Probing prefrontal cortex hemodynamic alterations during facial emotion recognition for major depression disorder through functional near-infrared spectroscopy

Major depression disorder (MDD), one of the most frequent psychiatric disorders, has high estimated lifetime prevalence rates of 16.6% [1], which is associated with considerable impairments in social functioning [2, 3]. Currently, the MDD diagnosis mainly depends on patients' reports of symptoms, observed behaviors and disease course. Establishment of clinically useful biomarkers for the MDD diagnosis would enhance patients' management and treatment effect, and lead to the therapies adjusted to the individual [4]. However, no such biomarkers have been established up to now. Therefore, the development of objective and feasible biomarkers is of special significance and a great challenge for accurate and early diagnosis and treatment of depression, which can overcome the limitations of relying on clinical interviews alone.

Impairments in processing of facial emotional expressions and biased facial emotion detection are frequently found, and produce barriers in interpersonal engagement and social functioning for the MDD patients [5–8]. Emotional face processing consistently activates brain regions in limbic and prefrontal areas including the amygdala, orbitofrontal cortex, ventrolateral prefrontal cortex, and anterior cingulate cortex [9, 10]. Functional imaging studies revealed that an increased activity in the MDD patients in medial prefrontal regions in negative emotions regulation tasks [11]. A compensatory recruitment of parietal and lateral prefrontal cortex has been reported while redirecting attention away from interfering emotions for the MDD patients [12]. Findings regarding the involvement of dorsolateral prefrontal cortex (PFC) have been controversial, showing increased [13], decreased [14] or similar activity compared to the healthy controls (HC) [15]. Thus, the emotion recognition deficit appears greatly related to the abnormal activation of the prefrontal cortex.

Up to now, functional near-infrared spectroscopy (fNIRS) has not been applied for the facial emotion recognition for the depression patients. Previous studies on neural mechanism of the facial emotion recognition of the MDD patients were mainly through functional magnetic resonance imaging (fMRI). The fMRI technique for functional imaging is limited by the fact that the individuals need to be placed in an uncomfortable setting and foreign, noisy, dark, or claustrophobic environment (e.g. lying in a supine position in a narrow gantry with the head restrained during the entire examination), for accurate measurement during the procedure, with relatively lower temporal resolution. In contrast, a multi-channel fNIRS machine provides a completely non-invasive, quiet and mobile measurement of brain function in ordinary clinical settings and allows patients to be comfortably seated in a normal posture in a well-lit room, with higher temporal resolution making it possible to obtain a recording of the actual time course of a hemodynamic epoch in response to a specific cognitive task in a single trial [16–21]. More importantly, due to the small operating and maintenance costs associated with fNIRS, it is possible to run fNIRS study with a large sample of participants [22]. Thus, the fNIRS is particularly well suited for repeated studies and diagnosis practice involving patients with psychiatric disorders.

So far as we know, it is the first time using the fNIRS signals to investigate the neural mechanisms during facial emotion recognition for the MDD patients. We measured and analyzed the difference of the median, the Mayer wave power, the mean cross wavelet coefficient, and the mean wavelet coherence coefficient from hemodynamic signals in prefrontal cortex induced by cognitive activation in facial emotion recognition tasks. In the present study, we hypothesized the physiological features of the hemodynamic responses in prefrontal cortex measured by fNIRS during face emotion recognition could provide reliable and feasible diagnosis biomarkers of underlying major psychiatric disorders with depression.

2.1. Participants

2.2. Experimental procedure

The experiment began and ended with a 3 min rest, and consisted of three runs separated by 1 min intervals, during which the participants were passively presented with greyscale emotional faces (happy, fearful, and sad), respectively (figures 1(a) and (b)). The pictures were taken from the extended-Cohn-Kanade face database [23]. We used the same procedures for each run of three emotional face types (figure 1(c)). In brief, the runs began with a two faces appearing for 1 s. These were replaced by an asterisk in one hemi-field for less than 3 s. For the behavioral measure of attention, there were two conditions of interest: (1) congruent trials, in which an emotional/neutral face pair was followed by an asterisk on the same side on which the emotional face had appeared; and (2) incongruent trials, in which an emotional/neutral face pair was followed by an asterisk on the same side on which the neutral face had appeared. Participants were instructed to press one button with their thumb under the congruent condition or a second button with their index finger under the incongruent condition as quickly as possible. The inter-trial interval was 2 s–4 s. Each run consisted of 40 trials for each of three emotional face types, including 20 congruent trials and 20 incongruent trials. Before scanning, participants were trained on the task. The pictures of the emotional faces were randomly selected from the database to eliminate the adaptive behaviour caused by the practice effect. During the experiment, the accuracy and response time of key pressing were automatically recorded for the further analysis.

Zoom In Zoom Out Reset image size

2.3. fNIRS data acquisition

A fNIRS measurement equipment (CW-NIRS) developed by our laboratory in Xi'an Jiaotong University was applied to continuously monitor the hemodynamic changes in the prefrontal area of the subjects in real-time in the present study (see figures 2 and S1(stacks.iop.org/JNE/16/026026/mmedia)) [24]. The near-infrared instrument had two probes with a light source and three detectors. The light source sent out two different near infrared light in an orderly way, with wavelengths of 735 nm and 850 nm, respectively. The distances between the three detectors and the light source were 12 mm, 20 mm and 30 mm, respectively, to measure the near infrared signal of the brain tissues at different depths in the prefrontal area. During the experiment, the near-infrared probes were symmetrically distributed on the left and right forehead area of the subjects. The modified Lambert–Beer law can be adopted to convert the near infrared signal to the blood oxygen signal. The measured blood oxygen signals include surface oxygenated hemoglobin (oxy-Hb), surface deoxygenated hemoglobin (deoxy-Hb), proximal oxy-Hb, proximal deoxy-Hb, distal oxy-Hb, distal deoxy-Hb, and local tissue blood oxygen saturation in the left and right prefrontal cortex. The units of oxy-Hb and deoxy hemoglobin are µmol l⁻¹, and local tissue blood oxygen saturation is percentage. The sampling frequency was 10 Hz.

Zoom In Zoom Out Reset image size

2.4. fNIRS data analysis

2.4.1. Preprocessing.

A signal ${{y}_{ref}}(n)$ from the surface detectors mainly included global interference ${{y}_{global}}(n)$ (such as heart rate pulse, respiration, and low frequency oscillation) and noise $\varepsilon (n)$, while a far detector signal ${{y}_{standard}}(n)$ contained evoked hemodynamic response ${{y}_{deep}}(n)$ (EHR), global interference ${{y}_{global}}(n)$ and noise $\delta (n)$. We applied improved complete ensemble empirical mode decomposition with adaptive noise and recursive least square [25, 26] to decompose the signal. Then the surface detector signal was dynamically adjusted and subtracted from the far detector signal to extract the EHR related to functional activation in response to the external stimuli [27]. The mathematical procedure is described as follows [25, 26].

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{y}_{ref}}(n)={{y}_{global}}(n)+\varepsilon (n)\nonumber \end{align} \tag{ 1 }$$

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{y}_{standard}}(n)={{y}_{deep}}(n)+w(n)\times {{y}_{global}}(n)+\delta (n)\nonumber \end{align} \tag{ 2 }$$

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{y}_{deep}}(n)={{y}_{standard}}(n)-w(n)\times {{y}_{global}}(n)-\gamma (n)\nonumber \end{align} \tag{ 3 }$$

$n$ represents the time. $w(n)$ is the weight of global interference, $\gamma (n)$ is the noise.

2.4.2. Median filter.

For each measurement channel, we applied a one-dimensional median filter (window width  =  600, step  =  10) to the raw temporal data to eliminate outliers such as those with sudden jumps or drops and remove impulsive noise and motion artifacts. The median differences of the blood oxygen signals between emotional face recognition stage and rest stage were calculated for all the subjects. And the independent-sample t test was performed between the MDD and the HC groups.

2.4.3. Mayer wave power analysis.

We used Welch method to calculate the power spectral density of the blood oxygen signal, and the frequency range is 0.08–0.12 Hz, corresponding to the Mayer wave (window width  =  600, step  =  10). The blood oxygen signals were standardized to eliminate the effect of amplitude variation. The ratio of the Mayer wave power of the blood oxygen signals at emotional face recognition stage to the rest stage (short of fearful/rest, happy/rest, and sad/rest, respectively) was calculated, and the statistical test was also performed between the MDD and the HC groups.

2.4.4. The continuous wavelet transform (CWT).

The wavelet transform is achieved by correlating the signal with a wavelet $\varphi \left( t \right)\in {{L}_{2}}\left( \mathbb{R} \right)$ that satisfies the admissibility condition [28]:

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle \int_{\mathbb{R}}{\frac{{{\left| \psi (\,f) \right|}^{2}}}{\left| f \right|}df<\infty }.\nonumber \end{align} \tag{ 4 }$$

$\psi \left(\,f \right)$ is the Fourier transform of $\varphi \left( t \right)$. The admissibility condition ensures the invertibility of the wavelet transform. Additional properties are also required: (a) Since the coherence estimator is very sensitive to phase information, the wavelet should be complex. (b) $\varphi \left( t \right)\in {{L}_{2}}\left( \mathbb{R} \right)$ necessitates that ${{\lim }_{\left| t \right|\to \infty }}\varphi (t)=0$ and ${{\lim }_{\left| f \right|\to \infty }}\psi (\,f)=0$. (c) The spectrum of the wavelet should be unimodal [28].

The wavelet transform is defined as

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{W}_{\varphi ,x}}(s,\tau)=\frac{1}{\sqrt{s}}\int_{\mathbb{R}}{x(t)\varphi (\frac{t-\tau }{s})dt}\nonumber \end{align} \tag{ 5 }$$

with $s$ representing reciprocal frequency, $\tau$ representing time. $x(t)$ is the signal for analysis. $\varphi (t)$ is a wavelet with zero mean, and that is localized in both frequency and time. The classical version of the Heisenberg uncertainty principle points out that there is always a tradeoff between localization in time and frequency. One particular wavelet, the Morlet, is defined as [29]

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{\varphi }_{{{\omega }_{0}}}}(\eta)={{e}^{\,j{{\omega }_{0}}\eta }}{{e}^{-\frac{1}{2}{{\eta }^{2}}}},\quad {{\omega }_{0}}\geqslant 5\nonumber \end{align} \tag{ 6 }$$

where ${{\omega }_{0}}$ is dimensionless frequency, and $\newcommand{\e}{{\rm e}} \eta$ is dimensionless time.

The CWT of a time series $x(n)$ with uniform time steps $\delta t$, is defined as the convolution of $x(n)$ with the scaled and normalized wavelet [29].

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle W_{n}^{X}(s)=\sqrt{\frac{\delta t}{s}}\sum\limits_{{{n}^{\prime }}=1}^{N}{{{x}_{{{n}^{\prime }}}}{{\varphi }_{{{\omega }_{0}}}}[\frac{({{n}^{\prime }}-n)\delta t}{s}]}.\nonumber \end{align} \tag{ 7 }$$

The wavelet power is defined as ${{\left| W_{n}^{X}(s) \right|}^{2}}$. The cross wavelet transform (XWT) of two time series ${{x}_{n}}$ and ${{y}_{n}}$ is defined as ${{W}^{XY}}={{W}^{X}}\overline{{{W}^{Y}}}$, where $\overline{{{W}^{Y}}}$ denotes complex conjugation of ${{W}^{Y}}$. The cross wavelet power is defined as $\left| {{W}^{XY}} \right|$. The complex argument $\arg \left( {{W}^{XY}} \right)$ can be interpreted as local relative phase between ${{x}_{n}}$ and ${{y}_{n}}$ in time.

In the present study, the signals were standardized to make the standard deviation equal to 1, so as to eliminate the effect of amplitude variation. We restricted our further analysis to this Morlet wavelet (with ${{\omega }_{0}}=6$) with a balance between time and frequency localization. The cross wavelet power spectrum between oxy-Hb and deoxy-Hb over the left and right prefrontal brain areas respectively, and the cross wavelet power spectrum between the bilateral oxy-Hb, and also between the bilateral deoxy-Hb were calculated for all the subjects.

2.4.5. Wavelet coherence.

The wavelet coherence of two time series ${{x}_{n}}$ and ${{y}_{n}}$ is define as:

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle R_{n}^{2}(s)=\frac{{{\left| S({{s}^{-1}}W_{n}^{XY}(s)) \right|}^{2}}}{S({{s}^{-1}}{{\left| W_{n}^{X}(s) \right|}^{2}})\cdot S({{s}^{-1}}{{\left| W_{n}^{Y}(s) \right|}^{2}})}\nonumber \end{align} \tag{ 8 }$$

where $S$ is a smoothing operator. It can be considered as a localized correlation coefficient in time frequency space. The smoothing operator S is written as [29].

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle S\left( W \right)={{S}_{scale}}\left( {{S}_{time}}\left( {{W}_{n}}(s) \right) \right)\nonumber \end{align} \tag{ 9 }$$

where ${{S}_{scale}}$ denotes smoothing along the wavelet scale axis, and ${{S}_{time}}$ is smoothing in time axis. For the Morlet wavelet a smoothing operator is given as

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{S}_{scale}}\left( W \right)\left| _{s} \right.=\left( {{W}_{n}}(s)\ast {{c}_{1}}^{{-{{t}^{2}}}/{2{{s}^{2}}}\;} \right)\left| _{s} \right.\nonumber \end{align} \tag{ 10 }$$

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {{S}_{time}}(W){{\left| _{n}=({{W}_{n}}(s)*{{c}_{2}}\Pi (0.6s)) \right|}_{n}}\nonumber \end{align} \tag{ 11 }$$

where ${{c}_{1}}$ and ${{c}_{2}}$ are normalization constants and $\prod$ is the rectangle function.

The factor of 0.6 was the empirically determined scale decorrelation length for the Morlet wavelet [30]. The wavelet coherence between oxy-Hb and deoxy-Hb over the left and right prefrontal brain areas respectively, between the bilateral oxy-Hb, and also between the bilateral deoxy-Hb were calculated for all the subjects.

3.1. Sample characteristics

3.2. Behavioural results

3.3. Median results

The average median changes of the blood oxygen signal in the MDD and the HC groups under the emotional face recognition task are shown in figures 3(a) and (b), respectively. The three shadow periods in each plot represent the task periods from the fearful face recognition, happy face recognition to the sad face recognition, and the non-shadow periods demonstrate the rest periods. The median value of oxy-Hb for the MDD patients was larger over the rest periods compared with the task stages; the median deoxy-Hb of the MDD patients was smaller over the rest stages compared with the task periods. The trend of the median changes of the blood oxygen signal was almost opposite for the HC group compared with the MDD patients.

Zoom In Zoom Out Reset image size

The median differences of the blood oxygen signals between emotional face recognition stage and rest stage, that is, the difference between the average median of the fearful face recognition task stage and the rest stage (simply written as fearful-rest), the difference between the average median of the happy face recognition stage and the rest stage (happy-rest), and the difference between the average median of the sad face recognition stage and the rest stage (sad-rest), and the independent-sample t test results between the groups are shown in table 3. Except for the right deoxy-Hb for the fearful-rest (p   =  0.050), there were significant differences in all the median differences in fearful-rest (left oxy-Hb: p   =  0.005, right oxy-Hb: p   =  0.001, left deoxy-Hb: p   =  0.001), happy-rest (left oxy-Hb: p   =  0.017, right oxy-Hb: p   =  0.000, left deoxy-Hb: p   =  0.004, right deoxy-Hb: p   =  0.039) and sad-rest (left oxy-Hb: p   =  0.046, right oxy-Hb: p   =  0.002, left deoxy-Hb: p   =  0.007, right deoxy-Hb: p   =  0.034) between the MDD and the HC groups. The median difference of the oxy-Hb of the MDD patients was smaller than zero, while it was larger than zero for the HC group. On the contrary, the median difference of the deoxy-Hb of the MDD subjects was larger than zero, while it was smaller than zero for the HC subjects.

Table 3. Difference of the medium value of the blood oxygen signals between the emotional face recognition stage and the rest stage for the MDD and HC groups (mean  ±  standard deviation).

Variable	Prefront-al brain area	The median difference of the oxy-Hb, µM		p  value	The median difference of the deoxy-Hb, µM		p  value
		MDD	HC		MDD	HC
Fearful-rest	Left	0.98  ±  1.81	0.71  ±  1.82	0.005b	0.44  ±  0.69	0.18  ±  0.40	0.001b
	Right	1.21  ±  2.67	1.39  ±  1.50	0.001b	0.19  ±  0.93	0.29  ±  0.52	0.050
Happy-rest	Left	0.58  ±  1.55	1.06  ±  2.50	0.017a	0.44  ±  0.59	0.19  ±  0.72	0.004b
	Right	1.00  ±  2.52	1.92  ±  2.21	0.000b	0.35  ±  1.01	0.29  ±  0.90	0.039a
Sad-rest	Left	0.44  ±  2.06	1.40  ±  3.40	0.046a	0.49  ±  0.73	0.25  ±  0.89	0.007b
	Right	0.74  ±  2.73	2.62  ±  3.52	0.002b	0.36  ±  1.10	0.42  ±  1.13	0.034a

^ap   <  0.05. ^bp   <  0.01.

3.4. Mayer wave power results

The average Mayer wave power results of the MDD and the HC groups are shown in figures 4(a) and (b), respectively. The Mayer wave power of the oxy-Hb and deoxy-Hb of the MDD patients was relatively larger in the rest stage, while it was smaller during the emotional face recognition task. On the contrary, it was relatively smaller for the HC group in the rest stage, and larger during the emotional face recognition.

Zoom In Zoom Out Reset image size

The ratios of the Mayer wave power at emotional face recognition stage to the rest stage, and the statistical results are shown in table 4. There were significant differences in the Mayer wave power ratios of the oxy-Hb and deoxy-Hb over the left prefrontal brain area between the MDD and the HC groups (fearful-rest: left oxy-Hb p   =  0.001, left deoxy-Hb p   =  0.000; happy-rest: left oxy-Hb p   =  0.007, left deoxy-Hb p   =  0.015; sad-rest: left oxy-Hb p   =  0.040, left deoxy-Hb p   =  0.045). Regarding the right prefrontal brain area, we could only find significant differences in the Mayer wave power ratios of the oxy-Hb at the fearful face recognition stages to the rest periods (p   =  0.023), and the happy face recognition to the rest stages (p   =  0.049).

Table 4. The ratios of Mayer wave power during the emotional face recognition to the rest stage for the MDD and the HC groups (mean  ±  standard deviation).

	Prefrontal brain area	The ratio of the Mayer wave power of the oxy-Hb		p  value	The ratio of the Mayer wave power of the deoxy-Hb		p  value
		MDD	HC		MDD	HC
Fearful-rest	Left	0.80  ±  0.45	1.81  ±  1.12	0.001a,d	0.83  ±  0.34	1.29  ±  0.39	0.000a,d
	Right	1.48  ±  1.82	3.03  ±  2.28	0.023a,c	2.03  ±  2.34	1.70  ±  1.68	0.606a
Happy-rest	Left	0.86  ±  0.52	2.28  ±  2.07	0.007b,d	0.79  ±  0.30	1.33  ±  0.87	0.015b,c
	Right	1.74  ±  2.38	4.01  ±  3.11	0.049a,c	3.03  ±  4.93	1.30  ±  1.30	0.143b
Sad-rest	Left	1.03  ±  0.69	2.65  ±  3.24	0.040b,c	0.79  ±  0.32	1.36  ±  1.18	0.045b,c
	Right	1.83  ±  2.24	6.38  ±  10.32	0.068b	3.74  ±  7.19	1.62  ±  1.75	0.216b

^aIndependent-sample t test. ^bMann–Whitney U test. ^cp   <  0.05. ^dp   <  0.01.

3.5. The cross wavelet power spectrum results

The average cross wavelet power spectrum between the oxy-Hb and deoxy-Hb in the left prefrontal brain area of the MDD and HC subjects is shown in figures 5(a) and (b), respectively. For the MDD subjects, the high power area focused in the time period of 8–128 s at the emotional face recognition stages, while it was distributed between the period of 16–128 s for the HC group, and it was obviously larger than the MDD patients. The distributions of the cross wavelet power spectrum between the oxy-Hb and deoxy-Hb over right prefrontal area, and between the bilateral oxy-Hb and the bilateral deoxy-Hb for the MDD and HC groups showed the consistent characteristics. The figures are omitted here.

Zoom In Zoom Out Reset image size

The average cross wavelet coefficients at different stages (fearful, happy and sad face recognition stage, and rest stage) within the frequency band of 0.03–0.2 Hz were statistically compared by independent-sample t test between the MDD and the HC groups, and the results are shown in table 5. We could see no significant difference in the mean cross wavelet coefficients between the MDD and the HC groups at the rest stage. In contrast, the coefficient between the oxy-Hb and the deoxy-Hb of the MDD group were significantly smaller than the HC group over the left prefrontal area (fearful: p   =  0.007, happy: p   =  0.005, and sad: p   =  0.049) and right prefrontal brain area (fearful: p   =  0.001, happy: p   =  0.003, and sad: p   =  0.048), and also the coefficient between the bilateral prefrontal oxy-Hb (fearful: p   =  0.000, happy: p   =  0.000, and sad: p   =  0.001) during the emotional face recognition tasks. The coefficients between the bilateral prefrontal deoxy-Hb of the MDD group were statistically smaller than the HC group only during the fearful (p   =  0.046) and happy face recognition tasks (p   =  0.047).

Table 5. Average cross wavelet coefficients during the emotional face recognition and the rest stages for the MDD and the HC groups (mean  ±  standard deviation).

Signal	Stage	Average cross wavelet coefficients		p  value
		MDD	HC
Left prefrontal oxy-Hb and deoxy-Hb	Rest	36.34  ±  28.37	40.19  ±  15.45	0.546
	Fearful faces	20.69  ±  10.13	29.26  ±  11.63	0.007b
	Happy faces	22.57  ±  13.03	32.94  ±  12.55	0.005b
	Sad faces	21.61  ±  11.01	27.43  ±  9.72	0.049a
Right prefrontal oxy-Hb and deoxy-Hb	Rest	33.66  ±  20.18	36.00  ±  16.46	0.648
	Fearful faces	21.72  ±  8.28	31.13  ±  11.37	0.001b
	Happy faces	23.41  ±  11.49	32.62  ±  9.91	0.003b
	Sad faces	23.29  ±  10.98	29.05  ±  9.42	0.048a
Bilateral prefrontal deoxy-Hb	Rest	36.03  ±  15.23	38.27  ±  22.79	0.679
	Fearful faces	23.85  ±  10.36	29.60  ±  9.91	0.046a
	Happy faces	26.50  ±  14.25	34.54  ±  14.23	0.047a
	Sad faces	26.00  ±  13.40	31.52  ±  13.84	0.150
Bilateral prefrontal oxy-Hb	Rest	31.82  ±  25.43	42.71  ±  20.31	0.094
	Fearful faces	19.70  ±  8.25	34.77  ±  13.53	0.000b
	Happy faces	21.18  ±  10.11	34.61  ±  11.68	0.000b
	Sad faces	20.47  ±  9.48	29.76  ±  9.98	0.001b

^ap   <  0.05. ^bp   <  0.01.

3.6. Wavelet coherence results

The average wavelet coherence between the oxy-Hb and the deoxy-Hb over the left prefrontal brain area of the MDD and the HC subjects is shown in figures 6(a) and (b), respectively. For the MDD subjects, the significantly higher coherence areas did not center in period and time, and varied with different emotional face recognition task. In contrast, the significantly higher coherence areas of HC group were larger, and mainly focused in the period of 1–4 s and 16–32 s. The distributions of the wavelet coherence between the oxy-Hb and deoxy-Hb over right prefrontal brain area, and between the bilateral oxy-Hb, and the bilateral deoxy-Hb for the MDD and the HC groups showed the similar characteristics. The figures are omitted here.

Zoom In Zoom Out Reset image size

The results of the average wavelet coherence coefficients at different stages (fearful, happy and sad face recognition stage, and rest stage) within the frequency band of 0.03–0.2 Hz, and the statistical analysis results between the MDD and the HC groups by independent-sample t test are shown in table 6. We could see significant differences between the MDD and the HC groups in the coefficients between the oxy-Hb and the deoxy-Hb over the left prefrontal area (fearful: p   =  0.044, happy: p   =  0.004, and sad: p   =  0.007) and between the bilateral oxy-Hb (rest: p   =  0.001, fearful: p   =  0.006, happy: p   =  0.003, and sad: p   =  0.002).

Table 6. Average wavelet coherence coefficients between deep signals during the emotional face recognition stages and the rest stages for the MDD and the HC groups (mean  ±  standard deviation).

Signal	Stage	Average wavelet coherence coefficients		p  value
		MDD	HC
Left prefrontal oxy-Hb and deoxy-Hb	Rest	0.51  ±  0.07	0.55  ±  0.08	0.053
	Fearful faces	0.48  ±  0.10	0.53  ±  0.10	0.044a
	Happy faces	0.46  ±  0.11	0.54  ±  0.08	0.004b
	Sad faces	0.48  ±  0.09	0.55  ±  0.09	0.007b
Right prefrontal oxy-Hb and deoxy-Hb	Rest	0.48  ±  0.11	0.50  ±  0.07	0.407
	Fearful faces	0.45  ±  0.08	0.48  ±  0.08	0.108
	Happy faces	0.47  ±  0.09	0.51  ±  0.08	0.144
	Sad faces	0.46  ±  0.08	0.50  ±  0.09	0.075
Bilateral prefrontal deoxy-Hb	Rest	0.59  ±  0.13	0.60  ±  0.11	0.703
	Fearful faces	0.56  ±  0.12	0.61  ±  0.09	0.111
	Happy faces	0.60  ±  0.12	0.61  ±  0.10	0.661
	Sad faces	0.60  ±  0.12	0.61  ±  0.12	0.629
Bilateral prefrontal oxy-Hb	Rest	0.52  ±  0.10	0.62  ±  0.11	0.001b
	Fearful faces	0.52  ±  0.10	0.60  ±  0.11	0.006b
	Happy faces	0.53  ±  0.09	0.62  ±  0.09	0.003b
	Sad faces	0.53  ±  0.12	0.62  ±  0.10	0.002b

^ap   <  0.05. ^bp   <  0.01.

Applying a portable neuroimaging system using near-infrared spectroscopy, the present study revealed hemodynamic changes in the left prefrontal cortex showed significant differences in the median and the Mayer wave power ratios during the emotional face recognition in a sample of patients diagnosed with MDD compared to well-matched HC subjects. We also found the mean cross wavelet coefficients and the average wavelet coherence coefficient between the oxy-Hb and the deoxy-Hb over the left prefrontal cortex, and between the bilateral oxy-Hb in the MDD patients were significantly lower than the HC group, demonstrating abnormal locally functional connectivity over the left prefrontal cortex, and the inter-hemispheric connection between the bilateral prefrontal cortices. Thus, the hemodynamic changes over left prefrontal cortex and between the bilateral prefrontal cortices provided the diagnostic biomarkers for the depression. These findings also supported the application of excitatory (high-frequency) repetitive transcranial magnetic stimulation (rTMS) over the left dorsolateral prefrontal cortex in order to normalize the hypoactivity in the left prefrontal cortex for patients with depression, which is by far the most prevalent strategy for modulating brain activity in depression therapy [31].

Emotional face recognition task reflects the subjects' perception, attention and response ability to emotional faces. From a behavioural point of view, the patients performed worse across all trials compared to the controls (table 2). The accuracy in the MDD group was significantly lower than that of the HC group, while the response time of the MDD patients was statistically higher than that of the HC group. The results are in line with the findings that the depressed subjects recognized expressions both more slowly and less accurately than the HC [32], and also the previous findings of biased interpretation of emotional stimuli in the MDD [33]. Some studies supported that adolescents and young adults with depression demonstrated a speed-accuracy trade-off in that they responded faster to emotional stimuli but had more false, than the HC, which might be associated with the psychomotor retardation in the MDD patients [34]. Our findings contradicted previous work, showing that slower reaction time in the facial expression recognition, which indicated that the patients' misclassifications did not result from a trade-off between the accuracy and speed and thus did not reflect an impulsive response style. It might be related to the perception decline to the surroundings, and delayed response time of the MDD patients. In addition, the response pattern to emotional faces showed similar tendency in both groups pointing to a positivity bias in attention control with faster reaction and higher accuracy to happy faces than to faces with negative expressions as shown in table 2 (i.e. fearful and sad), suggesting faster attentional allocation to positive emotional stimuli.

The median filter preserves the slow change components in the blood oxygen signals, excluding the influence of fast variable components such as heart pulsation, respiration, and etc. The median changes can reflect the brain oxidative metabolism changes of the scalp and the skull. The statistical results of the median difference of blood oxygen signal in table 3 presented that the median of oxy-Hb over the prefrontal cortex of the MDD group during the emotional face recognition was smaller than the rest stages, while the deoxy-Hb was larger than the rest stages, indicating the amount of oxygen consumed in the prefrontal cortex was more than the oxygen supply for the MDD patients during the task. On the contrary, the median value results for the HC group exhibited the opposite changes, revealing that the oxygen consumption of the healthy people was less than the oxygen supply at the task stage. Thus, the oxygen supply of healthy people was insufficient for the MDD patients, suggesting the abnormal oxidative metabolism of the MDD patients over the prefrontal cortex when performing the task of emotional face recognition.

Mayer waves are defined as spontaneous hemodynamic oscillations in arterial pressure with a frequency around 0.1 Hz. On the other hand, Zhang et al [35] found that Mayer wave fluctuations in the anterior region are relatively 'isolated' from the other regions. A study by Kirilina et al [36] has shown that task-evoked Mayer wave oscillations in the prefrontal cortex are not localized distributed in the scalp draining veins. Thus the Mayer power could represent the local activity in the prefrontal activity to some extent. The statistical results of the Mayer wave power ratios during emotional face recognition to the rest stages showed that the Mayer wave power ratios of the prefrontal oxy-Hb, and the left prefrontal deoxy-Hb for the MDD group were significantly lower than that of the HC group. The findings indicate the MDD patients had the relatively weaker local spontaneous hemodynamic oscillations over the bilateral prefrontal cortex during emotional face recognition, and it was more obvious in the left prefrontal cortex.

The cross wavelet transform exposes regions with high common power and further reveals information about the phase relationship [29]. As shown in table 5, the mean cross wavelet coefficients between the oxy-Hb and the deoxy-Hb over the left and right prefrontal cortex respectively in the MDD patients were significantly lower than those in the HC group, indicating relatively smaller functional connectivity existing in the left and right prefrontal cortex respectively for the MDD patients when performing emotional face recognition tasks. Meanwhile, the average cross wavelet coefficients between the bilateral oxy-Hb for the MDD patients were statistically lower than those in the HC group while performing the emotional face recognition tasks, suggesting that the MDD patients had the functional connection dysfunction between the bilateral prefrontal cortex.

Considering the correlation calculated from the cross wavelet transform analysis is related to the amplitude of the signals, while the local correlation obtained by the wavelet coherence analysis is independent of the amplitude, we proposed to quantitatively measure the connectivity between the two signals combining the cross wavelet transform with the wavelet coherence analysis. The wavelet coherence can be thought of as the local correlation between the two time series in time frequency space. Where cross wavelet transform unveils high common power, wavelet coherence finds locally phase locked behaviour [28]. The average wavelet coherence coefficients of the MDD patients between the oxy-Hb and deoxy-Hb over the left prefrontal cortex, and between the bilateral oxy-Hb over the prefrontal cortex were significantly lower than those in the HC group during the emotional face recognition, indicating abnormal function within left prefrontal cortex brain network and bilateral prefrontal cortex network for the MDD patients, respectively. Consequently, the mean cross wavelet coefficients and the mean wavelet coherence coefficients between the oxy-Hb and the deoxy-Hb over the left prefrontal cortex and between the bilateral oxy-Hb over the prefrontal cortex during the emotional face recognition task can be combined to use as quantitative functional connectivity indicators to distinguish the MDD patients from the HC subjects.

According to the analysis above, we could find the left prefrontal cortex showed less active in the MDD patients than in HC during processing of emotional stimuli, supporting the findings of hypoactivation over left-hemispheric dorsolateral prefrontal cortex in the previous studies [37, 38] and whole-brain meta-analysis comparing with the neural reactions of MDD patients and HC subjects in response to positive and negative stimuli [39, 40]. One cardinal function of the left dorsolateral prefrontal cortex is the top-down cognitive-emotional control of affect-generating structures, such as the amygdala, ultimately modulating sensory processing of affective material using suppression, attention redirection or reappraisal strategies [41–44]. Consequently, the MDD patients might use limited emotion regulation strategies during emotional perception. Furthermore, excitatory rTMS over the left dorsolateral prefrontal cortex is approved by the U.S. Food and Drug Administration for the MDD treatment, in order to restore excitability of prefrontal cortex regions [45, 46]. Our findings also provided evidence for the rTMS over the left prefrontal cortex.

As a study limitation, it should be noted that the results from the analysis on the data of six channels located in the prefrontal cortex could not reflect the brain function over the entire prefrontal cortex.

In conclusion, the present study revealed the significant hemodynamic changes over left prefrontal cortex and between the bilateral prefrontal cortices during emotional face recognition obtained from a portable neuroimaging system using near-infrared spectroscopy for the MDD patients compared with the HC subjects. The study demonstrates the feasibility of using fNIRS for objective MDD diagnosis in clinic. And also supports the rationale for use of transcranial magnetic stimulation over the left dorsolateral prefrontal cortex in the MDD patients.

This work was supported by the National Natural Science Foundation of China (61471291, 61601361, and 31571000), and Natural Science Foundation of Shaanxi Province in China (2017JM6013). The authors have confirmed that any identifiable participants in this study have given their consent for publication.

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.