42 Asian (Korean) participants (22 females, aged 22.357±2.639 years) took part in the experiment. Exclusion criteria included a history of psychiatric or neurological disorders, pregnancy, hearing problem, and a history of any blood diseases. One additional participant was excluded because she could not finish the experiment due to a claustrophobia-like symptom, which she had never experienced before. Although all participants have never been diagnosed as a patient before, the range of the BAI scores of the participants in the experiment was from 1 to 19 (please see ‘Psychometric tests’ below). This includes not only healthy/minimal (BAI score of 0~7) but also mild (8~15) and moderate (16~25) anxiety level. All participants were right-handed. They provided written informed consent for the procedure in accordance with protocols approved by the KAIST Institutional Review Board.

For the Asian face condition experiment, face images from 2 Asian men and 2 Asian women were used, and for the non-Asian face condition experiment, face images from 2 African-American men, 2 African-American women, 2 Caucasian men, and 2 Caucasian women were used. The African American and Caucasian face images were selected from Nimstim Face Stimulus Set (https://www.macbrain.org/resources.htm) [26]. For each face identity, the images of prototypical angry, happy, sad, and neutral facial expression were prepared. For Asian face stimuli, 1 female Asian face image was selected from Nimstim Face Stimuli Set [26] and the other images were obtained by taking pictures of volunteered actors. The volunteers were informed that photos of them will be used only for research and provided a written informed consent. In Fig. 1, the example face images that may be published from Nimstim Face Stimulus Set, were used for illustration purpose only and were not used in the experiment.

For each face identity, a graded stimulus series was generated by morphing faces with one of the prototypical emotional faces (angry, happy, or sad) and neutral face. Thus, a total of 11 intensities, from 0% emotional (neutral) to 100% emotional (prototypical angry, happy, or sad) were prepared for each emotion category of each identity (Fig. 1A). However, to measure the recognition sensitivity for facial emotions, we used only neutral face image and 8 steps of emotional images, starting with 10% emotional and ending with 80% emotional faces, for each emotion category of each identity in the experiment. Face Morpher (https://github.com/alyssaq/face_morpher) and STOIK Morph Man software (http://www.stoik.com/products/video/STOIK-Morph-Man-2016/) were used to morph the face images and the contours of the faces were post-edited with Adobe Photoshop CC 2017. The size of each face image was set to 500×600 pixels.

Participants conducted the emotion recognition task in a sound-attenuated experiment room. The tasks were performed in 2 separate blocks: Asian face condition block and non-Asian face condition block. In the Asian face condition block, only Asian face stimuli were presented while the face stimuli except for Asian faces were used for the non-Asian face condition block.

Each trial of the emotion recognition task consisted of a face phase and a subsequent orientation phase. In the face phase, the participants first saw a face stimulus and indicated their recognition of emotion category by pressing one of the four buttons (1=Angry, 2=Happy, 3=Sad, and 4=Neutral) (Fig. 1B). The participants were instructed to press the button as accurately and quickly as possible with a maximum response time of 2.5 sec. The face image was shown up until the response was recorded. After a 1-sec interval, the orientation phase began. In this phase, the participants were asked to determine the precise orientation of a grating while they were seeing one of four orientation gratings (0°, 45°, 90°, 135°) (Fig. 1B). This orientation phase was inserted to prevent any effect of the previously shown face stimulus on the recognition of the next facial emotion. The participants exhibited strong performance in this orientation phase: 93.744±15.478 % for Asian face condition and 95.359±8.950 % for non-Asian condition. In both phases, each image of the face or grating subtending approximately 8° of visual angle was viewed via a monitor (1920×1080 resolution, 60 Hz refresh rate), and the choice alternatives were displayed under each image.

After the tasks, participants completed K-BAI (Korean version of Beck Anxiety Inventory) and K-BDI-II (Korean version of Beck Depression Inventory) questionnaires (www.koreapsy.co.kr). The BAI and BDI are self-report inventories created by Aaron T. Beck for measuring the severity of anxiety and depression [27,28]. BAI scores in range 0~9 mean minimal anxiety, 10~16 mean mild anxiety, 17~29 mean moderate anxiety and 30~63 mean severe anxiety. BDI scores in range 0~9 indicate minimal depression, 10~18 indicate mild depression, 19~29 indicate moderate depression, and 30~63 indicate severe depression. The range of the BAI scores of the participants was from 1 to 19 (mean=5.26, std.=4.79), and the range of the BDI scores was from 0 to 23 (mean=9.64, std.=5.53).

To investigate the sensitivity of the three different facial emotions, we defined all faces used in the experiment except for the neutral faces as emotional (angry, happy, or sad) faces as a previous study [9]. Thus, the correct answer for the emotionally graded faces (from Step 1 to 8) in the angry, happy, or sad category was angry, happy, or sad, respectively. The percentage of the correct answers for the faces of each emotion category was measured.

To compare the average answer rate for each of the emotions with the chance level (25%), two-tailed t-test was used. Repeated-measures ANOVAs (tests of within-subjects effects) were used to determine the statistical significance of the stimuli race type effects. For all ANOVAs with factors having more than two levels, Greenhouse-Geisser Corrections were used. For the correlation analysis, one-tailed Spearman correlation was used with the assumption of predicted positive direction. Statistical analysis was done with MATLAB and SPSS.

To investigate the recognition of emotional facial expressions, we asked participants to conduct four alternate forced choice tasks in which they judged whether expression of a presented face was angry, happy, sad, or neutral. Each presented face image was one of the graded face images ranging from neutral (0% emotional intensity) to 80% angry, happy or sad expression (100% emotional intensity indicates prototypical angry, happy, or sad expression) (Fig. 1). There were 8 steps of emotional intensity by 10% increase from 10% to 80% for each emotion category (e.g., step 5 of angry face indicates 50% angry face).

We first examined how sensitively the participants detected the Asian, African-American, or Caucasian facial expressions of emotion. The rate of accurate responses (averaged rate across the rates from 8 steps for each emotion category) was greater than chance accuracy in all emotion categories (angry, happy, sad) and all races of face stimuli (all t(41)>6.773, p<0.001) (Fig. 2A), indicating that the participants successfully performed the task. To directly compare the responses between different races of face stimuli in each emotion category, we conducted a two-way ANOVA with races of stimuli (Asian, African-American, Caucasian) and emotion categories (angry, happy, sad) as within-subject factors. This analysis revealed a main effect of races of stimuli (F(1.887, 77.369)=96.049, p<0.001), emotion categories (F(1.862, 76.334)=17.942, p<0.001) and a significant interaction between race and emotion (F(3.584, 146.935)=11.240, p<0.001). A follow-up post-hoc Bonferroni pairwise comparison showed that the recognition accuracy for Asian faces was greater than the accuracy for Caucasian in all emotion categories (p<0.001 for angry, p<0.001 for happy, p<0.001 for sad) (Fig. 2A). In addition, the accuracy for Asian faces was significantly greater than African-American in angry (p<0.001) and happy conditions (p=0.028) (Fig. 2A). Thus, the accuracy for emotional facial expression was better for Asian faces than non-Asian faces. Given that the participants were Asians, this result is consistent with the prior research that people are generally better at the recognition of facial expressions for the faces from in-group than from other groups [22,24].

We also analyzed the recognition accuracy for the faces with relatively low-intensity (faces with 20%, 30%, and 40% emotional intensity) and high-intensity (faces with 60%, 70%, and 80% emotional intensity) emotional expressions separately (Fig. 2B). For both low- and high-intensity emotional faces, we found the same tendency observed in the responses for total faces; a two-way repeated-measure ANOVA with races and emotion categories as factors revealed a main effect of races (F(1.986, 81.422)=93.571, p<0.001 for low-intensity; F(1.886, 77.311)=44.024, p<0.001 for high-intensity), emotion categories (F(1.726, 70.763)=19.077, p<0.001 for low-intensity; F(1.315, 53.909)=25.660, p<0.001 for high-intensity) and significant interaction between race and emotion (F(3.184, 130.534)=12.662, p<0.001 for low-intensity; F(3.160, 129.548)=8.988, p<0.001 for high-intensity). Additionally, the accuracy rate for Asian faces were significantly greater than Caucasian faces in all emotion categories for both low- (Bonferroni post hoc tests, p<0.001 for angry, p<0.001 for happy, p=0.015 for sad) and high-intensity (p<0.001 for angry, p=0.002 for sad) except for high-intensity happy faces (p=0.236). Taken together, these results suggest that the sensitivity to facial emotions for the faces from the participant's own-race is higher than that for other race faces.

To investigate the influence of individual anxiety level on the recognition of emotional facial expressions, the participants were asked to respond to the BAI questionnaire. We derived correlation between the mean accuracy for the facial emotion recognition and BAI scores. Because the participants showed the best performance in the emotional face recognition for Asian faces, we first focused on the responses for Asian faces. We found that the recognition accuracy for angry faces was positively correlated with BAI scores (r=0.375, p=0.007) (Fig. 3A). Moreover, we also derived the correlations between BAI scores and the recognition accuracies for low and high intensity angry faces separately, and found the same tendency for both low and high intensities (r=0.266, p=0.044 for low-intensity; r=0.342, p=0.013 for high-intensity) (Fig. 3B). This tendency was also observed while controlling the effect of age or gender. We conducted partial correlation analyses between individual anxiety level and sensitivity to the recognition of angry expression with the control variables of age or gender, and found the same tendency observed in Fig. 3 (Age controlled: r=0.372, p=0.009 for total intensity; r=0.266, p=0.047 for low-intensity; r=0.347, p=0.013 for high-intensity; Gender controlled: r=0.329, p=0.018 for total intensity; r=0.218, p=0.085 for low-intensity; r=0.307, p=0.026 for high-intensity). The positive relationship between the recognition of facial emotions and the BAI scores was not observed in other emotions (Table 1). Thus, these suggest that during the recognition of Asian facial emotion, the recognition accuracy for angry faces, but not other emotional faces, depends on individual anxiety level.

We next examined the relationship between individual anxiety level and recognition of facial emotion for non-Asian faces. To increase the power of non-Asian face results, we combined the responses to African-American faces and Caucasian faces. In non-Asian face condition, we found statistically significant correlation for low-intensity angry faces (r=0.285, p=0.034), but not for high-intensity angry faces (r=0.010, p=0.474) (Fig. 4). This tendency was also observed while controlling for age (r=0.275, p=0.041 for low intensity; r=-0.059, p=0.644 for high intensity), and similar but weaker tendency was observed while controlling for gender (r=0.248, p=0.059 for low intensity; r=0.066, p=0.340 for high intensities). Additionally, there were no significant correlations between the emotion recognition and the BAI scores for other emotion categories regardless of emotional intensities (Table 2). Thus, these showed that while there was a positive relationship between the BAI scores and the emotion recognition for both low- and high-intensity angry faces in the Asian face condition (Fig. 3), during the recognition of non-Asian facial emotions, the positive relationship was observed only for low-intensity angry faces (Fig. 4).

Since anxiety and depression frequently co-occur [29,30], we also investigated the relationship between the sensitivity to facial expressions and the individual depression level measured by BDI scores (Table 3 and 4). However, we did not find any significant effect of individual depression on the recognition of facial emotions for both Asian (Table 3) and non-Asian faces (Table 4). Additionally, we also investigated the relationship between the sensitivity to facial expressions and the BAI scores while controlling individual BDI scores, and found the same tendency shown in the Fig. 2 and 3 (Table 5). These suggest that the perceptual sensitivity to angry expressions depends on mainly individual anxiety level rather than depression level.