Impact of depressed state on attention and language processing during news broadcasts: EEG analysis and machine learning approach

Fuseda, Kohei; Watanabe, Hiroki; Matsumoto, Atsushi; Saito, Junpei; Naruse, Yasushi; Ihara, Aya S.

doi:10.1038/s41598-022-24319-x

Download PDF

Article
Open access
Published: 28 November 2022

Impact of depressed state on attention and language processing during news broadcasts: EEG analysis and machine learning approach

Kohei Fuseda¹^na1^nAff2,
Hiroki Watanabe¹^na1,
Atsushi Matsumoto¹^nAff3,
Junpei Saito¹,
Yasushi Naruse¹ &
…
Aya S. Ihara¹

Scientific Reports volume 12, Article number: 20492 (2022) Cite this article

3396 Accesses
4 Citations
3 Altmetric
Metrics details

Subjects

Abstract

While information enriches daily life, it can also sometimes have a negative impact, depending on an individual’s mental state. We recorded electroencephalogram (EEG) signals from depressed and non-depressed individuals classified based on the Beck Depression Inventory-II score while they listened to news to clarify differences in their attention to affective information and the impact of attentional bias on language processing. Results showed that depressed individuals are characterized by delayed attention to positive news and require a more increased load on language processing. The feasibility of detecting a depressed state using these EEG characteristics was evaluated by classifying individuals as depressed and non-depressed individuals. The area under the curve in the models trained by the EEG features used was 0.73. This result shows that individuals’ mental states may be assessed based on EEG measured during daily activities like listening to news.

Brainwave activities reflecting depressed mood: a pilot study

Article Open access 04 September 2023

Technical and clinical considerations for electroencephalography-based biomarkers for major depressive disorder

Article Open access 25 October 2023

Neural activity during inhibitory control predicts suicidal ideation with machine learning

Article Open access 08 July 2024

Introduction

In the current era of information overload, individuals receive an excessive amount of information daily, regardless of their wish to receive such news. While information enriches daily life, it can also have a negative impact such as development of stress and eventually mental illness, depending on an individual’s mental state. Negative information can influence human mental health negatively^1,2. For example, online surveys for college students revealed that more exposure to information on COVID-19 was associated with more significant self-reported depression³ and anxiety⁴. To promote the comfortable and healthy use of information for everyone, it is important to assess the individual’s mental state objectively when they are receiving information. In this study, we focused on brain activity measured during the receipt of information while participants were in a depressed state. Supposing the individuals’ states of depression may be predicted from their brain activities, this work could lead to an application based on EEG that encourages behavioral change; for example, reducing the amount of time spent on exposing oneself to negative information via the Internet or social media.

Focusing on how individuals attend to the negative information may contribute to our purpose. Attentional bias for negative information (also called negativity bias)^{5,6,7,8,9,10,11} has been considered to contribute to the cause and maintenance of depression^12,13,14,15. Furthermore, such attentional bias has been found in people who are at high risk of depression such as subclinically depressed individuals^16,17, people with remitted depression¹⁸, and children with depressed mothers¹⁹. Pronounced attentional bias for negative information is due to the mood congruency effect^14,20 which entails enhanced information processing when the affective valence of input information is congruent with the information receiver’s mood²¹. Conversely, some reports show that patients with depression²² and subclinically depressed individuals^17,23 reduce the allocation of attentional resources to positive information, whereas other studies found no attentional bias for negative information in patients with depression^24,25,26.

It cannot be ruled out that the inconsistent results of attentional bias for negative information may stem from the short duration of exposure to negative information. Previous studies that focus on attentional bias used traditional psychological paradigms with visual stimuli such as the emotional Stroop task^{24,27,28,29,30}, Posner task^16,17, and dot-probe task^{22,25,31,32,33}, and examined attention to affective stimuli over a short duration (i.e., a few hundred milliseconds to a few seconds^17,23). Koster et al.¹⁷ showed that negative bias occurred in the dysphoric individuals when the stimulus duration was 500 ms and 1500 ms, not when it was 250 ms. Based on these results, in this study, we aimed to investigate the attentional biases of people who have not been diagnosed with mental diseases when exposed to affective natural speech. The task of listening to natural speech enables participants to be exposed to affective stimuli for a longer duration than the traditional paradigms. Furthermore, it allows us to examine information processing in situations that are as close to everyday life as possible, rather than in paradigms designed for experiments. Thus, research using affective natural speech may further shed light on attentional biases for negative information.

In addition, our research focused on the effects of depression on language comprehension processing as depression is reported to cause difficulties in the non-affective processing of information with negative valence. For example, the reaction time to the lexical judgment of negative words was longer for depressed individuals than it was for non-depressed individuals³⁴. This is interpreted by the affective interference theory, which states that prioritizing the processing of negative information causes non-affective processing to function improperly^27,34,35,36. Despite this, whether depression affects the language comprehension process—especially semantic processing—remains unclear.

Therefore, we adopted positive, neutral, or negative radio news broadcasted in Japan (Fig. 1, Supplementary Table S1) as stimuli and investigated whether the attention to affective information differed between depressed and non-depressed individuals who have not been diagnosed with mental diseases. In addition, how attentional bias influences language processing, especially semantic processing, was also elucidated. In this study, participants were divided into depressed individuals and non-depressed individuals based on the Japanese version of the Beck Depression Inventory (BDI)-II³⁷ (Table 1). Based on the above-mentioned items, we formulated the following two hypotheses. First, the attentional bias for negative information is stronger in depressed individuals than it is in non-depressed individuals with exposure to the affective news content. Subsequently, the depressed individuals allocate more attentional resources to negative news than to positive news. Second, due to affective interference, the depressed individuals will require more load on language comprehension (semantic processing on words) while listening to negative news than non-depressed individuals.

Table 1 Characteristics of the depressed and non-depressed individuals.

Full size table

One method of investigating the attentional biases and semantic processing on words is to observe event-related potential (ERP) components time-locked to stimuli onset by measuring electroencephalograms (EEG). For example, the N1 and P2 reflect perceptual processing³⁸ and are negative and positive components that peak at approximately 100 and 200 ms after the onset of the auditory stimuli, respectively. Both responses are exogenous components elicited by the presentation of a physical stimulus. Their amplitudes are affected by the change of attentional resources allocation to the stimuli^38,39,40. Thus, the changes in these responses can be used to examine the perceptual processing characteristics of depression. In contrast to these exogenous components, one of the endogenous components is N400 which is negativity that peaks at approximately 400 ms after the onset of meaningful stimuli. This component reflects semantic processing^41,42,43, with amplitudes that vary based on the ease of accessing information from long-term memory and integrating semantic representations into preceding contexts⁴⁴. Therefore, by using N1, P2, and N400 as indices, differences in perceptual and semantic processing between depressed and non-depressed individuals can be considered.

Despite the effectiveness of measuring ERP for our research purpose, it is generally difficult to observe ERP response time-locked to word onset included in natural speech like news broadcasts. Such natural speech contains only a small interval between words and, thus, EEG responses to each word are overlapped. Calculating temporal response function (TRF) weights⁴⁵ to word onset is one solution to address this issue. TRF describes the linear mapping between ongoing stimuli (i.e., words within the news in case of news stimuli) and ongoing EEG data. Since TRF allows the separation of overlapping responses due to different stimuli close in time, this method is suitable for the analysis of EEG signals to words in natural speech. Our previous study demonstrated that the TRF weights time-locked to word onset included in natural speech showed components corresponding to N1, P2, and N400⁴⁶. Thus, measuring these TRF components corresponding to ERP components possibly allows for quantifying the attentional resource allocation or load on semantic processing on words. To clarify the effect of depression and/or news contents on attentional bias or the load on semantic processing, we investigated group effects (depressed individuals and non-depressed individuals) and news effects (negative, neutral, and positive) on the latency and amplitude of each component.

Another advantage of using the news broadcast as stimuli over those used in the traditional paradigms was that listening to a news broadcast is a daily activity. If a depression differentiates the attentional biases or language processing, the depressed states can be predicted by simply listening to natural speech stimuli in an environment that is close to their reality. Thus, in addition to clarifying the characteristics of depressed individuals’ EEG responses to natural speech stimuli, we verified the usefulness of these responses’ applications. To this end, we evaluated the performance of classifying individuals into depression and non-depression, using EEG features and linear support vector machine (SVM), to investigate the feasibility of detecting depressed individuals.

Results

Subjective evaluation of news

In the EEG experiment, participants performed subjective evaluations on each news item immediately after listening to it, based on the following items: affective valence, arousal, interest, and comprehensibility. We investigated the effect of Group (depressed and non-depressed individuals) and News (positive, neutral, and negative news) on these evaluations using a two-way mixed-design analysis of variance (ANOVA). Concordant with the preliminary survey results (Table 2; see “Materials” for details), news contents influenced affective valence, arousal, interest, and comprehensibility. First, a significant main effect of News on affective valence (F (2, 266) = 554.73, ε = 0.92, p = 9.50 × 10^–88, η_p² = 0.807) was found (Fig. 2a). The affective valence increased in the following order: positive, neutral, and negative news—that for positive news was significantly higher than that for neutral (p = 6.18 × 10^–33) and negative news (p = 1.74 × 10^–59), and that for neutral news was significantly higher than that for negative news (p = 4.41 × 10^–40).

Table 2 Characteristics of news selected based on the preliminary survey (N = 160).

Full size table

Second, the main effect of News on arousal was significant (F (2, 266) = 27.02, ε = 0.95, p = 6.57 × 10^–11, η_p² = 0.169) (Fig. 2b), indicating that the negative and positive news showed significantly higher arousal than the neutral ones did (p = 2.35 × 10^–10 and p = 1.97 × 10^–8, respectively). Importantly, for our experiment that focused on the attentional bias regarding negative information, no significant difference was noted between the positive and negative news (p = 0.697). In addition, a significant main effect of Group was found (F (1, 133) = 7.79, p = 0.006, η_p² = 0.055): arousal in the non-depressed individuals was significantly higher than that in depressed individuals. The subjective evaluation results showed that negative news was evaluated as negative content and positive news as positive content, confirming the validity of the news conditions used in this study. In addition, as no difference in arousal between positive and negative news was noted, it was ensured that the difference in EEG responses when listening to positive and negative news was not due to arousal.

Moreover, a significant main effect of News on interest (F (2, 266) = 22.91, p = 6.64 × 10^–10, η_p² = 0.147) was found (Fig. 2c), revealing that the negative news showed significantly higher interest than the positive (p = 3.83 × 10^–4) and neutral ones did (p = 2.12 × 10^–10) and that the positive news was more interesting than neutral news was (p = 0.021).

Finally, a significant main effect of News on comprehensibility (F (2, 266) = 114.63, p = 1.25 × 10^–36, η_p² = 0.463) was found (Fig. 2d), revealing that the positive news was significantly more comprehensible than the negative (p = 7.29 × 10^–10) and neutral (p = 3.14 × 10^–28) items and the negative news was more comprehensible than the neutral news (p = 7.37 × 10^–15). In addition, a significant main effect of Group (F (1, 133) = 20.86, p = 1.11 × 10^–4, η_p² = 0.136) was found: news was more comprehensible to the non-depressed individuals than to the depressed individuals. There was no interaction between News and Group for affective valence (p = 0.105), arousal (p = 0.758), interest (p = 0.258), or comprehensibility (p = 0.075).

EEG responses while listening to news

To investigate brain activity in response to each piece of news, we calculated TRF weights that describe the linear mapping between ongoing stimuli (words within news) and ongoing EEG data. The grand-averaged TRF waveforms clearly showed three components, corresponding to N1, P2, and N400 as reported in ERP studies (Fig. 3a). To clarify the effect of the group and/or news contents on these types of processing, we performed the two-way mixed-design ANOVA, using the between factor of Group and within factor of News, on the latency and amplitude of each component.

Peak latencies

Significant interactions, Group × News, were found for the peak latencies of N1 (F (2, 266) = 3.38, p = 0.036, η_p² = 0.025) and P2 (F (2, 266) = 4.78, p = 0.009, η_p² = 0.035) at the Fpz (Fig. 3b). This revealed that, compared with the non-depressed individuals, the depressed individuals had averages of 21 ms and 27 ms longer latencies of N1 (p = 0.034) and P2 (p = 5.79 × 10^–3), respectively, in the positive news. In addition, the peak latency for N400 at the Cz showed a significant main effect of News (F (2, 266) = 4.42, p = 0.013, η_p² = 0.032) (Fig. 4b), which revealed that the latency for negative news was on average 33.3 ms shorter than that for positive news (p = 0.002).

Amplitudes

The significant main effect of Group was found in the amplitudes of N1 at the Pz (F (1, 133) = 6.45, p = 0.012, η_p² = 0.046): the non-depressed individuals showed more negative deflection than the depressed individuals did (Fig. 3c). A significant main effect of News on the P2 amplitude (F (2, 266) = 7.62, p = 0.001, η_p² = 0.054) was found at the Fpz (Fig. 4a). This revealed that the negative news elicited a larger P2 than the positive news (p = 2.95 × 10^–4).

Regarding amplitudes of N400, the significant interaction Group × News was found at Pz (F (2, 266) = 4.05, ε = 0.95, p = 0.020, η_p² = 0.030) (Fig. 3d). This revealed that the depressed individuals showed more negative deflection in the N400 amplitude than the non-depressed individuals did while listening to the negative news (p = 0.034). In addition, the negative news caused smaller negative deflection than the positive news, only in the non-depressed individuals (p = 0.017).

Classification of depressed or non-depressed individuals

We evaluated the performances to classify depressed or non-depressed individuals, based on our subjective evaluation and EEG measures. Between the EEG and subjective evaluation data, for which the main effect of Group or the interaction of Group × News was significant in the ANOVA, we prepared three feature sets. These included subjective evaluation-only features (six features: arousal and comprehensibility for each news condition), EEG-only features (12 features: peak latencies of N1 and P2 at Fpz and mean amplitudes of N1 and N400 at Pz, for each news condition), and the Combination features of EEG and subjective evaluation (18 features). For classification, linear SVM was trained and evaluated using the leave-one-out cross-validation (LOOCV). The EEG-only and the Combination features were selected using the recursive-feature elimination method, and six features were used. To determine whether the performances of the classifiers trained using each feature were by chance, the permutation test was performed by randomizing the labels 1000 times, and the empirical p-values were calculated for each feature⁴⁷.

The area under the receiver operating characteristic curves (AUCs) of the Subjective evaluation-only, EEG-only, and Combination features achieved 0.735, 0.730, and 0.832, respectively (Fig. 5a), indicating that all sets of the features can successfully discriminate depressed individuals. The permutation tests revealed that these performances of the classifiers were all significant (all ps = 0.001). The true positive rate (TPR), which is the rate of the depressed individuals, classified as the depressed individuals for the Subjective evaluation-only and EEG-only features was 0.719 (23/32 individuals) and 0.656 (21/32), respectively (Fig. 5b and c). The true negative rate (TNR), which is the rate of the non-depressed individuals, classified as the non-depressed individuals, was 0.592 (61/103) for the Subjective evaluation-only features and 0.660 (68/103) for the EEG-only feature. In the classification based on the Combination features, the TPR and TNR were 0.781 (25/32) and 0.718 (74/103), respectively (Fig. 5d). Figure 5e shows the number of depressed and non-depressed individuals per predicted label when using the Combination features.

The features’ contribution to the classification was analyzed by calculating the mean absolute values of the coefficients for the EEG-only features across all trained models (Fig. 5f), and the following six features were mainly selected: N1 amplitudes for negative, neutral, and positive news; P2 peak latencies for negative and positive news; and N400 amplitude for negative news. The feature with the highest coefficient values was the mean amplitude of N400 for negative news.

Discussion

We tested two hypotheses to clarify whether attention to affective information differed between depressed and non-depressed individuals, and how attentional bias influences language processing. The first hypothesis states that negativity bias is stronger for depressed individuals than for non-depressed individuals. So far, previous behavioral studies’ results are inconsistent, with some studies confirming a greater attentional engagement to negative information than to positive information^33,48 and difficulty in attentional disengagement from negative information^9,16,17, while in individuals with subclinical depression^16,17; others did not demonstrate such effects^24,25. In the present study, we focused on the N1 and P2 components to test the hypothesis. We found that news contents had a significant effect on the amplitude of P2, but not N1, with its amplitude being significantly larger for negative news than for positive news. Given that the P2 amplitude is enhanced by attention⁴⁹, which was confirmed especially when using affective stimulus^50,51,52 in visual studies, our result shows more attentional resources were allocated to negative information than to positive information, while participants were listening to the news. The P2 amplitude did not show a significant main effect of Group, which suggests that contrary to the hypothesis, negativity biases were observed regardless of the depressed state. Furthermore, from the results that the peak latency of N400, which reflects semantic processing^41,42,44, was shorter for negative news than for positive news in both groups, more attentional resources allocated to negative information could accelerate the subsequent semantic processing, regardless of the depressed state. To the best of our knowledge, this is a novel study on the effect of the attentional bias for negative information on the peak latency of semantic processing.

Other than negativity bias, delayed attention to positive information could be a characteristic of natural speech processing in depressed individuals: compared with the non-depressed individuals, the depressed individuals showed longer latencies in N1 and P2 to words within the positive news content. A previous ERP study using visual oddball paradigm reported a longer latency of early components (P1) to positive stimulus than to negative stimulus in individuals with major depression⁵³. Compared with the controls, behavioral data also showed delayed attention to positive stimulus in patients with major depression²² and in individuals with subclinical depression and anxiety¹⁷. Consistent with these results, the present result suggests that the attention to inputs is delayed in depressed individuals more than in non-depressed individuals, while they listen to positive information.

The second hypothesis states that owing to the affective interference, the depressed individuals require more load on language comprehension (semantic processing on words) while listening to negative news, than the non-depressed individuals. The result supports that the depressed individuals showed a negatively larger amplitude of N400 during the negative news than the non-depressed individuals. This result is consistent with a previous ERP study that used the emotional Stroop paradigm, indicating that negative words produced a negatively larger amplitude of N400-like component in participants with major depression and remitted depression than in healthy participants⁵⁴. These results support the affective interference theory³⁴: owing to the allocation of attentional resources to negative information, depressed individuals required a greater load on non-affective processing (i.e., semantic processing). Our series of results suggest that depressed individuals are characterized by a spillover of the effects of attentional bias for negative information to other types of processing.

To detect depressed individuals, the EEG-only features reached an AUC of 0.730, which is a metric in which the closer a number is to 1, the better the performance (0.5 = random classification). Combining the subjective evaluation features with the EEG features improved performance, and the highest performance of 0.832 in AUC was achieved. These classification performances highlight the effectiveness of EEG measures in successfully detecting depressed individuals. Previous studies on EEG-based depression patient detection have focused on resting-state data, using features such as power values and functional connectivity metrics across channels^55,56. In contrast, the current study focused on detecting depressed individuals based on their daily activities, and, for the first time, it demonstrates the feasibility of the EEG measured while listening to news.

The detection method used in this study was advantageous in that the cognitive function of the EEG features used was more clearly interpretable based on a huge number of previous ERP studies. The coefficient values of the EEG-only features indicated that the N400 amplitude to words within negative news contributed the most to the classification performance, suggesting that EEG features related to the semantic processing of natural speech are involved in the detection of depressed individuals. The contribution of EEG features in processing negative information is consistent with previous results that, the best classification performance for detecting individuals with mild depression in a non-clinical population was obtained when EEG features were used during the presentation of negative face stimuli⁵⁷. In addition, the N1 amplitude and P2 peak latency—modulated by attention to the stimuli³⁸—mainly for affective news contents (N1 amplitude: negative, neutral, and positive news, P2 peak latency: negative and positive news, cf. Fig. 5f) were also contributed to classification performance. These results suggest that individuals who increase the allocation of attentional resources and semantic processing load to affective speech stimuli are likely to be classified as having depression. In recent years, technologies to monitor cognitive or mental states such as cognitive load⁵⁸ and motivation⁵⁹ using brain activities under real-world conditions or situations close to the real have been widely studied. Consistent with this trend, we envision an application based on EEG that notifies individuals of their depressed state as predicted by their EEG so that it could lead to changing their behavior for using information, such as spending less time exposed to negative information on the Internet and social media. The promising classification performances in this study using natural speech as stimuli would lead to the development of such an application based on EEG and encourage individuals to change the way they use information according to their state of depression. In reality, the feasibility of such an application is high because a depressed state can be detected simply by listening to news auditory clips. In the future, it will be necessary to test whether providing feedback on the depressed state actually promotes such behavioral changes in individuals.

As a limitation, since the study ultimately aimed to develop an application based on EEG that provides feedback about individuals’ states of depression in daily life rather than to detect or diagnose clinical depression, we recruited people who had not been diagnosed with any mental disease in the study; depressed and non-depressed individuals were determined based solely on their BDI-II scores. Therefore, the depressed individuals in this study may have included those with clinical depression. It cannot be denied that there are differences in EEG responses between individuals with clinical depression and those who have not been diagnosed with depression, which may affect results obtained in this study. Future research will address this issue.

Conclusion

This study found that depressed individuals are characterized by the delayed attention to positive news and a higher semantic processing load for negative news. The classification model trained by these EEG features successfully classified depressed and non-depressed individuals, suggesting that EEG measures during daily life activities, such as listening to news items, provide a way to ascertain the depressed state.

Methods

Participants

Participants in the EEG experiment included 162 adult native Japanese speakers (88 women and 74 men, M_age = 34.2, SD_age = 10.1, range_age = 20–49). All participants had normal hearing and normal or corrected-to-normal vision. They had no history of neurological or psychiatric diseases. The data sets for participants whose EEG data contained large artifacts were excluded from the analysis; finally, data for 135 participants were used (69 women and 66 men, M_age = 33.2, SD_age = 10.1, range_age = 20–49). The participants’ handedness was estimated based on the laterality quotient (LQ) of the Edinburgh Handedness Inventory⁶⁰. Of the participants, 124 were classified as right-handed (M_LQ = 92.5, SD_LQ = 12.1), four as ambidextrous (M_LQ = − 18.2, SD_LQ = 24.6), and 7 as left-handed (M_LQ = − 83.0, SD_LQ = 18.0).

The participants’ depressed state was assessed using the BDI-II for Japanese³⁷ before data collection. The questionnaire comprised 21 questions regarding the participants’ depressed state in the past two weeks, including the day of the experiment. There was no significant gender difference in the BDI-II scores between women (M = 10.13, SD = 7.77) and men (M = 9.98, SD = 8.28) by an independent samples t-test (t (133) = 0.12, p = 0.908, Cohen’s d = 0.020). The BDI-II guidelines state that scoring 14 or higher is the threshold at which individuals are classified as having depression. As such, 32 participants were classified in the depressed individuals (M_age = 31.3, SD_age = 10.5, range_age = 20–49, M_LQ = 70.1, SD_LQ = 55.6) and 103 in the non-depressed individuals (M_age = 33.8, SD_age = 10.0, range_age = 20–49, M_LQ = 82.8, SD_LQ = 40.2) based on their BDI-II scores. The participants’ characteristics are summarized in Table 1.

The participants’ academic history was as follows: 1 junior high school graduate, 18 high school graduates, 19 junior college or vocational school students or graduates, 85 university students or graduates, and 12 graduate students, master’s, or doctoral degree holders. To confirm the relationship between academic history and the participants’ group (depressed or non-depressed), a chi-square test of independence was performed. No significant difference was confirmed (${\chi }^{2}$(4) = 5.54, p = 0.246, Cramer's V = 0.203).

This study was approved by the Ethics Committee for Human and Animal Research of the National Institute of Information and Communications Technology and was carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki). Written informed consent to participate in this study was obtained from all participants.

Materials

We prepared 30 news items from Japanese radio news programs (NHK Radio News; https://www.nhk.or.jp/radionews/) broadcasted by the Japan Broadcasting Corporation (NHK) in February 2019. Approximately one minute of the audio from the beginning of each news item was cut out for use as stimuli (ranged from 45 to 83 s).

To select positive, neutral, or negative news to be used in the EEG analysis from those 30 news items, we conducted a preliminary survey through subjective evaluations of 160 native Japanese speakers (82 women and 78 men, M_age = 35.9, SD_age = 8.4, range_age = 20–49) who did not participate in the EEG experiment. We selected each of five news items with (1) significant differences in affective valence among negative, neutral, and positive news items and (2) no significant differences in arousal among positive and negative news items. This was done as follows: participants read each news item and then subjectively evaluated the affective valence (1: negative–5: positive), arousal (1: low arousal–5: high arousal), interest (1: not interest at all–5: high interest), and comprehensibility (1: very difficult–5: very easy) on a 5-point scale. Based on this evaluation, we selected five news items with high (3.6 or more), moderate (3.0–3.4), and low (2.2 or less) affective valence, to be used as positive, neutral, and negative news conditions. The topics of the selected news are shown in Fig. 1.

To confirm whether the subjective evaluation in the preliminary survey differed across news conditions, the subjective evaluation ratings were subjected to one-way repeated measures ANOVA. As the subjective evaluation values ranged from 1 to 5 and were non-normally distributed, an aligned rank transformation of the subjective evaluation values was performed using the ARTool^61,62. To control for familywise error rates in multiple comparisons, p-values were adjusted based on the Bonferroni correction. The ANOVA results showed significant differences across news conditions in all items: the affective valence (F (1, 318) = 447.17, ε = 0.81, p = 7.87 × 10^–76, η_p² = 0.738), arousal (F (1, 318) = 31.34, ε = 0.93, p = 2.11 × 10^–12, η_p² = 0.165), interest (F (1, 318) = 25.95, p = 3.62 × 10^–11, η_p² = 0.140), and comprehensibility (F (1, 318) = 66.91, p = 5.60 × 10^–25, η_p² = 0.296). The affective valence of positive news was significantly higher than that of neutral and negative ones, and the affective valence of neutral news was significantly higher than that of negative news. Arousal due to both negative and positive news was significantly higher than that due to neutral news. The interest in negative news was significantly higher than that in neutral and positive ones, and positive news was significantly more than a neutral one. The comprehensibility of positive news was significantly higher than that of neutral and negative ones, and negative news was significantly more than that of a neutral one. All post-hoc test results for each item are summarized in Table 2. The differences in the ratings between the preliminary survey and EEG experiment are presented in Supplementary Table S2.

The EEG responses, time-locked to each word onset included in each news item, were analyzed. A Japanese native speaker identified the time of onset of each content word in each news item by listening to the audio and visually inspecting the sound waveform and spectrograms using WaveSurfer (https://sourceforge.net/projects/wavesurfer/), an open-source speech and acoustic analysis tool.

EEG experimental procedure

The participants sat in chairs, in front of a monitor display and numeric keypad. In the EEG experiment, they listened to 30 news items (15 selected news in total; the remaining 15 news were included in the stimuli set and were used for other research purposes) and evaluated them subjectively. In each trial, the participants started a trial by pressing a key “1” on the numeric keypad, and immediately after, they listened to an audio clip of one news item binaurally using earphones (RHA Technologies Ltd., United Kingdom). To suppress their eye movements, participants were instructed to gaze at a fixation point (+) presented at the center of the display while listening to the news. After each news item, participants were asked to subjectively evaluate the affective valence, arousal, interest, and comprehensibility of the news, based on a 5-point scale. The procedure for the subjective evaluation was the same as in the preliminary survey. Participants performed a total of six blocks (five trials/block). The data collection lasted approximately 1 h.

EEG and electrooculogram (EOG) signals were continuously measured throughout all blocks using an eight-channel wireless EEG device and measurement software (Polymate Mini AP108 and Mobile Acquisition Monitor 2.02, Miyuki Giken Co. Ltd., Japan). Because this study envisioned an application based on EEG that could be used in real-world environments, we focused on a few electrodes that were relatively easy to place rather than a high-density electrode system. Since N1/P2 and N400 to the auditory stimuli are dominantly observed in the fronto-central regions⁶³ and centro-parietal regions⁶⁴, respectively, active electrodes were placed on Fpz, Cz, and Pz locations, according to the International 10–10 system for EEG measurement. The Fpz location was chosen instead of Fz because it is easier to place the electrode on the forehead, where interference by hair can be avoided. Previous research has demonstrated that the Fpz electrode can successfully observe the N1 and P2 components to auditory stimuli⁶⁵. Two electrodes were placed on the lateral of the left outer canthus and above the left eye to measure horizontal and vertical EOGs, respectively. All signals were sampled at 500 Hz using reference and ground electrodes placed at the right and left earlobes, respectively.

EEG analysis

MATLAB (MathWorks Inc., USA) and EEGLAB toolbox⁶⁶ were used to preprocess EEG signals. A band-pass finite impulse response (FIR) filter between 1 and 50 Hz (3300th order) was applied to the signals, and it was resampled at 200 Hz. The transient, large-amplitude artifacts were removed by the artifact subspace reconstruction^67,68. Artifacts of eye blink and eye movements were excluded from the data using an independent component analysis. Finally, a further band-pass FIR filter of 1–8 Hz (1320th order) was applied to the signals to improve the signal-to-noise ratio.

To estimate the TRFs that describe the linear mapping between a stimulus and the preprocessed EEG data for each channel, we prepared a stimulus matrix at the same sampling rate as the preprocessed EEG data for each news item. The stimulus matrix comprised time-aligned impulses with a value of 1 at the content word onset time points and 0 at other time points. We used the mTRF toolbox in MATLAB⁴⁵ for the estimation. The continuous EEG response ${r}_{ch}\left(t\right)$ at a time point t was assumed to comprise convolutions of the stimulus vector s(t) and TRF weights ${w}_{ch}$($\tau$) at ch-th channel:

$$r_{ch} \left( t \right) = \mathop \sum \limits_{\tau } w_{ch} \left( \tau \right)s\left( {t - \tau } \right) + \varepsilon_{ch} \left( t \right),$$

(1)

where ε_ch(t) is the residual response in the ch-th channel. Ridge regression, with a regularization parameter λ, was used to estimate the TRF weights over a range of time lag τ from − 100 to 800 ms, relative to each stimulus onset. Using the mTRFcrossval function⁴⁵, we estimated the optimal regularization parameters for each participant based on the LOOCV, in which one trial is evaluated as test data and the remaining trials as training data, and the evaluation is repeated until all trials are used as test data. In this procedure, the following steps were applied to each participant’s channel data: (1) the TRF weights were estimated for each single-trial data, (2) the single-trial weights were averaged over the training data, and (3) the mean squared errors (MSEs) between the actual responses and the responses predicted using the averaged weights were computed using the test data. In the range [2¹–2²¹], for each participant, we determined the optimal regularization parameter λ that resulted in the smallest MSEs, averaged over the test data and the channel. Finally, the regularization parameter, λ = 2¹³, which was the most frequent value in the distribution of regularization parameters estimated by each participant (N = 42), was adopted for all participants. Finally, the single-trial TRF weights for each participant, estimated using the regularization parameter, were averaged by news condition per channel.

For each channel, the mean amplitude per component of a single participant’s TRF weight for each news condition was calculated using time windows of [100, 160 ms], [170, 230 ms], and [300, 600 ms] for N1, P2, and N400, respectively. These time windows were determined through the visual inspection of grand-averaged TRF weights. The peak latency of each component was also detected per channel in each news condition. To detect peak latencies, we used a dynamic time warping (DTW) algorithm, which is a technique used to non-linearly map two temporal sequences that vary in time or speed⁶⁹. DTW has been used to automatically detect peak latencies of ERP components^70,71. First, we calculated the grand-averaged TRF waveforms for all participants and each news condition for each channel, from which we detected the peak latencies for the three components. Thereafter, for each channel, the TRF for each participant and the grand-averaged TRF were mapped onto a common time axis using DTW, as such, the time point corresponding to the peak latency obtained from the grand-averaged TRF was automatically identified.

Statistical analysis

To analyze subjective evaluation, a two-way mixed-design ANOVA was employed to analyze whether subjective evaluation was affected by News (within-subjects factor: negative, neutral, and positive) and Group (between-subjects factor: depressed individuals and non-depressed individuals). For analysis using ANOVA, each subjective evaluation was performed on an aligned rank transformation as well as the preliminary survey. To investigate the effect of Group and News on the amplitude or peak latency of each component corresponding to N1, P2, and N400, a two-way mixed-design ANOVA was conducted. When the main effect of News was statistically significant, multiple comparisons of paired t-tests across news conditions were conducted. When significant interaction of Group × News was obtained, multiple comparisons of paired t-tests across news conditions were conducted for each group. In addition, to investigate the simple main effects of Group, unpaired t-tests between the groups were administered per the new condition.

In all statistical test procedures, the significance level was set at 0.05. If Mauchly’s test showed that homogeneity of variance was violated, the degree of freedom was adjusted using the Greenhouse–Geisser procedure. When the degree of freedom was adjusted using the procedure, the original degree of freedom and ε were reported. We report partial eta squared as an effect size. For all multiple comparisons, p-values were adjusted using the Bonferroni procedure.

Classification with a linear SVM

To investigate whether EEG measurements play a role in the detection of depressed individuals, we performed a binary classification of depressed and non-depressed individuals based on the EEG features and/or subjective evaluation of the news. As features for classification, we used explanatory variables showing the significant main effect of Group or interaction of Group × News in the ANOVA: 12 EEG features (peak latencies of N1 and P2 at Fpz and mean amplitudes of N1 and N400 at Pz, for each news item) and 6 subjective evaluation features (arousal and comprehensibility for each news item). To compare the results, based on the features used, three sets of features were prepared: 12 features of only EEG, 6 features of only subjective evaluation, and 18 features combining EEG and subjective evaluation (hereinafter referred to as the EEG-only features, Subjective evaluation-only features, and Combination features, respectively). The feature vectors were standardized per feature by subtracting the mean value and then dividing the difference by the SD. The mean values and SD were calculated using the training data. For classification, we used a linear SVM, which is widely used in the Brain-Computer Interface context because of the robustness to outliers, due to the regularization and applicability to high-dimensional data sets⁷², with the squared hinge loss and the L2 penalty.

The performance of the linear SVM was evaluated using the LOOCV, whereby one participant’s data was reserved for a test and the remaining data were used for model training. In the LOOCV, the procedure was repeated so that every participant’s data was evaluated as test data once. For the EEG-only features and the Combination features, the feature selection procedure involved the recursive-feature elimination method. In each LOOCV procedure, six features were selected based on the coefficients of the linear SVM trained on the training data, and the selected features were used for model training and evaluation. A cost parameter C, which is a hyperparameter that determines the degree of tolerance for classification errors in the training data, was optimized using the grid-search. The best performing cost parameter in the stratified threefold cross-validation of the training data was selected from [C = 10^–3, 10^–2, 10^–1, 10⁰, 10¹, 10², 10³].

The number of data samples per class was imbalanced (non-depressed individuals: N = 103, depressed individuals: N = 32). Therefore, the costs of the classification errors were adjusted per class by multiplying the cost parameter C by the weights, based on the number of training samples in the class (i.e., the total number of samples/(the number of classes × the number of samples in the class)). Another widely used approach to handle imbalanced data is resampling, which artificially balances the class distribution of the training data by over- (i.e., increasing the number of samples belonging to the minority class) and/or under-sampling (i.e., removing samples belonging to the majority class). The classification performance was evaluated using AUC and those using the approach are provided in Supplementary Fig. S1. The permutation test was conducted to determine whether the classification performances using each future were by chance. In the test, a null distribution was constructed under the null hypothesis that there was no dependence between the features and the labels on each other by randomizing the labels, an then calculating an empirical p-value of the obtained AUC against that null distribution, which is the proportion of the randomized labels that performs as well or better than the original label⁴⁷. The total number of randomizations was set to 1000.

To clarify which features contributed to the classification performances, among the features that showed differences in EEG responses to affective news depending on individuals’ depressed state, we computed the absolute values of the coefficients assigned to each feature of the models trained by the EEG-only features. As LOOCV repeats the training and evaluation of the same number of data samples, the values of the coefficients were averaged over all models trained in LOOCV. If a feature was not selected in the feature selection procedure applied to the EEG-only features, its coefficient was treated as having zero value in this analysis. The scikit-learn library for Python was used for the classification procedures⁷³.

Data availability

The data set presented in this research is partially available upon request to the corresponding author, limited to the data that the participant has agreed to make publicly available.

References

Johnston, W. M. & Davey, G. C. The psychological impact of negative TV news bulletins: The catastrophizing of personal worries. Br. J. Psychol. 88, 85–91 (1997).
Article PubMed Google Scholar
Silver, R. C., Holman, E. A. & Garfin, D. R. Coping with cascading collective traumas in the United States. Nat. Hum. Behav. 5, 4–6. https://doi.org/10.1038/s41562-020-00981-x (2021).
Article PubMed PubMed Central Google Scholar
Zhao, N. & Zhou, G. Social media use and mental health during the COVID-19 pandemic: Moderator role of disaster stressor and mediator role of negative affect. Appl. Psychol. Health Well-Being 12, 1019–1038 (2020).
Article PubMed PubMed Central Google Scholar
Kecojevic, A., Basch, C. H., Sullivan, M. & Davi, N. K. The impact of the COVID-19 epidemic on mental health of undergraduate students in New Jersey, cross-sectional study. PLoS ONE 15, e0239696. https://doi.org/10.1371/journal.pone.0239696 (2020).
Article CAS PubMed PubMed Central Google Scholar
Ito, T. A., Larsen, J. T., Smith, N. K. & Cacioppo, J. T. Negative information weighs more heavily on the brain: the negativity bias in evaluative categorizations. J. Pers. Soc. Psychol. 75, 887–900 (1998).
Article CAS PubMed Google Scholar
Rozin, P. & Royzman, E. B. Negativity bias, negativity dominance, and contagion. Pers. Soc. Psychol. Rev. 5, 296–320 (2001).
Article Google Scholar
Taylor, S. E. Asymmetrical effects of positive and negative events: the mobilization-minimization hypothesis. Psychol. Bull. 110, 67–85 (1991).
Article CAS PubMed Google Scholar
Fossati, P. Is major depression a cognitive disorder?. Rev. Neurol. 174, 212–215 (2018).
Article CAS PubMed Google Scholar
LeMoult, J. & Gotlib, I. H. Depression: A cognitive perspective. Clin. Psychol. Rev. 69, 51–66 (2019).
Article PubMed Google Scholar
Roiser, J. P., Elliott, R. & Sahakian, B. J. Cognitive mechanisms of treatment in depression. Neuropsychopharmacology 37, 117–136 (2012).
Article CAS PubMed Google Scholar
Nakamura, A. et al. The cerebellum as a moderator of negative bias of facial expression processing in depressive patients. J. Affect. Disord. Rep. 7, 100295 (2022).
Article Google Scholar
Beck, A. T. Cognitive Therapy and the Emotional Disorders (Penguin, 1979).
Google Scholar
Gotlib, I. H. & Joormann, J. Cognition and depression: Current status and future directions. Annu. Rev. Clin. Psychol. 6, 285–312 (2010).
Article PubMed PubMed Central Google Scholar
Teasdale, J. D. Negative thinking in depression: Cause, effect, or reciprocal relationship?. Adv. Behav. Res. Therapy 5, 3–25 (1983).
Article Google Scholar
Mennen, A. C., Norman, K. A. & Turk-Browne, N. B. Attentional bias in depression: Understanding mechanisms to improve training and treatment. Curr. Opin. Psychol. 29, 266–273 (2019).
Article PubMed PubMed Central Google Scholar
Baert, S., De Raedt, R. & Koster, E. H. W. Depression-related attentional bias: The influence of symptom severity and symptom specificity. Cogn. Emot. 24, 1044–1052 (2010).
Article Google Scholar
Koster, E. H. W. et al. Mood-congruent attentional bias in dysphoria: maintained attention to and impaired disengagement from negative information. Emotion 5, 446–455 (2005).
Article PubMed Google Scholar
Joormann, J. & Gotlib, I. H. Selective attention to emotional faces following recovery from depression. J. Abnorm. Psychol. 116, 80–85 (2007).
Article PubMed Google Scholar
Joormann, J., Talbot, L. & Gotlib, I. H. Biased processing of emotional information in girls at risk for depression. J. Abnorm. Psychol. 116, 135–143 (2007).
Article PubMed Google Scholar
Ingram, R. E. Toward an information-processing analysis of depression. Cognit. Ther. Res. 8, 443–477 (1984).
Article Google Scholar
Bower, G. H. Mood and memory. Am. Psychol. 36, 129–148 (1981).
Article CAS PubMed Google Scholar
Mingtian, Z. et al. Do the early attentional components of ERPs reflect attentional bias in depression? It depends on the stimulus presentation time. Clin. Neurophysiol. 122, 1371–1381 (2011).
Article PubMed Google Scholar
Bodenschatz, C. M. et al. Implicit negative affect predicts attention to sad faces beyond self-reported depressive symptoms in healthy individuals: An eye-tracking study. Psychiatry Res. 265, 48–54 (2018).
Article PubMed Google Scholar
Cheng, P. et al. Evidence against mood-congruent attentional bias in Major Depressive Disorder. Psychiatry Res. 230, 496–505 (2015).
Article PubMed Google Scholar
Neshat-Doost, H. T. et al. Lack of attentional bias for emotional information in clinically depressed children and adolescents on the dot probe task. J. Child Psychol. Psychiatry 41, 363–368 (2000).
Article CAS PubMed Google Scholar
McCabe, S. B. & Gotlib, I. H. Selective attention and clinical depression: Performance on a deployment-of-attention task. J. Abnorm. Psychol. 104, 241–245 (1995).
Article CAS PubMed Google Scholar
Kaiser, R. H. et al. Distracted and down: neural mechanisms of affective interference in subclinical depression. Soc. Cogn. Affect. Neurosci. 10, 654–663 (2015).
Article PubMed Google Scholar
Sass, S. M. et al. Electrophysiological evidence of the time course of attentional bias in non-patients reporting symptoms of depression with and without co-occurring anxiety. Front. Psychol. 5, 301. https://doi.org/10.3389/fpsyg.2014.00301 (2014).
Article PubMed PubMed Central Google Scholar
Mitterschiffthaler, M. et al. Neural basis of the emotional Stroop interference effect in major depression. Psychol. Med. 38, 247–256 (2008).
Article CAS PubMed Google Scholar
Joyal, M. et al. Characterizing emotional Stroop interference in posttraumatic stress disorder, major depression and anxiety disorders: A systematic review and meta-analysis. PLoS ONE 14, e0214998. https://doi.org/10.1371/journal.pone.0214998 (2019).
Article CAS PubMed PubMed Central Google Scholar
Mogg, K., Bradley, B. P. & Williams, R. Attentional bias in anxiety and depression: The role of awareness. Br. J. Clin. Psychol. 34, 17–36 (1995).
Article CAS PubMed Google Scholar
Li, X. et al. Attentional bias in MDD: ERP components analysis and classification using a dot-probe task. Comput. Methods Programs Biomed. 164, 169–179 (2018).
Article PubMed Google Scholar
Trapp, W. et al. Attentional biases in patients suffering from unipolar depression: Results of a dot probe task investigation. Psychiatry Res. 261, 325–331 (2018).
Article PubMed Google Scholar
Siegle, G. J., Ingram, R. E. & Matt, G. E. Affective interference: An explanation for negative attention biases in dysphoria?. Cogn. Ther. Res. 26, 73–87 (2002).
Article Google Scholar
Shimony, O. et al. The association between implicit and explicit affective inhibitory control, rumination and depressive symptoms. Sci. Rep. 11, 11490. https://doi.org/10.1038/s41598-021-90875-3 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Dell’Acqua, C. et al. Depressive symptoms and cognitive control: The role of affective interference. Cogn. Emot. https://doi.org/10.1080/02699931.2022.2128065 (2022).
Article PubMed Google Scholar
Kojima, M. et al. Cross-cultural validation of the Beck Depression Inventory-II in Japan. Psychiatry Res. 110, 291–299 (2002).
Article PubMed Google Scholar
Hillyard, S. A., Hink, R. F., Schwent, V. L. & Picton, T. W. Electrical signs of selective attention in the human brain. Science 182, 177–180 (1973).
Article ADS CAS PubMed Google Scholar
Luck, S. J. & Kappenman, E. S. ERP components and selective attention. In (eds Kappenman, E. S. & Luck, S. J.) The Oxford Handbook of Event-Related Potential Components (Oxford University Press, 2012).
Marzecová, A. et al. Interrelation of attention and prediction in visual processing: Effects of task-relevance and stimulus probability. Biol. Psychol. 125, 76–90 (2017).
Article PubMed Google Scholar
Friederici, A. D. Towards a neural basis of auditory sentence processing. Trends Cogn. Sci. 6, 78–84 (2002).
Article PubMed Google Scholar
Kutas, M. & Hillyard, S. A. Reading senseless sentences: Brain potentials reflect semantic incongruity. Science 207, 203–205 (1980).
Article ADS CAS PubMed Google Scholar
Rabovsky, M., Hansen, S. S. & McClelland, J. L. Modelling the N400 brain potential as change in a probabilistic representation of meaning. Nat. Hum. Behav. 2, 693–705. https://doi.org/10.1038/s41562-018-0406-4 (2018).
Article PubMed Google Scholar
Kutas, M. & Federmeier, K. D. Electrophysiology reveals semantic memory use in language comprehension. Trends Cogn. Sci. 4, 463–470 (2000).
Article CAS PubMed Google Scholar
Crosse, M. J., Di Liberto, G. M., Bednar, A. & Lalor, E. C. The multivariate temporal response function (mTRF) toolbox: A MATLAB toolbox for relating neural signals to continuous stimuli. Front. Hum. Neurosci. 10, 604. https://doi.org/10.3389/fnhum.2016.00604 (2016).
Article PubMed PubMed Central Google Scholar
Ihara, A. S. et al. Prediction of second language proficiency based on electroencephalographic signals measured while listening to natural speech. Front. Hum. Neurosci. 15, 665809. https://doi.org/10.3389/fnhum.2021.665809 (2021).
Article PubMed PubMed Central Google Scholar
Ojala, M. & Garriga, G. C. Permutation tests for studying classifier performance. J. Mach. Learn. Res. 11, 1833–1864. https://doi.org/10.5555/1756006.1859913 (2010).
Article MathSciNet MATH Google Scholar
Leyman, L., De Raedt, R., Schacht, R. & Koster, E. H. Attentional biases for angry faces in unipolar depression. Psychol. Med. 37, 393–402 (2007).
Article PubMed Google Scholar
Luck, S. J. & Hillyard, S. A. Electrophysiological correlates of feature analysis during visual search. Psychophysiology 31, 291–308 (1994).
Article CAS PubMed Google Scholar
Carretié, L. et al. Automatic attention to emotional stimuli: Neural correlates. Hum. Brain Mapp. 22, 290–299 (2004).
Article PubMed PubMed Central Google Scholar
Carretié, L., Mercado, F., Tapia, M. & Hinojosa, J. A. Emotion, attention, and the ‘negativity bias’, studied through event-related potentials. Int. J. Psychophysiol. 41, 75–85 (2001).
Article PubMed Google Scholar
Carretié, L., Ruiz-Padial, E., López-Martín, S. & Albert, J. Decomposing unpleasantness: Differential exogenous attention to disgusting and fearful stimuli. Biol. Psychol. 86, 247–253 (2011).
Article PubMed Google Scholar
Yang, W. et al. Time course of affective processing bias in major depression: An ERP study. Neurosci. Lett. 487, 372–377 (2011).
Article CAS PubMed Google Scholar
Dai, Q. & Feng, Z. Deficient interference inhibition for negative stimuli in depression: An event-related potential study. Clin. Neurophysiol. 122, 52–61 (2011).
Article PubMed Google Scholar
Duan, L. et al. Machine learning approaches for MDD detection and emotion decoding using EEG signals. Front. Hum. Neurosci. 14, 284. https://doi.org/10.3389/fnhum.2020.00284 (2020).
Article PubMed PubMed Central Google Scholar
Khodayari-Rostamabad, A. et al. Diagnosis of psychiatric disorders using EEG data and employing a statistical decision model. Annu. Int. Conf. IEEE Eng. Med. Biol. 2010, 4006–4009 (2010).
Article Google Scholar
Li, X. et al. A deep learning approach for mild depression recognition based on functional connectivity using electroencephalography. Front. Neurosci. 14, 192. https://doi.org/10.3389/fnins.2020.00192 (2020).
Article PubMed PubMed Central Google Scholar
Dehais, F. et al. Monitoring pilot’s cognitive fatigue with engagement features in simulated and actual flight conditions using an hybrid fNIRS-EEG passive BCI. IEEE Trans. Syst. Man Cybern. Syst. 2018, 544–549 (2018).
Google Scholar
Watanabe, H. & Naruse, Y. P300 as a neural indicator for setting levels of goal scores in educational gamification applications from the perspective of intrinsic motivation: An ERP Study. Front. Neuroergonomics 3, 948080. https://doi.org/10.3389/fnrgo.2022.948080 (2022).
Article Google Scholar
Oldfield, R. C. The assessment and analysis of handedness: the Edinburgh inventory. Neuropsychologia 9, 97–113 (1971).
Article CAS PubMed Google Scholar
Elkin, L. A., Kay, M., Higgins, J. J. & Wobbrock, J. O. An aligned rank transform procedure for multifactor contrast tests. In Proc. 34th Annu. ACM Symp. User Interface Softw. Technol. 754–768 (2021).
Wobbrock, J. O., Findlater, L., Gergle, D. & Higgins, J. J. The aligned rank transform for nonparametric factorial analyses using only anova procedures. In Proc. of the SIGCHI Conference on Human Factors in Computing Systems 143–146 (2011).
Picton, T. W., Hillyard, S. A., Krausz, H. I. & Galambos, R. Human auditory evoked potentials. I: Evaluation of components. Electroencephalogr. Clin. Neurophysiol. 36, 179–190 (1974).
Article CAS PubMed Google Scholar
McCallum, W., Farmer, S. & Pocock, P. The effects of physical and semantic incongruites on auditory event-related potentials. Electroencephalogr. Clin. Neurophysiol. 59, 477–488 (1984).
Article CAS PubMed Google Scholar
Khatun, S., Morshed, B. I. & Bidelman, G. M. A single-channel EEG-based approach to detect mild cognitive impairment via speech-evoked brain responses. IEEE Trans. Neural Syst. Rehabil. Eng. 27, 1063–1070 (2019).
Article PubMed PubMed Central Google Scholar
Delorme, A. & Makeig, S. EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 134, 9–21 (2004).
Article PubMed Google Scholar
Blum, S., Jacobsen, N. S., Bleichner, M. G. & Debener, S. A Riemannian modification of artifact subspace reconstruction for EEG artifact handling. Front. Hum. Neurosci. 13, 141. https://doi.org/10.3389/fnhum.2019.00141 (2019).
Article PubMed PubMed Central Google Scholar
Mullen, T. R. et al. Real-time neuroimaging and cognitive monitoring using wearable dry EEG. IEEE Trans. Biomed. Eng. 62, 2553–2567 (2015).
Article PubMed PubMed Central Google Scholar
Sakoe, H. & Chiba, S. Dynamic programming algorithm optimization for spoken word recognition. IEEE Trans. Acoust. 26, 43–49 (1978).
Article MATH Google Scholar
Assecondi, S. et al. Automated identification of ERP peaks through dynamic time warping: An application to developmental dyslexia. Clin. Neurophysiol. 120, 1819–1827 (2009).
Article PubMed Google Scholar
Zoumpoulaki, A. et al. Latency as a region contrast: Measuring ERP latency differences with dynamic time warping. Psychophysiology 52, 1559–1576 (2015).
Article CAS PubMed Google Scholar
Lotte, F. et al. A review of classification algorithms for EEG-based brain–computer interfaces. J. Neural Eng. 4, R1–R13. https://doi.org/10.1088/1741-2560/4/2/r01 (2007).
Article CAS PubMed Google Scholar
Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830. https://doi.org/10.4236/ib.2022.142006 (2011).
Article MathSciNet MATH Google Scholar

Download references

Acknowledgements

We wish to thank Ms. Nagisa Kamioka for her assistance with data analysis. We also wish to thank Ms. Yuka Ishizawa, Ms. Akiko Yoshida, and Ms. Miyoko Ikeda for their assistances with data measurement.

Author information

Kohei Fuseda
Present address: Bunkyo Gakuin University, Fujimino, Saitama, Japan
Atsushi Matsumoto
Present address: Kansai University of Welfare Sciences, Kashiwara, Osaka, Japan
These authors contributed equally: Kohei Fuseda and Hiroki Watanabe.

Authors and Affiliations

Center for Information and Neural Networks, Advanced ICT Research Institute, National Institute of Information and Communications Technology, and Osaka University, 588-2 Iwaoka, Iwaoka-cho, Nishi-ku, Kobe, Japan
Kohei Fuseda, Hiroki Watanabe, Atsushi Matsumoto, Junpei Saito, Yasushi Naruse & Aya S. Ihara

Authors

Kohei Fuseda
View author publications
You can also search for this author in PubMed Google Scholar
Hiroki Watanabe
View author publications
You can also search for this author in PubMed Google Scholar
Atsushi Matsumoto
View author publications
You can also search for this author in PubMed Google Scholar
Junpei Saito
View author publications
You can also search for this author in PubMed Google Scholar
Yasushi Naruse
View author publications
You can also search for this author in PubMed Google Scholar
Aya S. Ihara
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

A.S.I. and Y.N. designed the study. A.S.I. prepared the experiment and collected data. A.M. and J.S. conducted a preprocessing of the subjective evaluation and EEG data. K.F. estimated TRF. and performed the statistical analysis. H.W. performed the classification analysis. K.F. and H.W. wrote the first draft of the manuscript. All authors revised the draft and approved the final version.

Corresponding authors

Correspondence to Hiroki Watanabe or Aya S. Ihara.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary Information.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Fuseda, K., Watanabe, H., Matsumoto, A. et al. Impact of depressed state on attention and language processing during news broadcasts: EEG analysis and machine learning approach. Sci Rep 12, 20492 (2022). https://doi.org/10.1038/s41598-022-24319-x

Download citation

Received: 21 July 2022
Accepted: 14 November 2022
Published: 28 November 2022
DOI: https://doi.org/10.1038/s41598-022-24319-x

This article is cited by

ChineseEEG: A Chinese Linguistic Corpora EEG Dataset for Semantic Alignment and Neural Decoding
- Xinyu Mou
- Cuilin He
- Haiyan Wu
Scientific Data (2024)
Delineating emotional differences between depressed and non-depressed individuals using a novel multimodal framework
- Rupali Gill
- Jaiteg Singh
- Durgesh Srivastava
Multimedia Tools and Applications (2024)
The power of humorous audio: exploring emotion regulation in traffic congestion through EEG-based study
- Lekai Zhang
- Yingfan Wang
- Fo Hu
EURASIP Journal on Audio, Speech, and Music Processing (2023)