2.1.

Participants

Twenty-nine healthy Japanese volunteers (12 females and 17 males, mean $\mathrm{age}=35.9\pm 7.7$) and twenty-one return-to-work (RTW) trainees in remission of mental disorders with depressive symptoms (5 females and 16 males, mean $\text{age}=40.9\pm 6.31$) participated in this study. The RTW trainees have attended this trial instead of attending some kind of normal training class for returning to work, which is given by clinical psychologists in an RTW facility. The facility obtains statements about each trainee’s mental states (e.g., depressive symptoms) and the will to return to work but does not obtain other private information such as education level and treatment, which are managed by the doctors in charge who are in different hospitals/clinics. Written informed consent was obtained from all participants after they were informed of the purpose, procedures, risks, benefits, and voluntary nature of the experiments. This study was conducted according to the regulations of the internal review board of the Central Research Laboratory, Hitachi, Ltd.

2.2.

Mood State Measurement

The participants’ mood state was assessed using the Japanese edition of a short form of the POMS^™.^26,27 The participants rated their mood using 30 mood-related adjectives (angry, sad, vital, etc.) on a 5-point scale ranging from 0 (“not at all”) to 4 (“extremely”). This rating enabled us to estimate six mood-related measures: tension-anxiety (T-A), depression-dejection (D), anger-hostility (A-H), vigor (V), fatigue (F), and confusion (C).

2.3.

Prefrontal Cortex Activity Measurement

We used two types of multichannel functional NIRS system for measuring the changes in the cerebral blood concentration of the participants.^11,12,29,30 One was a commercial model of an NIRS system (ETG-4000, Hitachi Medical Corp.),^11,12 and the other was a prototype of a wearable optical topography (WOT) system developed by the Central Research Laboratory, Hitachi, Ltd.^29,30 The former was used to measure healthy controls in an experimental room, and the latter was for used for participants in the RTW program because of its portability (we could use it at the other site where the program was operated). Each NIRS device emits near-infrared light (2 mW) at two wavelengths onto the scalp and detects the reflected light. The NIRS method measures changes in the product of hemoglobin (Hb) concentration and effective optical path length in human brain tissue. The unit of Hb change is molar concentration (mM) multiplied by optical path length (mm). The changes in oxygenated Hb (${\mathrm{HbO}}_2$) and deoxygenated Hb (HHb) signals are estimated using absorption coefficient spectra for the two wavelengths. Both systems have the same source–detector distance (3 cm) and the same laser power (2 mW). The differences of both systems were the sampling rate (the former was 10 Hz, and the latter was 5 Hz) and the pair of wavelengths (the former used 695 and 830 nm, and the latter used 754 and 830 nm). However, the waveforms of ${\mathrm{HbO}}_2$ and HHb measured with the latter prototype were confirmed as almost the same as those of the commercial model.^29,30 The ETG-4000 and WOT systems cover the entire forehead, and the measurement positions of each system were estimated by a probabilistic method used in previous studies.^31–33 The channel positions on the prefrontal area of both systems are shown in Fig. 1; however, the positions were not completely the same, so we chose the five channel positions numbered 1 to 5 that are closest to each other in the PFC for analysis.

Fig. 1

Channel positions of NIRS systems. Lilac and red circles indicate positions of ETG-4000 and WOT systems on cerebral cortex, respectively. Each channel position is defined as center point of each light source and detector. Five channels numbered 1 to 5 and surrounded by dashed line in prefrontal area were used for analysis.

2.4.

Working Memory Tasks

We used verbal and spatial delayed match-to-sample WM tasks to measure the PFC activities [Fig. 2(a)]. These were the same as those used in previous NIRS studies.^18–22 Each task trial started with a 1.5-s presentation of the target stimuli (target) on a PC screen [Fig. 2(b)]. After presenting a delay (black screen with white fixation cross) for the following 7 s, a test stimulus for retrieval (probe) was then presented for 2 s or until the participant responded. The participant responded by pressing a button on a handheld game pad connected to the PC, and the test stimulus disappeared when the button was pressed within 2 s. The intervals between the trial finish and the target onset in the next trial were randomized from 16 to 22 s for avoiding participant habituation or prediction of the target onset. For a verbal WM task, a set of four Japanese hiragana characters was presented as the target, and a Japanese katakana character was presented as the probe. The participants were asked to judge whether the probe character corresponded to one of the four target characters. For the judgment, they used their right index and middle fingers to press the “yes” and “no” button, respectively. Because the characters presented as target and probe were in different Japanese morphograms, the participants made their decisions on the basis of the phonetic information conveyed by the characters, not their form. Auditory cues of 1000- and 800-Hz pure tones of 100-ms duration were presented at the target and probe onsets, respectively. For a spatial WM task, the target was given by the locations of four squares randomly arranged at eight peripheral locations. The probe was given by presenting a square at one of the eight locations. The participant was asked to judge whether the probe square location corresponded to one of the target locations.

Fig. 2

(a) Spatial and verbal WM tasks and (b) time sequence of a trial.

2.5.

Experimental Procedure

The participants were seated in a comfortable chair in a quiet room. Their mood states were assessed with POMS^™26,27 at the beginning of the experiment. Next, the participants received computer-automated instructions that were followed by a brief practice session to familiarize the participants with the tasks. The practice session included two to five trials of each task depending on the participants’ familiarization with the tasks. Thereafter, NIRS measurements were conducted while the participants performed the WM tasks. The tasks were organized into two sessions, one for the verbal WM task and the other for the spatial WM task, with a counterbalanced order across participants. Each session included five trials of either one of the tasks, and the sessions were separated by a short break ($\sim 1\min$). The duration of the NIRS measurements was $\sim 15\min$, and the whole experiment took about 30 min.

2.6.

Near-Infrared Spectroscopy Data Analysis

For the NIRS data analysis, we used MATLAB (The MathWorks, Inc.) and the plug-in-based software we developed. We defined a block for each signal as the period from 1 s before presenting the target (which lasts 1.5 s followed by a 7-s delay for memorization) to 16 s after starting presenting the test, 25.5 s in total. We conducted five blocks of spatial WM tasks and five blocks of verbal WM tasks. To remove the components originating from slow fluctuations and the high-frequency noise of blood flow, a 1.5-Hz low-pass filter and 0.02-Hz high-pass filter were applied to the signals. Then, the ${\mathrm{HbO}}_2$ signal in each block was baseline corrected by subtracting the average value of the first 1 s. Finally, we defined the PFC activity period as the period from 5 s after presenting the target to just before starting presenting the test (3.5 s).¹⁶ We focused on the ${\mathrm{HbO}}_2$ signal in the statistical analysis as the previous studies found main effects in this signal.^18–20 We used the $t$-statistic value of ${\mathrm{HbO}}_2$ signals during the PFC activity period from the five blocks for spatial and verbal WM tasks and denote them as ${\mathrm{HbO}}_2^{\mathrm{spat}}$ and ${\mathrm{HbO}}_2^{\mathrm{verb}}$, respectively, as the measures estimating the PFC activity for each channel.

2.7.

Modeling Indication of Healthy Control/Return-to-Work Trainee

To estimate the indication of the difference between healthy controls and RTW trainees, we applied multivariable logistic-regression analysis to the POMS^™ scores, ${\mathrm{HbO}}_2^{\mathrm{spat}}$, and ${\mathrm{HbO}}_2^{\mathrm{verb}}$ (Ch 1 to 5, see Fig. 1). In this analysis, each participant’s data were labeled with a binary value, which is 1 if the participant is a healthy control and 0 if the participant is an RTW trainee. Next, manual backward selection was used to eliminate nonsignificant components from each logistic-regression model and repeat the regression analysis using only significant component(s) ($p<0.05$).

To evaluate the performance of logistic-regression models from the three sets of POMS^™, ${\mathrm{HbO}}_2^{\mathrm{spat}}$, and ${\mathrm{HbO}}_2^{\mathrm{verb}}$, we used ROC analysis.²⁶ The area-under-curve (AUC) was used for quantitative comparison of the ROC curve obtained from those logistic-regression models, and its 95% confidence interval (CI) was calculated by bootstrap estimation with 10,000 replications.

3.1.

Time Courses of HbO₂ and HHb Signals

Figure 3 shows the mean time courses of ${\mathrm{HbO}}_2$ and HHb signals of two channels (Ch 1 and 5) in response to spatial and verbal WM tasks for the healthy control and RTW trainee groups as typical waveforms, where PFC activation of these channels (Ch 1 and 5) are statistically significant ($p<0.05$) for the ${\mathrm{HbO}}_2^{\mathrm{verb}}$ of the healthy control, as shown in Table 1. In Fig. 3, the horizontal and vertical axes indicate the time in which the zero value means the target onset and the ${\mathrm{HbO}}_2$ and HHb signals, respectively. Changes in the ${\mathrm{HbO}}_2$ signals during spatial and verbal WM memory tasks for both groups increase after the target onset and decrease around the baseline in the several seconds after the probe stimuli, and those in the HHb signals are almost flat compared to those in the ${\mathrm{HbO}}_2$ signals. However, the amplitudes in the first peak (around 7 to 9 s) after the target onset for the RTW trainees are less than half of those for the healthy controls.

Fig. 3

Mean time courses of ${\mathrm{HbO}}_2$ and HHb signals of two representative channels (Ch 1 and 5) to spatial and verbal WM memory tasks for (a) healthy controls and (b) RTW trainees.

Table 1

The t-statistic value of PFC activity (HbO2) of all participants.

3.2.

Logistic-Regression Models of Prefrontal Cortex Activity and Profile of Mood States

The logistic-regression results from ${\mathrm{HbO}}_2^{\mathrm{spat}}$ and ${\mathrm{HbO}}_2^{\mathrm{verb}}$ (calculated from ${\mathrm{HbO}}_2$ signals) are summarized in Table 2. This table shows the regression coefficients and $p$ values of the target channels. No significant components are observed in the results from ${\mathrm{HbO}}_2^{\mathrm{spat}}$. The statistically significant components are Ch 1 and 3 of ${\mathrm{HbO}}_2^{\mathrm{verb}}$. We applied logistic-regression analysis to those two channels again since we assumed that the other channels contribute less to discriminate the RTW trainee group and healthy control group than those two channels (Ch 1 and 3), and then we obtained significant results (Ch 1: regression $\mathrm{coef}=0.19$, $p=0.0065$, Ch 3: $-0.70$, $p=0.035$).

Table 2

Logistic-regression results for HbO2 data.

The logistic-regression results from POMS^™ are shown in Table 3. The statistically significant component is the vigor score (POMS_V). We applied logistic-regression analysis to only POMS_V, and then we obtained significant results (POMS_V: regression $\mathrm{coef}=0.29$, $p=0.0062$).

Table 3

Logistic-regression results for POMS™ data.

3.3.

Receiver Operating Characteristic Analysis

To assess the performance of discrimination between healthy controls and RTW trainees based on logistic-regression models of ${\mathrm{HbO}}_2^{\mathrm{verb}}$ and POMS_V, we used ROC curve analysis. One ROC curve shows the relationship of sensitivity versus 1 — specificity when the criterion to discriminate the binary state based on the designed model is changed; thus, the model with higher sensitivity and higher specificity is closer to the upper left corner. Figure 4 shows the ROC plots derived from ${\mathrm{HbO}}_2^{\mathrm{verb}}$ and POMS_V. The solid diagonal line from (0, 0) to (1, 1) indicates the no-discrimination line (chance level), and its AUC is equal to 0.5. In Fig. 3, the ROC curves of both ${\mathrm{HbO}}_2^{\mathrm{verb}}$ and POMS_V fall between the upper left corner and the no-discrimination line. The AUC of ${\mathrm{HbO}}_2^{\mathrm{verb}}$ is 0.792 (95% CI 0.651 to 0.918) and that of POMS_V is 0.755 (95% CI 0.591 to 0.891), indicating moderate power for discrimination, but the AUC of ${\mathrm{HbO}}_2^{\mathrm{verb}}$ is slightly higher than that of POMS_V. To estimate the best discrimination point of each AUC curve, the point nearest to the upper left point was selected and marked with a circle on the curve. At the best discrimination point of ${\mathrm{HbO}}_2^{\mathrm{verb}}$, the sensitivity and specificity are 75.9% and 81%, respectively. At that of POMS_V, those values are 90% and 57.1%, respectively.

Fig. 4

ROC curves.