Estimating Brain Health from Facial Expressions: An Exploratory Study

Abstract

In recent years, increasing attention has been paid to the effects of reading and art on the human brain. However, how these activities are associated with brain structure in healthy middle-aged adults remains unclear, partly because structural brain measures are not easily accessible outside MRI-based settings. In Study I, we analyzed the correspondence between gray matter volume (GMV) derived from brain MRI and facial-expression data obtained while participants imitated four facial expressions (happiness, anger, sadness, and surprise). Based on these data, we developed an exploratory algorithm and a digital application to estimate brain-health-related indices from facial expressions. In Study II, we examined correlations between the estimated brain-health-related indices and questionnaire-based measures of creative behavior and reading habits in 113 self-reported healthy middle-aged adults. The results showed that estimated indices related to the default mode network (DMN) and central executive network (CEN) were positively associated with creative behavior and reading habits, respectively. To our knowledge, this is among the first studies to explore whether facial-expression-based estimates of brain-health-related indices may be used to examine associations between everyday intellectual activities and brain-health-related characteristics. However, the findings should be interpreted cautiously because the estimation model was evaluated within a limited sample, included repeated observations, and has not yet been externally validated.

1. Study I: Development of Digital Technology to Estimate Standardized Brain Health from Facial Expressions

1.1. Introduction

In recent years, in developed countries, the reduction of medical and nursing care costs due to the decline in the working-age population and the aging of society has become a social issue, and extending healthy life expectancy is considered to be an important factor. In addition, lifestyle habits, aging, and mental and neurological diseases account for a large portion of medical expenses, and there is a demand for the establishment of health care services that improve lifestyle habits. In the healthcare field, inferring a patient’s health through face-to-face meetings is common, but it is not necessarily scientific.

A study by Cannavacciuolo et al. [1] compared facial emotional expressivity in patients with Parkinson’s disease, Alzheimer’s disease, and healthy controls, and found that patients with Parkin-son’s disease and Alzheimer’s disease shared a common tendency to have lower facial emotional expressivity compared to controls [1]. These symptoms are associated with atrophy of cortical and subcortical areas in the limbic system [2]. Inability to regulate emotional facial expressions can have a negative impact on an individual’s socio-emotional environment and has been linked to depression, schizophrenia [3], flattened affect [4], and reduced levels of trait empathy in healthy adults [5]. Alterations in socioemotional attitudes, including deficits in spontaneous emotion expression, empathy, emotion reading, and expression of blunt emotion, are also found in patients with behavioral variant frontotemporal dementia (bvFTD), right temporal frontotemporal dementia (rtvFTD), semantic variant primary progressive aphasia (svPPA), and Alzheimer’s disease (AD) [6,7,8,9,10,11]. Impaired intentional emotion expression was predicted by general expressiveness and emotion reading deficits across neurodegenerative patients, and imitation abilities were significantly impaired in bvFTD and rtvFTD patients compared to controls [12].

Blunted facial affect, or reduced facial expression of emotion, is a transdiagnostic component of serious mental illness [13,14,15,16,17]. Blunted facial affect is a key component of blunted affect, which involves reduced facial expression through many channels, including facial, vocal, and gestural expressions. Blunted affect is a symptom of schizophrenia and is also commonly seen in mood disorders (e.g., psycho-motor retardation in depression), autism spectrum, and neurodegenerative disorders [17,18]. Blunted facial affect is associated with poorer social and occupational functioning [19,20,21] and increased risk of suicide [22]. Therefore, in recent years, research has been conducted using AI technology to identify facial expression features characteristic of people with severe mental illness [23], identify depression using facial expressions and images [24,25], and predict neural networks from facial expressions [26].

On the other hand, there is a certain accumulation of research on predicting brain age from GM images [27]. The difference between brain age and chronological age is quantitatively associated with several age-related risk factors and general physical health, such as reduced grip strength, reduced lung function, history of stroke, increased frequency of alcohol consumption, and increased risk of death [28], as well as declines in cognitive functions such as fluid intelligence, processing speed, semantic language fluency, visual attention, and cognitive flexibility [29,30,31]. Brain age prediction models tend to perform better using GMV than white matter volume [32]. Therefore, it is possible to predict the relationship between facial expressions, which have been shown to be related to dementia and mental illness, and brain age, as expressed by GMV. However, no algorithms have been developed to predict GMV from facial expressions. Considering the cost and time required to measure brain state using MRI, it is difficult to measure frequently, and the development of technology to estimate GMV from facial expressions could greatly contribute to people’s health and welfare.

To date, the authors have been developing evaluation technologies to visualize and quantify the state of the human mind and body, including human senses, emotions, and physical strain, with the mission of contributing to society by creating better products and services through the use of analytical technology to add new value to our customers’ products and services. They also possess original facial expression analysis technology [33]. On the other hand, focusing on the “brain” that controls the human mind and body, Yamakawa et al. [34] have proposed the Brain Healthcare Quotient (BHQ) as an index to measure brain health [35]. The BHQ was internationally standardized in 2018 [36] and is easy to use even for non-experts, so it is expected to be used in a variety of solutions, such as allowing users to understand their own brain health and take action to improve their lifestyle habits. However, the BHQ is an index calculated from Magnetic Resonance Imaging (MRI) images of the brain, and requires MRI imaging at a hospital or other facility, so users cannot easily measure it.

To solve this problem, we used our facial expression analysis technology to analyze the correspondence be-tween BHQ and facial expression data, and examined the relationship between the two. Using the results, we built an algorithm to estimate BHQ from facial expression data, and developed a digital application that allows users to easily and quickly learn about their standardized brain health.

1.2. Materials and Methods

Data were acquired to analyze the relationship between the BHQ calculated from brain MRI images and facial expression measurement data.

1.2.1. Participants

A total of 58 healthy male and female subjects (47 males and 11 females) aged 25 to 63 years were randomly selected from employees of Panasonic Holdings Corporation. Data acquisition was performed after obtaining informed consent from all subjects. Participants were employees who were able to perform their routine duties on the day of measurement and reported no neurological or psychiatric disorders in a screening questionnaire. Therefore, the sample is described as self-reported healthy adults. According to self-report, none of the recruited participants had a record of neurological, psychiatric, or other medical conditions that could affect the central nervous system. This study was approved by the Ethics Committee of the Institute of Science Tokyo (approval numbers 2023137), and all methods were carried out in accordance with the relevant guidelines, regulations, and principles of the Declaration of Helsinki. All participants provided written informed consent before participation, and anonymity was maintained.

1.2.2. BHQ Indicators

The BHQ mentioned above was used as an index of brain health. There are two types of BHQ calculated from brain MRI images: GM-BHQ (Grey Matter-BHQ), which indexes the amount of gray matter where many brain neurons are present, and FA-BHQ (Fractional Anisotropy-BHQ), which indexes the quality of white matter where many brain nerve fibers are present [37]. GM-BHQ represents the average of 116 regions in the AAL atlas [38]. We also calculated the average values for regions associated with the Default Mode Network (DMN), Central Executive Network (CEN), and Salience Network (SN). The AAL atlas regions used for each network are shown in

Table 1. Network and region name.

Network	AAL Code	Region Name
DMN	AAL023	Superior Medial Frontal Gyrus (Left)
	AAL024	Superior Medial Frontal Gyrus (Right)
	AAL035	Posterior Cingulate Gyrus (Left)
	AAL036	Posterior Cingulate Gyrus (Right)
	AAL061	Inferior Parietal Lobule (Left)
	AAL062	Inferior Parietal Lobule (Right)
	AAL067	Precuneus (Left)
	AAL068	Precuneus (Right)
CEN	AAL003	Superior Frontal Gyrus (Left)
	AAL004	Superior Frontal Gyrus (Right)
	AAL059	Superior Parietal Lobule (Left)
	AAL060	Superior Parietal Lobule (Right)
SN	AAL029	Insula (Left)
	AAL030	Insula (Right)
	AAL031	Anterior Cingulate Gyrus (Left)
	AAL032	Anterior Cingulate Gyrus (Right)

.

1.2.3. MRI Data Acquisition

A 3 Tesla MRI scanner (MAGNETOM Prisma, Siemens, Munich, Germany) with a 32-channel head array coil, three-dimensional (3D) T1-weighted magnetization-prepared rapid acquisition gradient echo pulse sequence, and spin-echo echo-planar imaging (SE-EPI) with generalized autocalibrated partially parallel acquisition (GRAPPA) were used for MRI data collection and structural imaging. The following parameters were used: repetition time (TR) 1900 ms, echo time (TE) 2.52 ms, inversion time (TI) 900 ms, flip angle 9°, matrix size 256 × 256, field of view (FOV) 256 mm, and slice thickness 1 mm.

1.2.4. MRI Data Analysis

T1-weighted images were segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) using Statistical Parametric Mapping 12 (SPM12; Wellcome Trust Center for Neuroimaging, London, UK) running on MATLAB R2022b (MathWorks Inc., Sherborne, MA, USA). Segmented GM images were then spatially normal-ized using diffeomorphic anatomical registration using the exponential lie algebra (DARTEL) algorithm [39], which included incorporating a modulation step in-to the preprocessing model and smoothing images using a Gaussian kernel with a full width at half maximum (FWHM) of 8 mm. The smoothed GM images were then converted to proportional GM images by dividing by the intracranial volume (ICV) and used to create mean and standard deviation (SD) images. By averaging these in-formation and the local GM quotients using the automatic anatomical labeling (AAL) atlas [40], the GM-BHQ was created with a mean of 100 and a SD of 15 points.

Previous research has reported that BHQ tends to decline with age and can be an indicator of the youthfulness of the brain. Research into GM-BHQ’s relationship with lifestyle has also been conducted; for example, resting and doing housework on weekdays has a positive effect on GM-BHQ, while taking too much time off on weekends tends to be bad for GM-BHQ. It has been shown that lifestyle is related to brain health, suggesting that lifestyle habits may be related to the maintenance of brain health. In addition, in previous studies, the whole-brain GM-BHQ has been related to dietary balance [41], lifestyle [42], social performance [43], curiosity and grit [44], and diversity understanding [45]. These studies were conducted in healthy middle-aged adults, not elderly people or people with underlying diseases, so it is expected that using the same index will add depth to the interpretation of the study. Our study, which focuses on GMV, a function of surface area and cortical thickness, is more comprehensive than studies using other neural biomarkers such as functional MRI (fMRI), which is excellent for examining brain responses [46], and cortical thickness, which is suitable for observing genetic influences [47].

1.3. Analysis

Two types of measurements were performed: MRI brain imaging and facial expression measurement.

(a)

MRI brain imaging

With the cooperation of Brain Impact, a general incorporated association, brain MRI imaging was performed twice, in July and December 2023.

(b)

Facial expression measurement

For subjects who underwent brain MRI imaging, facial expression measurements were also performed at the same time, with 37 subjects being measured twice, in July and December, and 21 subjects being measured once, in July or December.

For facial expression measurement, an original PC application developed by the authors was used. This application imitates facial expressions in photos displayed on a PC screen in front of a camera, and as shown in Figure 1, there are four types of facial expressions: joy, anger, sadness, and surprise. The photos were displayed randomly for each subject, and facial expression data was obtained from the subjects when they continued to imitate each facial expression for 5 s.

A total of 95 sets of data from the same subject, each consisting of BHQ from MRI brain images and facial expression data, were analyzed.

(a)

BHQ from MRI brain images

With the cooperation of the Brain Impact Association, GM-BHQ values for the entire brain and for each brain region, DMN, CEN, and SN, were calculated from MRI brain images and used for analysis.

(b)

Facial expression measurement data

In total, 468 facial landmarks were detected from the measured facial expression data (see Figure 2), and the time when the landmark change was maximum for each facial expression when imitating four types of facial expressions, joy, anger, sadness, and surprise, was used as the reaction time for analysis.

1.4. Results

1.4.1. Association Between BHQ and Age

As mentioned above, previous research has reported that BHQ tends to decrease with age. First, we used BHQ data from brain MRI images of 362 healthy men and women aged 22 to 67 (262 men, 100 women) that we had obtained separately to examine the association between BHQ and age.

As an example, Figure 3 shows the relationship between whole-brain GM-BHQ and age. The correlation coefficient between whole-brain GM-BHQ and age was negative and substantial (r = −0.71), indicating a strong age-related association.

In analyzing the relationship between whole-brain GM-BHQ and facial expressions, we therefore adjusted for age by using the residual of GM-BHQ from the regression line with age (hereinafter referred to as res_GM-BHQ) as the objective variable, and searched for the facial-expression parameters that were optimal for explaining res_GM-BHQ.

We also examined the relationship between age and GM-BHQ in the DMN, CEN, and SN brain regions, finding substantial age-related associations in all three networks (DMN: r = −0.67; CEN: r = −0.61; SN: r = −0.67).

Based on these results, we used a similar approach to extract facial expression features that estimate each “res” component for the DMN, CEN, and SN, using res_DMN, res_CEN, and res_SN as objective variables.

1.4.2. Building a BHQ Estimation Model

Using each “res” component as the objective variable and the reaction time in the X and Y directions of each facial landmark as the variables, the analysis was carried out in the following steps to extract facial expression features that can estimate each “res” component. Reaction time was defined as the time point at which the magnitude of landmark displacement reached its maximum for each facial expression.

Step 1: Extract variables correlated with the objective variable.

Step 2: In order to exclude variables with multicollinearity, variables with high correlation coefficients were grouped together, and only the variable most correlated with the objective variable for each group was extracted.

Step 3: Construct a linear multiple regression model using the forward variance method.

Variables with a correlation coefficient |r| ≥ 0.2 with the objective variable were first extracted. Variables with correlation coefficients ≥ 0.5 were grouped together to address multicollinearity. The final models included 5 predictors for res_GM-BHQ, 4 for res_DMN, 8 for res_CEN, and 6 for res_SN.

In selecting the model, cross-validation was performed on linear regression models, linear mixed models, and random forest models using the Leave One Group Out method, and it was decided to adopt the linear regression model with the highest accuracy.

In the present study, cross-validation was performed at the measurement level, where each participant–session pair was treated as one validation unit. Because some participants were measured twice, repeated observations from the same individual could appear in both training and test folds when originating from different sessions. Therefore, the reported performance metrics should be interpreted as within-sample predictive performance rather than definitive evidence of generalizable prediction accuracy for completely unseen individuals.

More specifically, the 95 observations consisted of repeated measurement sessions from 58 participants, meaning that some individuals contributed more than one session to the dataset. This design increases the possibility that stable participant-specific facial characteristics may have contributed to model performance. Therefore, the present validation should be interpreted as preliminary internal validation rather than subject-independent external validation.

Facial landmarks were extracted using MediaPipe FaceMesh (Google LLC, Mountain View, CA, USA), which provides 468 facial landmarks. Participants were instructed to maintain a direct frontal orientation toward the camera, and facial position was aligned within the application interface before recording to ensure consistent spatial positioning.

We acknowledge that subject-level cross-validation would provide a more rigorous estimate of generalization performance. However, due to the limited sample size and repeated-measures structure of the current dataset, implementing a fully independent subject-level validation would substantially reduce statistical power. Therefore, we retained the measurement-level validation as an exploratory approach and explicitly frame the reported performance as preliminary and not indicative of generalizable prediction accuracy. Future studies with larger samples will employ subject-level validation. Importantly, the performance values reported in

Table 2. Comparison of predictive performance for residualized brain-health-related indices estimated from facial expressions and age alone.

	Estimated Results from Facial Expressions		Estimated Results from Age
Regions of brain	Correlation coefficient	MAE	Correlation coefficient	MAE
res_GM-BHQ	0.75	3.12	0.66	3.54
res_DMN	0.58	4.14	0.53	4.34
res_CEN	0.63	3.87	0.50	4.25
res_SN	0.74	4.05	0.68	4.53

Note: All indices shown in this table are age-adjusted residualized values (i.e., residuals from regression models with age), rather than raw GM-BHQ values.

are based on this measurement-level validation and should not be interpreted as subject-independent predictive accuracy.

1.4.3. Extracted Features

As a result of the analysis, Figure 4 shows an example of facial expression features extracted to estimate res_GM-BHQ. Figure 4 shows the facial landmarks that are key when imitating the expression of “happiness,” and it can be seen that the area around the eyes is important.

Figure 5 also shows the percentage of people who showed a strong correlation with the happiness scale, with a correlation coefficient of 0.7 or more, when imitating the ex-pression of “happiness,” with darker colors indicating a higher percentage showing a stronger correlation. We can see that when imitating the expression of “happiness,” many people make the expression with movements around the mouth, and the percentage of people who make the expression with movements around the eyes tends to be lower. From the above, it can be assumed that people who are able to imitate facial expressions while moving parts of the face that are difficult for most people to move as widely and quickly as possible, rather than parts of the face that are easy to move, have a high res_GM-BHQ.

As a result of the analysis, five items were extracted for res_GM-BHQ, four items for res_DMN, eight items for res_CEN, and six items for res_SN as facial expression features for estimating each “res” component, and a linear multiple regression model was constructed.

1.4.4. Predictive Performance for Whole-Brain res_GM-BHQ

Figure 6 shows the results of estimating res_GM-BHQ for the whole brain using the estimation model constructed from the extracted facial-expression features. The correlation coefficient between MRI-derived res_GM-BHQ and facial-expression-estimated res_GM-BHQ was 0.75, the corresponding coefficient of determination (R²) was 0.56, suggesting a substantial within-sample association. In addition, the mean absolute error (MAE) for res_GM-BHQ was 3.12. For the 95 datasets analyzed in this study, the correlation coefficient between age-based estimated res_GM-BHQ and MRI-derived res_GM-BHQ was 0.66, and the MAE was 3.54. These results suggest that adding facial-expression features may improve within-sample predictive performance beyond age alone. Confidence intervals for the correlation estimates were not calculated in the present study and should be examined in future work to assess statistical robustness.

1.4.5. Predictive Performance for res_DMN, res_CEN, and res_SN

Similarly, the results of estimating the age-adjusted residualized indices for the DMN, CEN, and SN are shown in Figure 7. The correlation coefficients between MRI-derived res_DMN, res_CEN, and res_SN and those estimated from facial expressions were 0.58, 0.63, and 0.74, R² values were 0.34 (DMN), 0.40 (CEN), and 0.55 (SN), respectively, suggesting moderate to substantial within-sample associations. In addition, the MAE values for res_DMN, res_CEN, and res_SN were 4.14, 3.87, and 4.05, respectively. For the 95 datasets analyzed in this study, the correlation coefficients between the values of res_DMN, res_CEN, and res_SN estimated from age alone and those derived from MRI images were 0.53, 0.50, and 0.68, respectively, and the corresponding MAE values were 4.34, 4.25, and 4.53. Compared with estimation based on age alone, the correlation coefficients were higher and the MAE values were improved, suggesting that adding facial-expression features may improve within-sample predictive performance for these regional indices as well. The predictive performance results for res_GM-BHQ, res_DMN, res_CEN, and res_SN are summarized in Table 2.

1.4.6. Developing Applications

Using the estimation model described above, we developed a PC application called “Estimated BHQ” that provides an exploratory estimate of brain-health-related indices from facial expressions. With this application, users simply follow the instructions on the screen in front of the camera and imitate four types of facial expressions—happiness, anger, sadness, and surprise (see Figure 8)—to obtain an exploratory estimate of whole-brain and network-related brain-health indices, including indices corresponding to the DMN, CEN, and SN (see Figure 9). This allows users to easily and quickly obtain an exploratory estimate of brain-health-related indices and brain-related characteristics.

1.5. Discussion

By analyzing the correspondence between GM-BHQ, which indicates the state of brain health, and facial ex-pression data when four types of facial expressions (happiness, anger, sadness, and surprise) were made, we were able to extract facial expression features that are important for estimating GM-BHQ. In addition, we constructed an estimation model using these features to estimate the GM-BHQ of the entire brain, and the correlation coefficient between the GM-BHQ from brain MRI images and the GM-BHQ estimated from facial expressions was 0.75, and the MAE was 3.12. Similarly, when we estimated the GM-BHQ of each brain region, the DMN, CEN, and SN, the correlation coefficients between the DMN, CEN, and SN from brain MRI images and the DMN, CEN, and SN estimated from facial expressions were 0.58, 0.63, and 0.74, respectively, and the MAE was 4.14, 3.87, and 4.05, respectively, although the estimation accuracy was lower than that of the GM-BHQ of the whole brain.

Furthermore, we developed a PC application called “Estimated BHQ” that estimates brain health from facial expressions, and have conducted various demonstration experiments to confirm user needs and acceptability for measuring brain health [48,49,50]. Through these demonstration experiments, we have preliminarily explored the possibility that by providing unprecedented new value of being able to measure brain health in a fun and easy way, we can arouse users’ interest in brain health, and by visualizing brain health, we can encourage them to try measuring it again. These results suggest that presenting brain-health-related information may encourage user interest and engagement, although this possibility requires further study.

Currently, it is possible to visualize brain health, but it is not yet possible to present specific actions on how users can maintain and improve their brain health based on the measurement results. In the future, we plan to explore measures to improve brain health while also pro-posing improvement actions tailored to each user. To our knowledge, this study is among the first exploratory attempts to develop technology that estimates brain-health-related indices from facial expressions. However, it should be noted that the regression-based models capture statistical associations rather than causal effects. Because the design is cross-sectional, the analyses do not permit causal inference. We hope to continue developing brain health estimation technology and proposing measures to improve brain health, thereby potentially contributing to future research on brain-health-related assessment and well-being. The present findings should therefore be interpreted as exploratory and hypothesis-generating rather than confirmatory.

1.6. Limitations

The current study has several important limitations. First, the algorithm developed in Study I provides an exploratory estimate of GMV rather than an exact neuroimaging-based value. Therefore, it should be used with caution in settings requiring high individual-level precision, such as medical or diagnostic applications. Second, the study was conducted in Japanese participants, and the findings cannot be directly generalized to other cultural or national contexts. Third, the gender imbalance in Study I (47 males and 11 females) may limit external validity. Fourth, the sample size was relatively small compared with the dimensionality of the facial landmark features, which increases the risk of overfitting. Although dimensionality reduction was performed through correlation-based screening and grouping, the number of initial candidate facial features was large relative to the sample size, and therefore the possibility of overfitting cannot be fully excluded. Fifth, repeated measurements from the same participants were included in the dataset, and the internal validation procedure was performed at the measurement level rather than strictly at the participant level. As a result, participant-specific facial characteristics may have contributed to the reported model performance. Sixth, the estimation model has not yet been externally validated, and its predictive performance in completely unseen individuals remains unclear. Finally, camera-related and environmental factors were not systematically modeled. Future studies should examine the robustness of the model using larger and more diverse samples, participant-level validation, and independent external cohorts.

2. Study II: An Application of Estimated GMV to Neuroaesthetics Research

2.1. Introduction

In recent neuroscience, the effects of artistic activities and reading have been attracting attention as a way to reduce the risk of dementia in the elderly. Sensing beauty activates the entire brain, including the brain’s reward system, such as the frontal orbital cortex and striatum [51,52]. Therefore, it has been reported that there is a relationship between high creativity and several brain regions. In particular, it has been revealed that DMN cooperates with CEN to improve creativity and artistic performance [53,54]. Creative activities may have a positive effect on the brain not only by performing them yourself, but also by watching and feeling the state of the creative activity. It has been reported that the DMN is engaged when people are deeply moved by a work of art or when they feel it is beautiful [55,56]. On the other hand, when learning something from literature, we connect various events and imagine the mental scenes of people at that time. Such cognitive acts are related to the activity and connectivity of the hippocampus and prefrontal cortex [57,58]. Previous evidence suggests that there is a relationship between the retrieval of autobiographical memories and CEN [59,60]. However, the effects of such artistic activities and reading on the brain structure of healthy middle-aged adults remain unclear. The possibility that engaging in intellectual activities at a young age may prevent or delay the onset of dementia has been asserted and supported by research on “cognitive reserve” [61,62,63].

Therefore, in this study, we focus on the GMV of the Healthy adult brain. GMV was determined as the GM-BHQ value originally based on MRI image analysis. This value can be estimated by measuring facial expressions as shown in Study I. The representative brain networks are three networks that control rational judgment, emotion control, other recognition, self-recognition, and behavior evaluation. Of these, CEN, composed of the dorsolateral prefrontal cortex and posterior parietal cortex, is critical for the active retention and manipulation of information in working memory, attention, problem solving, decision-making, and self-awareness [64,65]. The SN is also a network that includes the ventrolateral prefrontal cortex (VLPFC) and anterior insula (collectively referred to as the fronto-insular cortex FIC), as well as the anterior cingulate cortex (ACC) [66], and responds to subjective salience, whether cognitive, homeostatic, or emotional [67]. The SN also acts as a switch between the CEN and the DMN, inhibiting the latter and activating the former when salient stimuli or cognitive tasks are at hand, a process essential for attention and flexible cognitive control [68,69,70,71,72,73]. On the other hand, the DMN includes the medial posterior cortex, including the posterior cingulate cortex (PCC) and parts of the frontal lobe, the medial prefrontal cortex (MPFC), and the posterior temporal region around the temporoparietal junction (TPJ), including the inferior parietal lobe (IPL) [74,75,76]. The DMN is preferentially activated when individuals are not focused on the external environment [75] and are engaged in various areas of cognitive and social processing. That is, the medial prefrontal cortex (MPFC) plays a key role in the social understanding of others, and the connections between the medial prefrontal cortex (aMPFC) and the posterior cingulate cortex and anterior cingulate cortex mainly contribute to self-other discrimination. On the other hand, the connection between the dorsal MPFC (dMPFC) and the temporoparietal junction (TPJ) is mainly related to the understanding of others’ mental states [77].

The interactions between these networks are related to daily self-regulation and empathy for others [78,79,80,81,82,83]. However, to our knowledge, there have been no studies that have addressed the relationship between the brain structure centered on these triple networks and artistic activities and reading habits. Therefore, in this study, we contribute to the development of related research by exploring how estimated brain-health-related indices, particularly those related to the DMN and CEN, are associated with creative behavior and reading habits in healthy adults using questionnaire data and facial-expression-based estimation.

2.2. Materials and Methods

2.2.1. Participants

A priori power analysis was conducted for Study II because it involved hypothesis-driven inferential analysis. In contrast, Study I focused on exploratory model development and therefore did not include a priori power analysis. The sample size required to perform correlation analysis at a significance level of 5% and power of 80%, with an effect size of r = 0.3, which corresponds to a “moderate correlation” according to Cohen’s criterion, was calculated as 109 samples using G*Power 3.1.9.7. A total of 113 people (58 men, 54 women, and 1 unknown) participated in the study, with a mean age of 39.0 ± 10.1 years. Participants were randomly selected from among employees of a single private company and asked to participate. Participants completed online questionnaires and facial expressions measurement from March to April 2025. According to self-report, none of the participants had a history of neurological, psychiatric, or other medical conditions that could affect the central nervous system. All methods were carried out in accordance with relevant guidelines and regulations, and all participants provided written informed consent before participation, and anonymity was maintained. This study was approved by the Institute of Science Tokyo’s Brain Information Cloud (Ethics Committee for Human Research: Permit Number 2023137) Ethics Committee.

Table 3. Participant information (frequency).

	N	%
Sex
Male	58	51.3
Female	54	47.8
No answer	1	0.9
Position
Executives/executives	6	5.3
Department head	2	1.8
Manager/leader	40	35.4
General/professional	62	54.9
Contract/individual business	3	2.7
Occupation
Managerial occupations	17	15
Research/technical occupations	9	8
Professional occupations such as legal affairs, management, culture/arts, etc.	17	15
Clerical occupations	26	23
Sales/sales occupations	5	4.4
Service occupations	39	34.5

and

Table 4. Participant information (minimum, maximum, and average value).

	Mean	N	SD	Min	Max
Age	39.0	113	10.1	24	68
BMI	22.8	113	4.5	17.2	55.2
Years of education	16.7	113	2.0	6	22
Years of employment	3.5	113	1.7	1	6

SD: standard deviation.

show information on various variables of participants.

2.2.2. Estimating Brain Information Using Facial Expression Information

In this study, we used facial-expression-based estimated GM-BHQ-related indices derived from the model developed in Study I. These values should be interpreted as indirect exploratory estimates of brain-health-related characteristics rather than as direct MRI-derived structural measures. The reason for using the estimated GM-BHQ is that GM-BHQ has been used primarily in studies of healthy middle-aged adults, as in this study, rather than in elderly people or people with underlying diseases, and correlations with various psychological indices have been confirmed. In our previous studies, whole-brain GM-BHQ was positively correlated with cognitive ability [84], job satisfaction [85], dietary balance [41] and negatively with unhealthy lifestyle [42], fatigue and stress [86], and polluted residential environment [87]. Furthermore, in a recent pilot study, the DMN, CEN and SN of GM-BHQ were positively correlated with understanding of gender and origin diversity [45]. Also, consistent with neuroplasticity claims [88], GM-BHQ has been shown to change with certain interventions [89,90]. Accordingly, the present study uses these indices as exploratory proxies for brain-health-related variation in a non-MRI setting.

2.3. Psychological Test

2.3.1. Creative Behavior

This study used the Aesthetic Responsiveness Assessment developed by Schlotz et al. [91] and adapted by Magsamen and Ross [92]. The AReA originally is a short (14-item) questionnaire for assessing how responsive a person is to aesthetic experiences, such as with visual arts, music, dance and performance. Among the three subscales of AReA, namely Aesthetic Appreciation (8 items), Intense Aesthetic Experience (3 items), and Creative Behavior (3 items including “I sculpt, paint, draw, direct films, or do design work”), we used a revised creative behavior scale consisting of four items: the original three Creative Behavior items plus one item from Aesthetic Appreciation, “I visit museums or go to musical/dance performances.” This revision was made because the added item is an item that represents behavior, like the three items in Creative Behavior, and a factor analysis using test data from 80 people collected through general recruitment showed that the factor was more stable when it was included in Creative Behavior rather than Aesthetic Appreciation, and also because an improvement was observed in the Cronbach’s alpha reliability coefficient of the factor (available upon request). Each item was answered on a 5-point Likert scale, and each response was assigned a score from 1 to 5. The creative behavior was calculated as the average of 4 items. Cronbach’s alpha reliability coefficients of creative behavior were 0.662.

2.3.2. Reading Habits

The study used two items from the adult version of the OECD Programme for International Student Assessment (PISA): “How often do you read books in your current job?” and “How often do you read novels or non-fiction books outside of work?” [93]. Each item was answered on a 5-point Likert scale, and each response was assigned a score from 1 to 5. Reading habits were calculated as the average of the 2 items.

2.3.3. Control Variables

Sex (male = 1; female = 0), age (years old), educational background (years) and BMI (kg/m²) were used as control variables.

2.4. Data Analysis

Correlation and partial correlation analyses were performed between the estimated whole-brain and regional GMVs of the CEN, SN, and DMN, creative activities and reading habits variables. Partial correlation analyses were performed controlling for sex, age, educational background, and BMI. The standard for statistical significance was 5% two-sided. All statistical analyses were performed using IBM SPSS Statistics Version 28 (IBM Corp., Armonk, NY, USA). Because the analyses were exploratory and limited to a small number of theoretically motivated comparisons, no formal correction for multiple testing (e.g., Bonferroni or FDR) was applied, and the results should therefore be interpreted cautiously.

A total of 4 statistical tests were conducted for creative activities and reading habits variables, respectively. If a Bonferroni correction were applied, the adjusted significance threshold would be p < 0.05/4 = 0.0125. Under this criterion, none of the observed associations remain statistically significant. Therefore, the reported findings should be interpreted as exploratory.

2.5. Results

In

Table 5. Correlation coefficient.

		1	2	3	4	5	6
1	GM-BHQ		0.650 ***	0.568 ***	0.737 ***	−0.012	0.186
2	DMN	0.923 ***		0.643 ***	0.610 ***	0.095	0.231 *
3	CEN	0.858 ***	0.878 ***		0.617 ***	0.231 *	−0.028
4	SN	0.940 ***	0.910 ***	0.872 ***		0.076	0.166
5	Reading habits	−0.074	−0.042	0.060	−0.044		0.253 **
6	Creative behavior	0.164	0.187 *	0.065	0.167	0.195 *

N = 113; * p < 0.05; ** p < 0.01; *** p < 0.001. The figures below the diagonal are Pearson’s correlation coefficients. The figures above the diagonal are Pearson’s correlation coefficients controlling for demographic variables.

, the figures below the diagonal show Pearson correlation coefficients, and the figures above the diagonal show partial correlation coefficients controlling for demographic variables. Zero-order correlations between the estimated brain-health-related indices and the behavioral variables were generally small and mostly non-significant. Therefore, the partial correlation results are emphasized below.

Estimated DMN-related indices showed a small but statistically significant positive partial correlation with creative behavior (r = 0.231, p = 0.016), but not with reading habits (r = 0.095, p = 0.329). Estimated CEN-related indices showed a small but statistically significant positive partial correlation with reading habits (r = 0.231, p = 0.016), but not with creative behavior (r = −0.028, p = 0.777). No evidence supporting a broader triple-network pattern was observed, as no significant partial correlations were found for the other estimated indices. Figure 10 and Figure 11 illustrate these two significant associations.

2.6. Discussion

The brain-health-related indices used in Study II were estimated from facial expressions using the model developed in Study I rather than directly measured from MRI. Therefore, these variables should not be interpreted as equivalent to true MRI-derived GMV measures. Instead, they should be regarded as indirect exploratory estimates that may contain prediction error and participant-specific bias. Accordingly, the observed associations should be interpreted cautiously.

In addition, the analyses were exploratory and no formal correction for multiple comparisons was applied. The statistically significant associations observed in this study were small in magnitude, and should therefore be interpreted as preliminary rather than conclusive evidence. The present findings are better understood as hypothesis-generating observations that may inform future MRI-based and longitudinal studies.

The results of this study suggest that facial-expression-based estimated brain-health-related indices, particularly those related to the DMN and CEN, may be associated with intellectual activities such as creative behavior and reading habits in healthy adults. This result is broadly consistent with previous studies that have linked the DMN to artistic and aesthetic engagement [53,54,55,56] and the CEN to reading and related cognitive processes [57,58,59,60]. The participants in this study were self-reported healthy working adults. Therefore, the present findings may suggest that engagement in intellectual activities such as art and reading is associated with brain-health-related characteristics in healthy adults, although longitudinal and interventional studies are needed to clarify causal direction. This interpretation is also broadly relevant to the concept of cognitive reserve [61,62], but the present findings should be regarded as exploratory.

In a previous study in which 16 young adults were tasked with rating images of artworks, areas of the medial prefrontal cortex, measured in an fMRI scanner, were positively activated in the best-rated trials. This means that the rating-specific neural response is attributable not to the characteristics of the particular artwork, but to the aesthetic experience. In other words, DMN activity is due to the fact that a particular artwork fits the unique structure of the individual so well that it accesses neural substrates that are involved in the self, leading to a feeling of being “moved.” This explanation is consistent with the contemporary idea that an individual’s taste in art is linked to a sense of identity [94,95]. The involvement of the DMN in painting appreciation is further supported by the results of a recent meta-analysis [96].

On the other hand, an experimental study of the reading comprehension of 60 school-age children showed that executive functions make a significant contribution to reading comprehension [97]. Subsequent meta-analysis also supported a moderate positive association (r = 0.36) between executive function and reading comprehension [98]. Consistent with these, a recent neuroscience study of 55 school-age children showed that the discrepancy between the ability to convert printed words into sounds and reading comprehension of text was associated with a frontoparietal network including the GMV of the left dorsolateral prefrontal cortex [99].

These exploratory findings may be relevant to the broader concept of cognitive reserve, suggesting that different forms of intellectually and aesthetically engaging behavior may be associated with distinct large-scale brain-network-related characteristics. However, the present data are cross-sectional and indirect, and therefore cannot determine whether such activities shape brain health or whether individuals with certain brain-related characteristics are more likely to engage in them.

Thus, despite the use of a simplified facial-expression-based estimation approach, the present findings are broadly consistent with prior neuroscience literature on intelligence, brain function, and neurofeedback-related brain monitoring [100,101,102]. Conventional neuroscience research often relies on expensive and less accessible methods such as MRI, which limits feasibility in everyday settings. In contrast, the simplified approach used in this study may offer an exploratory low-burden tool for studying brain-health-related characteristics outside laboratory or hospital environments. Therefore, the present findings may provide a preliminary basis for future low-burden approaches to brain-health-related research in daily life. Taken together, the present findings should be regarded as exploratory and hypothesis-generating rather than confirmatory evidence.

2.7. Limitations

This study has some limitations. Firstly, it should be noted that the results of this study were estimated from facial expressions, not GMV calculated from MRI images, so the findings should ideally be re-examined using directly MRI-derived GMV measures. Secondly, this study was conducted on Japanese employees from a single company, and caution is required if the results are applicable to other countries. Japanese people have a unique sense of beauty, such as wabi-sabi and kawaii, and such peculiarities may have biased the results. They are also known to be conservative and, compared to other developed countries, enjoy reading but are not very active in artistic activities. Thirdly, because general measures of art and reading were used, the results of this study cannot suggest specific activities. In addition, the creative behavior scale showed only moderate internal consistency (Cronbach’s α = 0.662), and this measurement limitation should be considered when interpreting the corresponding associations. Fourthly, this study is a cross-sectional analysis and does not show a causal relationship. For example, rather than the arts and reading enhancing brain health, it may be that brain health drives people’s choices about the arts and reading. Fifthly, the small sample size limits the generalizability of the results. Sixthly, in this study, sex, age, educational background and BMI were used as control variables. However, background information not adopted in this study, such as socioeconomic status such as income and differences in the community to which the participants belong, may have been related to the results. Beauty is not a distinct, independent element embedded in a color, food, nature, landscape, idea, or even a face, nor is it something an artist “puts” into a work as a distinct entity; it is a property that emerges in the beholder’s brain. That aesthetic response is shaped by the beholder’s cultural experiences, life events, education, genetic inheritance, and perhaps gender, age, and health. Individual variation in the intensity and extent of the response has not been systematically explored, and there remains much to be explored [103]. Seventhly, this study was unable to clarify the relationship between SN and art and reading. This is an issue that remains for future efforts to improve brain health at the whole-brain level. Eighth, even if art and reading are good for the brain, this study does not allow us to infer the extent to which their effects are compared to other activities. With the advancement of neuroscience, there are many activities that are said to be good for the brain (exercise, diet, brain training, etc.). Since the general public has limited free time and money, the relative importance of neuroaesthetic-related activities, how often and for how long they are effective, and above what level the incremental effect disappears, need to be gradually clarified in the future. Ninth, the present study was limited to a triple network: a recent meta-analysis showed that neural responses to artistic beauty recruit a broadly distributed network in both hemispheres, including content-dependent brain regions in the ventral visual stream [104]. Previous findings have also shown that the originality of art is associated with the DMN, while beauty is associated with a network of areas including the occipital and frontal lobes [105]. As such, the latest findings in brain research indicate the possibility of a more diverse and complex network of neural responses to beauty, leaving room for further development of the simplified measuring device used in this study. In addition, because the estimated indices used in Study II were derived from facial-expression data, any uncontrolled variability in image capture conditions (e.g., camera quality, lighting, and positioning) may also have influenced the estimates. Finally, most previous studies have used fMRI, which differs from the data of this study, which is based on MRI. Although fMRI is excellent for measuring brain responses, there is insufficient evidence as to whether the responses have a lasting effect on brain structure. As neuroaesthetics research using MRI becomes more active and the relationship with brain structure becomes clearer, the value of the simple measurement used in this study is expected to increase even further. Future research should seek to verify the results of this study using other methods, such as MRI and fMRI, in a longitudinal manner that includes a variety of intervention methods and control variables, with a larger number of people in different cultures, in order to obtain further new findings.

3. Conclusions

In Study I, by analyzing the correspondence between the BHQ calculated from MRI images of the brain and the facial expression data when four types of facial expressions (happiness, anger, sadness, and surprise) are made, an algorithm was constructed to estimate brain health from facial expressions, and an application was developed to estimate brain health. This allows users to obtain an exploratory estimate of brain-health-related indices from facial expressions.

In Study II, a survey was conducted on 113 healthy Japanese adults using facial-expression-based estimated brain-health-related indices and questionnaires, and small positive associations were observed between DMN-related and CEN-related estimated indices and artistic activities and reading habits, respectively. These findings provide an exploratory basis for future research on whether facial-expression-based estimated brain-health-related indices may be useful for studying associations between everyday intellectual activities and brain-health-related characteristics in healthy adults.