 96 subjects including 47 MDDs and 49 healthy controls (HCs), and recorded their Magnetoencephalography data under a sad expressions recognition task. A hidden Markov model (HMM) was applied to separate the whole neural activity into several brain states, then to characterize the dynamics. To find the disrupted spatial-temporal features, power estimations and fractional occupancy of each state were contrasted between MDDs and HCs. Results Three states were found over the period of emotional stimuli processing procedure. The visual state was mainly distributed in early stage (0 - 270ms) and the limbic state in middle and later stage (270ms - 600ms) of the task, while the fronto-parietal state remained a steady proportion across the whole period. MDDs activated statistically more in limbic system during limbic state (p = 0.0045) and less in frontoparietal control network during fronto-parietal state (p = 5.38*10−5) relative to HCs. Hamilton-Depression-Rating Scale scores was significantly correlated with the predicted severity value using the state descriptors (p = 0.0062, r = 0.3933). Discussion As human brain exhibited varied activation patterns under the negative stimuli, MDDs expressed disrupted temporal-spatial activated patterns across varied stages involving the primary visual perception and emotional contents processing compared to HCs, indicting disordered regulation of brain functions. Furthermore, descriptors built by HMM could be potential biomarkers for identifying the severity of depression disorders.

pre-specifying the window length (Quinn et al., 2018;Vidaurre et al., 2016). Besides, it is convenient for HMMs to generate state-wise mean activation maps in large-scale network via the multivariate observation model, as well as to observe the processing of visual perception, decision making and motor response by the sequential spatiotemporal activation maps. Recent study also showed that impaired brain dynamics could be characterized not only in limited targeted regions but also in the large-scale brain networks via HMMs (Charquero-Ballester et al., 2020). Furthermore, dynamic descriptors such as fractional occupancy inferred from HMMs were found to correlate with symptoms of schizophrenia patients, emerging the potential of the promotion to other psychosis, like MDD (Zhi, Calhoun, Lv, Ma, & Ke, 2018).
In the present study, we aimed at exploring the abnormal spatiotemporal brain patterns in MDD patients under the negative emotional task. To achieve this, an AE-HMM model was applied on Magnetoencephalogram (MEG) data recorded under the stimulus of sad facial recognition task. MEG with high temporal resolution in milliseconds could be utilized to explore neuropsychological mechanisms of fast neural activity for MDD. The processing of chronological neural activity across visual task was characterized by several brain states whose dynamic descriptors are identifiable for MDD. Furthermore, regression analysis was adopted to explore the ability of these dynamic descriptors to indicate the severity of disorders in MDD. Our study might be a promotion to apply HMM to the field of MDD. We provided a new perspective to the evolution process of negative emotional stimulus over the visual task especially in MDD patients.

Participants
One hundred individuals (50 HC and 50 MDD) were enrolled in the Nanjing Brain Hospital between October 2011 and June 2016. All individuals were given Mini international Neuropsychiatric Interview (MINI) to exclude potential MDD from HC.
After exclusion in clinical judgment (potential bipolar disorders) and imaging quality (excessive head motion and other artifacts), ninety-six individuals (49 HC and 47 MDD) were recruited in this study. They were all matched in gender, age and education (Table1). For MDD, the severity of disease was assessed by professional psychiatrists using Hamilton Depression Rating Scale (HAM-D) and the Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-4). The inclusion criteria were no brain injury, alcohol or drug abuse. All individuals were right-handed and provided with written consent forms. This study was approved by the ethical committee at Nanjing Brain Hospital. Data were presented as the range of mean ± standard deviation (two tailed t-tests)

Stimuli and task
All individuals were engaged in a passively viewing task of emotional faces. Forty sad facial expressions were chosen from the Chinese Facial Expression Video System (Jiang et al., 2019;Jing-Lun, Yao, & Xie, 2007). Each video was presented lasting for 3 seconds. Then a fixation cross picture was also displayed during the rest. An random inter-trial interval of 0.5s, 1s, or 1.5s followed each facial expression.

Data acquisition and preprocessing
Task MEG data were recorded with a whole-head CTF275 MEG system (VSM Inc) with a 300 Hz sampling rate. Individuals were scanned lying in the supine position in a magnetically shield room. When scanned, individuals were required to adjust head positions if their head motions exceeded 5 mm compared to the initial position.
Individuals' structural T1 images were recorded using a 1.5T GE system and 3D gradient-echo pulse sequence.
MEG data were pre-processed using Matlab-based fieldtrip toolbox (Oostenveld,

Source reconstruction
The pre-processed data in sensor space were projected in source space onto a 6mm grid by a Linearly Constrained Minimum Variance (LCMV) beamformer (Woolrich, Hunt, Groves, & Barnes, 2011). The LCMV beamformer normalized grid in MNI space and constructed a realistic head model using participant's structural MRI. A covariance matrix was calculated across all trials using the spatial filters.
Subsequently, the spatial orientations of each epoch were rotated in order to maximize the variance. Source activity was estimated by multiplying spatial filters with sensor-level time series across the whole scale. Subsequently, a multivariate symmetric orthogonalization was applied to attenuate the spatial leakage interferences (Colclough, Brookes, Smith, & Woolrich, 2015).

Construction of AE-HMM over all subjects
To further explore dynamics in whole brain, MEG data in source space were processed using matlab-embedded HMMAR toolbox (Vidaurre et al., 2016). In this study, an AE-HMM was constructed to exhibit the process of dynamics transition during the emotional face task. Before estimating the model, whole-brain voxel set in source space were projected onto ninety brain regions based on AAL90 template.
Subsequently, source time-series were filtered between 1-120 Hz and the amplitude envelope was calculated using Hilbert transition. Data were concatenated across all individuals to constitute a big 3-dimension matrix whose first dimension is the number of individuals, and the second dimension is the number of brain regions, the third dimension is the number of time series. A schematic of the whole process was exhibited in Figure1. Figure 1 The architecture of the whole pipeline (A) A schematic for the preprocessing of MEG signal applied prior to HMM analysis. (B) A schematic for HMM model, in which X t denotes the brain resides at time point t, while the Y t denotes the observed data. (C) An illustration of brain activation patterns from HMM used to predict disorders severity with Support-Vector-Regression model.
Considering the variance distribution of our data and making sure the lowest free energy, this study denoted K=3 before training the model which also followed findings in Hirschmann's study (Hirschmann et al., 2020). The observation model subsequently was constructed by training 1000 iterations with pre-defined K states.

Statistical analysis over state-wise spatial-temporal characters between MDDs and HCs
To characterize the spatial power distribution of each state, covariance matrix was obtained from observation model. The mean activation map was calculated by averaging the prior distributions of the envelope value for brain regions in each state.
According to the study in 2019 (Luppi et al., 2019), regions in AAL90 template were divided into seven networks based on dynamic interactions and diversified functions of brain. Each inferred state was shown together with a mean activation for distinct brain networks. Using the time courses of posterior probabilities inferred from HMM, the power activation map of each subject was estimated. Next, non-parametric two-sample t tests were utilized over state-wise power activation in network level between MDD and HC groups. The multiple comparisons were controlled via a Bonferroni correction.
In addition, Fractional occupancy (FO) was computed to characterize the dynamics of inferred brain states in each participant. This descriptor is defined as the proportion of each state time spent in the whole time length (Zhi et al., 2018). To further analyze the dysfunction of brain dynamics, FO values in each state were then compared between MDDs and HCs by two-sample independent t test.

Correlational analysis between dynamic characters and MDD severity
To assess the relation between patients' disease severity and HMM descriptors, a support vector machine regression (SVR) algorithm was applied to regress disease severity in MDD using FO values and HAM-D scores. A Gaussian kernel regression was applied with a leave-one-out-cross-validation (LOOCV). Mean Absolute Error (MAE) was then calculated by average absolute bias between predictive value and real value to assess the predictive performances. Finally, correlational analysis was performed between predicted scores and real HAM-D scores via Pearson correlation.

State-wise temporal-spatial patterns for all subjects
In current study, three brain states were identified by AE-HMM over all subjects. As observed in Figure2

Disrupted spatial-temporal patterns associated with MDD
After comparing network power between MDD and HC groups, network with significant differences were found in two states: limbic system in state2 and FPN in state3. Brain regions in these two networks were illustrated in Figure3

Dynamic characters related to MDD severity
Since dysfunction of FO was proven to be associated with the depressed disorders, the descriptor was selected as the indicator to the severity of disorders in MDD group. As shown in Figure4

Discussion
This study characterized the aberrant brain activation patterns of mood disorders and found specific state regulating different brain systems through diverse temporal stages of the sad facial recognition task. Besides, dynamic descriptors of FO inferred from each state were regarded as valuable index for individualized severity of depression.
First, the current study revealed the common neural patterns under the negative stimuli for all subjects. We inferred the whole procedure of negative emotions processing into three separate brain states. Specifically, during the early stage of visual stimuli processing (around 0-270ms), subjects showed dominant activation in VN and DAN. During the middle and latter processing stage after the onset of the stimuli (around 270-600ms), subjects manifested more activation in limbic system and DMN. Besides, subjects exhibited a relative small but essential proportion of brain activation in FPN and VN in the whole task (0-600ms In summary, we reconstructed the brain states in the overall process of a passively sad expression recognition task via AE-HMM and found that aberrant temporal-spatial patterns in different process stages like primary visual processing and emotional contents processing were correspond to the different dysfunction of brain functions for MDD respectively. Furthermore, dynamic descriptors of FO values inferred from HMMs could reflect the aberrant dynamism and predict the severity of MDD  n  e  t  w  o  r  k  a  c  t  i  v  i  t  y  <  3  5  H  z  a  c  c  o  u  n  t  s  f  o  r  v  a  r  i  a  b  i  l  i  t  y  i  n  s  t  i  m  u  l  u  s  -i  n  d  u  c  e  d  g  a  m  m  a  r  e  s  p  o  n  s  e  s  .   N  e  u  r  o  i  m  a  g  e  ,  2  0  7   ,  1  1  6  3  7  4  .  d  o  i  :  1  0  .  1  0  1  6  /  j  .  n  e  u  r  o  i  m  a  g  e  .  2  0  1  9  .  1  1  6  3  7  4  J  i  a  n  g  ,  H  .  ,  H  u  a  ,  L  .  ,  D  a  i  ,  Z  .  ,  T  i  a  n  ,  S  .  ,  Y  a  o  ,  Z  .  ,  L  u  ,  Q  .  ,  &  P  o  p  o  v  ,  T  .  (  2  0  1  9  )  .  S  p  e  c  t  r  a  l  f  i  n  g  e  r  p  r  i  n  t  s  o  f  f  a  c  i  a  l  a  f  f  e  c  t  p  r  o  c  e  s  s  i  n  g  b  i  a  s  i  n  m  a  j  o  r  d  e  p  r  e  s  s  i  o  n  d  i  s  o  r  d  e  r  .   S  o  c  C  o  g  n  A  f  f  e  c  t  N  e  u  r  o  s  c  i  ,  1  4   (  1  1  )  ,  1  2  3  3  -1  2 l  e  c  t  r  o  c  o  r  t  i  c  a  l  r  e  a  c  t  i  v  i  t  y  t  o  n  e  g  a  t  i  v  e  a  n  d  p  o  s  i  t  i  v  e  f  a  c  i  a  l  e  x  p  r  e  s  s  i  o  n  s  i  n  i  n  d  i  v  i  d  u  a  l  s  w  i  t  h  a  f  a  m  i  l  y  h  i  s  t  o  r  y  o  f  m  a  j  o  r  d  e  p  r  e  s  s  i  o