EEG-Driven Personal Comfort Model for Cognitive Efficiency in Human-Centric Environments

Abstract

This study aims to develop a personal comfort model driven by real-time electroencephalogram (EEG) signals for constructing built environments customized to individual emotional states and preferences. EEG signals from a single subject were collected at regular intervals under controlled environmental conditions—temperature, humidity, and illumination. Real-time deep learning methods processed the sensor data, enabling effective prediction of the user’s preferred conditions. Model evaluation showed reliable predictions on the personal dataset, allowing for optimized lighting that enhanced concentration and reduced stress. These findings indicate that EEG can inform personalized environmental modifications. This integration of EEG and deep learning provides objective, precise comfort assessment and supports immediate environmental adaptation.

1. Introduction

Improving the user experience is crucial to human–computer interaction (HCI). It is imperative to comprehend and enhance the diversity of individual experiences [1]. Given that most individuals spend a substantial portion of their time within indoor environments—more than 90%—the quality of these spaces considerably affects various aspects of well-being, including health, productivity, and comfort [2]. Consequently, several initiatives have been undertaken to enhance building occupants’ comfort and well-being while increasing efficiency [3,4]. Assessing and optimizing human comfort in the built environment presents a complex challenge, given the extensive range of variables involved [5]. Research has demonstrated that perceptions of comfort are subjective and influenced by many factors, including thermal, visual, air quality, and acoustics [6].

These factors have been shown to affect health and productivity significantly [7,8,9]. Quantifying environmental factors in physical spaces is relatively straightforward, but accurately assessing comfort levels remains challenging [6]. Recent research has found a notable increase in interest in technologies capable of automatically recognizing emotional states. Emotions are essential in human physiological and mental states [10]. Electroencephalography (EEG) is a tool that can identify brain activity associated with different states over time [11]. Its potential applications have been recognized in various fields, including marketing, fatigue monitoring, e-learning, and healthcare [10,12,13,14,15,16].

Contemporary advancements in sensing and communication technologies have rendered the collection and processing of substantial data in real time a viable proposition [17]. This technology has demonstrated the capacity to detect human activity in smart homes via wearable devices and sensors distributed throughout the environment [18]. These developments have facilitated the integration of wireless sensor networks in the construction industry [19]. Regardless, existing research has concentrated on developing autonomous building systems for various applications, including health monitoring and fire detection [20,21]. These general applications do not reflect the diverse physiological conditions of individuals, which constrains their capacity to fulfill individual environmental preferences and needs [22,23,24].

This study aims to develop a real-time EEG-driven personal comfort model to create emotionally responsive architectural spaces, enabling enhanced built environments tailored to individual emotions and preferences. Specifically, it employs real-time deep learning methodologies to predict the preferences of space users effectively by utilizing EEG. The analysis methodology is appropriate for time series data and examines users’ EEG measurements in varying spatial contexts and circumstances. The research questions were derived by decomposing the expected outcomes regarding their psychological relationships with illumination as the primary environmental factor. The study’s hypotheses were:

• Specific illumination levels in indoor environments can reduce users’ stress, improve attention and engagement, and maximize comfort.
• Deep learning models utilizing EEG and illumination data can predict individual comfort in real time and recommend optimal lighting conditions.

After the introduction, we review the relevant literature to examine the need for personalized spatial environment design and real-time personal comfort models according to emotional cognition, laying the groundwork for the proposed research. Next, we provide a detailed account of the experimental setup, model structure, and implementation process. We present a comprehensive account of the methodology employed in constructing the real-time model and a detailed exposition of the experimental findings. In conclusion, we examine the implications, contributions, limitations, and potential avenues for future research presented in this paper.

2. Literature Review

2.1. Personalized Environmental Spaces for Occupants

The comfort level of a spatial environment significantly affects a user’s learning ability and productivity [25,26,27]. Many factors, including physiological, psychological, and physical elements, determine human comfort [28]. In this context, several studies have examined the relationship between physical measurements and subjective responses, thereby contributing to the development of indoor environmental control solutions that optimize user comfort [29,30]. The indoor environment exerts various health, well-being, and performance effects.

Likewise, subjective responses can vary significantly from one individual to another, even when the conditions remain identical [28]. In particular, light significantly influences various cognitive processes, including attention [31,32,33], working memory [34,35,36,37], and overall cognitive performance [38,39]. Bright light has been shown to enhance attention and working memory performance [31,34]. Lighting can influence emotional states, interacting with subjective mood and modulating emotional processing [40]. Occupants’ productivity is contingent on providing suitable indoor lighting conditions. Creating a pleasant and secure indoor environment can potentially enhance users’ comfort and well-being [41].

Incorporating personalized space design is crucial in optimizing spaces’ efficiency and user satisfaction by reflecting individuals’ specific needs and preferences [42]. A recent study indicates that implementing personalized design methodologies and utilizing computer graphics and machine learning techniques, can enhance users’ psychological comfort and task satisfaction [43]. Personalization of workspaces that automatically adjust lighting, temperature, desk height, and other factors has been shown to affect psychological comfort and job satisfaction positively [44]. Personalizing intelligent buildings is crucial for achieving an equilibrium between energy consumption and user well-being [45]. It enables developing novel systems that provide tailored feedback and environmental conditioning to diverse users [46]. Implementing personalized environmental conditions has been demonstrated to promote health, comfort, and productivity [47,48,49,50], underscoring the significance of customized environmental systems in sustainable buildings. It has been demonstrated that the ability to control indoor environmental conditions is closely associated with users’ perceived comfort and environmental satisfaction [51,52,53]. For example, Baird and Lechat [47] noted the significance of lighting control, and Ilvitskaya et al. [49] highlighted the pivotal role of natural elements for visual comfort. In their analysis of the sustainability of innovative interior design, Barbosa et al. [50] considered the effect of learning spaces on building comfort. Similarly, López-Chao et al. [51] explored the sustainability of learning spaces.

Accordingly, personalizing built environments necessitates a comprehensive approach to optimizing user experience. Custom design has emerged as a pivotal strategy to address the diverse needs of users [54]. However, existing personalized systems primarily prioritize system efficiency, which does not fully leverage the elements of the built environment. This can be enhanced by integrating architectural design with HCI to create novel interfaces [55]. Recent studies have exemplified this integration through practical frameworks. For example, Economidou et al. [56] proposed an HBI framework that emphasizes lived experience in human–building interaction and provides guidance for incorporating occupants’ subjective perspectives into design. Hamdani and Chihi [57] introduced an Adaptive HCI model for Industry 5.0 that dynamically adapts to user behavior, cognitive load, and environmental changes, demonstrating improved efficiency and reduced cognitive strain. Lin et al. [58] presented a framework that leverages multimodal data and neural correlates of spatial experience to inform AI-supported, user-centered design tools. These approaches demonstrate the potential of integrating architectural design with HCI to advance personalized environmental systems that more effectively address users’ cognitive requirements.

2.2. Emotional Cognition and Real-Time Personal Comfort Model

Emotions are a significant aspect of daily human life, and with the advancement of HCI applications, automated emotion recognition is becoming increasingly necessary. Recently, there has been a surge of interest in emotion recognition using EEG, with promising results in real-time emotional state determination [59,60,61,62]. EEG-driven technologies have historically been employed in the medical domain. However, with the advent of wireless headsets, which have enhanced wearability, cost-effectiveness, portability, and ease of use, their applications are expanding to diverse fields [63]. EEG can detect immediate responses to emotional stimuli with exceptional temporal resolution, thereby providing a comprehensive means of emotion recognition. Further, it is an effective method for determining concentration and stress levels [64,65]. EEG signals are significant in emotion recognition because they can monitor many physiological responses in real time. For example, an EEG fatigue detection system can monitor a driver’s fatigue status in real time [66,67]. Moreover, EEG is less susceptible to deception than behavioral responses, as it captures involuntary neurophysiological activity that is difficult to consciously manipulate [68]. It can also achieve higher classification accuracy than other physiological signals, such as electrocardiogram, skin temperature, and galvanic skin response, particularly in detecting cognitive or affective states where brain activity offers more direct markers [69]. The extant real-time emotion recognition systems employ stimuli such as pictures and sounds to induce emotions and are constrained to identifying emotions from the arousal-value matrix in isolation. In light of these considerations, the proposal of a real-time system for recognizing an individual’s emotional state according to EEG represents a potential avenue for overcoming the above-stated limitations, reducing the inaccuracy of subjective reports, and accurately assessing emotional states in real time [70].

EEG-driven systems for real-time monitoring and control of individual comfort in indoor environments underscore the significance of real-time application [71]. Such systems facilitate an enhanced user experience by processing data expeditiously and providing immediate feedback. Previous studies have employed EEG signals to track users’ cognitive state but have been constrained by limitations in real-time response speed and accuracy [72]. Recent studies have employed deep learning techniques, resulting in notable advancements in the real-time processing of EEG data [73]. Deep learning techniques enable the development of real-time systems according to EEG signals [74] and facilitate the automation of data preprocessing and feature extraction, thereby enhancing the real-time response speed [75]. Real-time EEG systems monitor user health and provide immediate feedback in the event of any abnormalities [76]. They also control various devices through brain–computer interface [77]. These systems necessitate the capacity for high-speed data processing, precise signal classification, and real-time responsiveness [78,79,80,81] to maintain individual comfort and facilitate immediate response in a multitude of indoor settings. Recurrent neural networks (RNNs) are frequently employed to address the time series data characteristics of EEG signals [82] because they effectively capture the temporal features of EEG signals [83]. As Chen et al. [84] demonstrated, a real-time emotion recognition system employing gated recurrent units (GRUs) can monitor stress levels by analyzing the user’s emotional state in real time from EEG data. This enables the system to recommend suitable rest periods or stress relief techniques to the user. Integrating EEG and deep learning is a highly effective real-time approach to monitoring and analyzing an individual’s emotional state, enhancing emotion recognition, and creating a personalized and comfortable environment [85,86]. Despite recent advances, most EEG-based systems have focused on general emotion or fatigue under controlled laboratory stimuli, limiting their applicability to built environment research.

While some recent studies have begun applying EEG directly to architectural contexts, their emphases have differed from cognitive performance. For instance, the concept of sustainable neuro-responsive habitats integrates brain–computer interfaces with building information modeling to create adaptive residential spaces, highlighting the role of neurophysiological data in shaping sustainable housing design [87]. Another study demonstrated the feasibility of real-time thermal regulation by using EEG-based assessments of individual thermal sensation, thereby linking neurophysiological responses to heating, ventilation, and air conditioning (HVAC) control strategies [88]. More recently, EEG-based emotion analysis combined with CNN–LSTM deep learning techniques has been employed to predict occupants’ architectural space preferences, suggesting the potential of EEG for data-driven spatial design decisions [89]. While these approaches highlight the increasing significance of EEG in the architectural field, their focus has mainly been on thermal comfort and overall spatial preferences, with limited consideration of illumination’s impact on cognitive performance. Although illumination is well established as a determinant of comfort and cognitive performance, its incorporation into EEG-driven personal comfort models has been limited. This study addresses this limitation by developing a real-time model centered on illumination, extending EEG applications toward adaptive and cognitively supportive architectural environments.

3. Materials and Methods

3.1. Experiment Setup

3.1.1. Subjects

The single-subject research design is a methodology well suited to the detailed observation and analysis of behavioral changes in specific individuals. It has been used effectively in a variety of fields. Given the complexity and variability of human behavior, encompassing both individual and contextual variations, a methodology that applies to diverse settings and individuals is essential for user-centered research [90,91,92,93]. Using single-subject studies enhances data consistency and accuracy because it accurately captures users’ needs and preferences. This approach is fundamental in implementing customized comfort environments, offering the advantage of detailed analysis of individual users’ needs and reactions. In this study, the single-subject design was adopted as an exploratory study to examine the feasibility of real-time EEG-based comfort modeling in the context of illumination. The findings are presented as illustrative, and future research should focus on accumulating data from diverse individuals to refine and extend personalized comfort models. The detailed insights derived from repeated and controlled measurements on a single individual nevertheless provide a valuable foundation for advancing user-specific environmental models.

Accordingly, a single subject (Participant A) was selected to conduct repeated experiments to control personalized spatial environments. The participant was a 28-year-old female architectural researcher with corrected vision and normal hearing, and no history of neurological or psychiatric disorders. She maintained a regular sleep schedule and refrained from consuming alcohol, tea, or coffee for at least 24 h before each session to minimize physiological variability. The selection was also based on her familiarity with EEG equipment and ability to remain still for extended periods, which helped ensure high-quality data acquisition with minimal artifacts. This study followed the ethical guidelines set forth by the Ethics Committee of Hanyang University (HYU-2023-280) and received the committee’s approval. The participant was required to sign an informed consent form confirming full understanding of the experiment.

3.1.2. EEG Equipment

In this study, data were collected using the Emotiv EPOC X EEG headset (Emotiv Inc., San Francisco, CA, USA), with the prominent EEG bands (δ, θ, α, β, γ) as the primary focus. The electrodes for EEG measurements were positioned by the 10 to 20 international standard system, comprising 14 channels and two reference channels. The data may be monitored, processed, and analyzed in real time via the EmotivPRO v.3.7.0 software [94].

Preprocessing the EEG data entails removing power noise and high frequencies within the EPOC X headset. The data are sampled at 2048 Hz and then passed through a high-pass filter of 0.16 Hz and a low-pass filter of 43 Hz. A dual-notch filter removes noise at 50 Hz and 60 Hz. Subsequently, the data are downsampled to 256 Hz or 128 Hz. Emotiv employs machine learning algorithms to categorize various emotional states and conditions, offering three distinct detection algorithms (Performance Metrics, Mental Commands, and Facial Expressions). The performance metrics include six emotion metrics, quantified on a scale of zero to 100, representing the user’s psychological and physiological response to the emotions depicted in

Table 1. Descriptions of Performance Metrics.

Category	Description
Stress (St)	Measures comfort level with a task. Low to moderate stress can enhance productivity, whereas high stress is detrimental and affects health negatively.
Engagement (En)	Reflects alertness and attention toward a task. Higher engagement indicates increased focus and immersion.
Interest (In)	Indicates attraction or aversion to a task. High interest shows a strong liking, low interest shows aversion, and mid-range shows neutrality.
Excitement (Ex)	Measures physiological arousal with a positive value. Higher excitement scores indicate increased arousal.
Attention (At)	Measures fixed attention to a task and the frequency of switching tasks. High task switching suggests poor focus.
Relaxation (Re)	Measures the ability to switch off and recover from intense concentration. High scores are seen in trained meditators.

[95].

3.1.3. Experiment Procedure and Environment

All windows were kept closed to eliminate external natural light, and the existing lecture hall lights were switched off. Illumination was provided solely by a single ceiling light temporarily installed in the experimental room, ensuring that only controlled lighting conditions were maintained throughout the study. Additionally, the participant was provided with earplugs to mitigate the potential effect of external noise. To collect a broad range of data under varying lighting conditions, the experiment spanned three days, with five daily sessions commencing at 9 am, 12 pm, 3 pm, 6 pm, and 9 pm. The primary focus of this study was on the effects of lighting conditions, rather than on variations across different times of day. Previous research supports this approach, showing that light can modulate cognitive and neural responses [39]. Each session comprised five trials, contingent on the prevailing light conditions, and lasted approximately 55 min (see Figure 1) [96].

The experiment was conducted in a lecture hall with experimental equipment and a simple room. For EEG measurements, an experimental room measuring 800 × 1200 × 2000 mm was placed in the center of the lecture hall. The space was covered with aluminum-plated film to provide a constant and uniform lighting environment and block external interference, thereby improving the accuracy of EEG measurements (see Figure 2).

The experimental procedure was as follows. At the commencement of the experimental session, a baseline was established utilizing the EmotivPRO appliance to eliminate any extraneous electrical noise from the participant’s EEG. Environmental sensor data were collected in conjunction with the EEG data from Participant A for 5 min [97,98]. The temperature and humidity were continuously monitored and maintained within a consistent range throughout the experiment to ensure environmental stability. Following data collection, the illuminance value was maintained for five minutes to stabilize the lighting device and ensure uniformity of the experimental environment. The experimenter provided no instructions beyond those necessary to operate the lighting device, and the participant was not permitted to look directly at the light.

Area A is the site of EEG experimentation and is equipped with standing desks. Illuminance sensors were installed in Area B to measure the lighting conditions continuously. These sensors are labeled by the Korean Industrial Standards (KS) C 7612:1987 (illuminance measurement for lighting installations) [99], and one is installed at each end of the upper surface (virtual surface within 5 cm) of the desk work surface. Intelligent temperature and humidity sensors are positioned in Area C to record environmental data. The ceiling luminaire in Area D is equipped with five levels of illuminance control, according to the KS A 3011:1998 (recommended levels of illumination) [100], offering 150, 300, 450, 600, and 750 Lx.

3.2. Data Acquisition and Preprocessing

The collection and preprocessing of various sensor data were divided into two principal stages for model training. In the initial phase, diverse sensor data, encompassing EEG-derived performance metrics and spatial environmental data, were gathered. The EEG dataset included metrics related to stress (St), engagement (En), and attention (At). These metrics serve as critical indicators of the user’s psychological and physiological state, collectively representing personal comfort by capturing negative and positive cognitive and emotional dimensions. Specifically, stress reflects the degree of tension experienced, while engagement and attention indicate the level of immersion and focus, respectively.

Prior studies support the rationale for this selection. Lang et al. [101] demonstrated that the relative power of alpha (α), theta (θ), and beta (β) EEG frequency bands is significantly and positively correlated with both thermal comfort and cognitive task performance, with α-band power in particular serving as a robust indicator of subjective comfort. Pino et al. [102] showed that increased EEG-derived engagement elevates cognitive function and is closely linked to a flow state—an optimal, low-stress condition of immersive, productive focus. Conversely, Veranic et al. [103] found that the suppression of α-band activity, particularly during the intrusion of personal space, reflects heightened attentional engagement associated with anxiety and vigilance rather than comfort per se. Their findings emphasize that increased attention is not inherently indicative of well-being; rather, it may reflect a defensive state elicited by environmental stressors.

Building on these findings, our model conceptualizes personal comfort as a positive state characterized by low stress and high levels of engagement and attention. That is, while both engagement and attention reflect increased cognitive focus on tasks or the environment, only engagement under low stress reliably corresponds to optimal comfort and performance. This distinction allows our model to capture the nuanced interplay between these neurophysiological states, thereby providing a comprehensive measure of individual well-being in dynamic environments. In summary, our approach defines optimal personal comfort as a state of high engagement and attention under minimal stress, ensuring that defensive or vigilance-driven attention is not conflated with comfort.

Thus, these three metrics provide a comprehensive real-time assessment of personal comfort. To operationalize this, the Personal comfort (PC) index can be conceptually modeled as a weighted combination of these EEG metrics as follows:

$$\mathrm{P}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{o}\mathrm{n}\mathrm{a}\mathrm{l} \mathrm{C}\mathrm{o}\mathrm{m}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{t} \left(\mathrm{P}\mathrm{C}\right)=\mathrm{\alpha }\times (1-\mathrm{S}\mathrm{t})+\mathrm{\beta }\times \mathrm{E}\mathrm{n}+\mathrm{\gamma }\times \mathrm{A}\mathrm{t}$$

(1)

where α, β, γ are weighting coefficients reflecting the relative influence of each metric (assigned as α = 0.5, β = 0.25, γ = 0.25 to emphasize negative effects (stress reduction) while equally accounting for positive effects (engagement and attention)), serving as the conceptual foundation. Building upon this, actual prediction is performed using a gated recurrent unit (GRU) deep learning architecture that integrates EEG metrics with environmental factors such as illumination, temperature, and humidity, allowing the model to capture complex nonlinear and temporal relationships for adaptive and individualized environmental optimization. The model is constructed to establish an optimal environment by identifying interrelationships between various environmental factors influencing illuminance and formulating corresponding recommendations. Temperature and humidity are additionally monitored to ensure their stability, minimizing interference from other environmental variables.

EEG data were collected using EmotivPRO v.3.7.0 software, while spatial environmental data were gathered in real time via the Home Assistant system. The collected data were integrated at 10 s intervals using Node-RED (OpenJS Foundation, San Francisco, CA, USA) v.3.1, resulting in a dataset of five-minute segments saved as comma-separated values (CSV) files. In the second phase, the data were transformed into formats conducive to model training. This preprocessing included the incorporation of emotional metrics (stress, engagement, and attention) alongside physical environmental factors (illumination, temperature, and humidity). After removing incomplete or corrupted data segments caused by sensor artifacts or communication failures via automated filtering procedures, the dataset was organized to facilitate analysis.

The preprocessed dataset was split into training and testing sets, with 80% of the data allocated for training and the remaining 20% reserved for testing. Due to the limited size of the dataset (7500 data points), a separate validation set was not used; model training and evaluation were conducted solely on the training and test subsets. To normalize the collected sensor data with varying characteristics, we employed the MinMaxScaler (see Equation (2)) to transform all values into numbers between zero and one. This approach reduces the discrepancy between EEG data and spatial environmental data, thereby facilitating comparative analysis and enhancing the stability and efficiency of the model training process.

$$\begin{gathered} X_{\mathit{std}}={\displaystyle \frac{\left(X-X_{\mathit{min}}\right)}{\left(X_{\mathit{max}}-X_{\mathit{min}}\right)}} \\ {X}_{\mathit{scaled}}=X_{\mathit{std}}\times \left(X_{\mathit{max}}-X_{\mathit{min}}\right)+X_{\mathit{min}} \end{gathered}$$

(2)

3.3. Model Construction

3.3.1. Model Selection

RNNs effectively process data continuity over time, particularly when analyzing time series data such as EEG. Among these, the GRU regulates data flow through update and reset gates, providing a faster learning speed and greater computational efficiency with a more straightforward structure and fewer parameters than long short-term memory networks (LSTMs). The GRU unifies the traditional input and forgetting gates into a single update gate, streamlining the model structure and facilitating rapid learning and prediction [82] (see Figure 3). The ability to process data rapidly and efficiently is crucial for sensors to operate in real time. GRUs are particularly suited to fulfilling this need. GRUs demonstrate comparable performance to LSTMs while utilizing fewer parameters [104]. While other recurrent models such as LSTM were considered, GRUs were chosen based on a balance of predictive performance and computational efficiency. The choice was also aligned with the real-time operational requirements of the system, where minimizing latency and model complexity was essential.

The structural properties of GRUs are well suited to efficiently processing real-time data streams. The need for rapid response and adaptation in real-time data processing is a significant challenge. The compact structure of GRUs allows them to incorporate new input information and identify time-varying data patterns with remarkable efficiency [105,106]. Further, GRUs can effectively learn time series data by addressing the long-term dependence issue, a limitation of RNN models. These characteristics are particularly advantageous for EEG data analysis, enabling rapid data integration and training and real-time model updates to maximize the efficiency and accuracy of the training process [107,108].

3.3.2. Hyperparameter Optimization

A feature selection process is employed to ensure the optimal time series data classification to ascertain that only pertinent features are applied in the model’s training. This approach has the dual benefit of reducing computation time and enhancing the model’s performance. Hyperparameters are variables in machine learning models that the user must define. Therefore, these parameters must be correctly set to optimize the model’s performance. The objective of this study is to construct a GRU-grounded neural network model that is optimized for time series data through the process of hyperparameter tuning.

The GRU model comprises two layers, with dropout applied to each layer to prevent overfitting and enhance generalization performance [109]. We employed the rectified linear unit (ReLU) activation function in the output layer to introduce nonlinearity and permit the model to discern more intricate patterns [110,111,112]. The Adam optimizer was employed for the optimization function to automatically adjust the learning rate and momentum, facilitating rapid and efficient learning [113,114].

The efficacy of a learning model is contingent on the specific combination of hyperparameters employed. Consequently, it is imperative to identify the optimal hyperparameter settings. This study employed the Optuna framework, which utilizes the tree-structured parzen estimator algorithm to efficiently explore and evaluate optimal hyperparameter combinations during the GRU model training process [115,116]. The principal hyperparameters are the dropout rate, which determines the percentage of neurons to be randomly deactivated to prevent overfitting; the learning rate, which determines the rate at which the model learns and the size of the weight updates; the number of epochs, which is the number of times the network is shown data during the entire training process; and the batch size, which is the number of data patterns shown before updating the weights (see

Table 2. Hyperparameter settings for GRU model optimization.

Parameter	Range	Hyperparameter Values
GRUs	16, 32, 48, 64, 80, 96, 112, 128	112
Dropout rate	0.1, 0.5, step = 0.01	0.19
Learning rate	1 × 10−4, 1 × 10−2, log = True	1.8769 × 10−3
Epochs	10, 200, step = 10	140
Batch size	16, 32, 48, 64	32
Window size	60	60
Activation functions	ReLU	ReLU
Optimization function	Adaptive Moment Estimation (Adam)	Adam

).

3.3.3. Model Evaluation

To assess the efficacy of the constructed GRU-grounded time series data prediction model, we undertook a process of hyperparameter optimization. The model’s performance was evaluated by measuring the loss function to identify the optimal combination of parameters. The results were obtained through 30 iterations of training (see Figure 4).

In the data regression analysis, the statistical performance of the model was evaluated using the performance metrics of mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE), as defined in Equation (3). The MSE is a statistical measure that quantifies the discrepancy between the predicted and actual values. The RMSE is a variant of the MSE that normalizes the error by the standard deviation, providing a measure of the average squared deviation. The MAE is a further variant of the MSE calculated as the average of the absolute deviations [117,118]. The overall performance evaluation of the model indicates that the MSE is approximately 0.0012, the RMSE is 0.0348, and the MAE is 0.0214. These results suggest the model can make accurate predictions on the dataset (see

Table 3. Evaluation of the GRU model.

MSE	RMSE	MAE
0.0012163493083789945	0.03487619916281371	0.021449482068419456

)

$$\begin{gathered} \mathrm{MSE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_t-y_{tP}\right)}^2 \\ \mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(\left|y_t-y_{tP}\right|\right)}^2} \\ \mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left(\left|y_i-\widehat{y_i}\right|\right) \end{gathered}$$

(3)

A graphical representation of the changes in MSE and MAE throughout the training process (see Figure 5) demonstrates a gradual reduction in the error rate as training progresses and the discrepancy between training and test metrics diminishes. This suggests that the model avoids overfitting and improves its prediction accuracy through generalization. The constructed GRU model receives real-time EEG and assorted environmental sensor data to forecast the user’s psychological and physiological state. The system can predict several variables according to EEG and environmental data, including stress index and concentration level, which can then be correlated with performance metrics. The GRU model considers the temporal continuity of input data, identifies the effect of data from previous times on the current prediction, and incorporates this information into its predictions.

By examining the interrelationship between patterns identified in EEG data and environmental data, the model can anticipate how a user’s psychological state may evolve when subjected to specific environmental conditions. By identifying patterns indicating an increase in stress levels under specific light or temperature conditions, the model can predict stress levels and alert when these conditions are present. This predictive capability enables rapid and precise responses to real-time data streams, facilitating immediate user feedback. In conclusion, the GRU model demonstrates effective real-time EEG and environmental sensor data processing, clearly identifying time-varying patterns and high prediction accuracy. This enables the model to provide personalized solutions across various applications.

4. Results

4.1. EEG-Driven Personal Comfort Model

In this study, we constructed the comfort model to predict and recommend the illuminance level of personal comfort in real time. As illustrated in Figure 6, the comfort model is operated in real time and implemented in two main frameworks. Initially, the pretrained GRU model is updated. This is a method of retraining whereby data collected in real time is added to the existing model. As the amount of data increases, the accuracy of the model improves. Subsequently, a prediction and ranking are conducted. The model suggests customized environmental conditions and calculates relative scores to provide the optimal environment, recommending the light level to reach the desired emotional metrics.

The model’s training process is integrated through the Google Cloud Platform (GCP), which provides the technical foundation for storing and processing the collected data and continuous training and updating of the GRU model. Additionally, pretrained GRU models and scalers are located and utilized on GCP. After accumulating data collected in real time for five minutes, comprising 60 data points, a mechanism integrated into the cloud storage system is triggered, initiating the preprocessing and conversion of the data into a format suitable for training the GRU model. During the transformation process, any missing values are removed from the dataset, ensuring only the necessary information is extracted. The system operates according to data measured in real time, offering an automated and efficient method of processing and predicting real-time data.

4.2. Updating Model

The adaptability and accuracy of the GRU model can be enhanced by continuously updating its conditions through interaction with real-time data (see Figure 7).

Following data preprocessing, the GRU model is updated continuously through the application of the continuous learning process. Continuous learning is defined as “learning over time by accepting new knowledge while retaining previously learned experience.” It is also known as “incremental learning” or “lifelong learning.” This learning method ensures the continuous incorporation of new data collected in real time into the model, enabling effective adaptation to evolving data patterns. The model is updated by learning from new data, which is particularly beneficial in contexts that continuously generate data streams [119,120,121].

Traditional machine learning models are trained on large datasets, and generalized patterns are learned from those datasets. However, models of personal comfort, which consider the interaction between users’ emotional responses and their spatial environments, require the learning process to use personal experience. This involves gradually collecting and fine-tuning data over an extended period to predict comfort metrics in a given space. In particular, incorporating spatial environment reactions can enhance the model’s predictive capabilities by updating only the user’s specific experiences in previous environments.

4.3. Prediction and Ranking

The GRU model is employed to make predictions from real-time data. The model artificially extends the dataset by duplicating the last 30 data points, creating the requisite space for the prediction process to generate 30 consecutive predictions (see Figure 8).

In the prediction process, each sequence of 60 data points—which is data collected in real time—is fed into the model. The model makes predictions regarding the occurrence of a subvalue according to the most recent data point in the sequence. Predictions are then made sequentially for each subsequent data point, commencing with the sixty-first data point. This prediction method considers the continuity of the time series data, enabling the formulation of predictions at various points shortly.

A relative score is calculated for the predicted temperature, humidity, and illumination values to rank them according to the user’s preferred conditions. This allows the user to receive suggestions regarding optimal environmental conditions. The score is determined by weighting the independent variables stress (St), attention (At), and engagement (En) with EEG emotion indicators. The relative scores for the concentration indicators (At, En) are set at a higher level for more positive emotional states. Further, as each indicator approaches the maximum value of that variable, the calculated score increases. In contrast, the stress indicator (St) is calculated using the formula 1 − (maximum value), whereby a lower stress value corresponds to a higher score (see

Table 4. Prediction relative score ranking.

Ranking	Lux	Temp	Humi	St	At	En	Relative Score
1	483.14	21.96	23.27	36	64	72	0.5432
2	465.89	22.00	22.99	32	46	74	0.5297
3	455.16	22.21	22.69	32	43	70	0.5045
4	457.55	22.28	22.60	33	36	71	0.4680

).

The final score is calculated by summing the concentration metric (At, En = 0.25 each) and the stress metric (St = 0.5), which determines the weight of the final score. The output dataset is then sorted in descending order according to the final score, with each data point assigned a rank. The relative score for each performance metric is calculated, and the previously set weights are applied to derive a composite score. Subsequently, the highest-ranked data are extracted, and the output is transmitted to the Google Cloud Pub/Sub service. It is then conveyed to Node-RED, which can be employed as input to control and operate spatial environment devices.

5. Discussion

This study developed and tested a real-time EEG-driven personal comfort model focusing on illumination. The results demonstrated the feasibility of combining EEG signals with environmental sensor data and deep learning techniques to predict and recommend individualized lighting conditions that enhance cognitive efficiency and comfort. However, the relationship between diverse spatial environmental variables and emotional responses must be more thoroughly considered to enhance the model’s precision. In addition, the relatively short experimental period limited the ability to observe long-term changes and patterns. Further research could be conducted to develop more sophisticated personal comfort models by incorporating additional environmental variables and psychological metrics. This approach will ultimately make a significant contribution to the fields of intelligent architecture and human-centered environmental design.

Nevertheless, the study is significant because it demonstrates the potential of combining EEG and deep learning techniques to enable objective and precise comfort assessment and immediate environmental adjustment, thereby establishing the feasibility of a real-time personalized environment system.

Based on these findings, several implications and contributions can be highlighted. First, unlike previous studies that used subjective surveys, self-assessments, and other similar techniques to assess psychological states, this study integrates EEG and data from multiple sensors to measure an individual’s psychological comfort more accurately. This indicates that designers and researchers may utilize integrated systems comprising diverse sensors to develop tailored user experiences, particularly in classrooms, offices, and other task-oriented environments where concentration and productivity are critical. Second, this study’s findings demonstrate that integrating wearable EEG headsets and environmental sensors enables the real-time monitoring of an individual’s psychological state, facilitating the provision of environmental conditions that enhance comfort. These findings have practical implications for the real-time management of user comfort. Third, this study’s insights can be employed in developing a framework capable of monitoring and adjusting environmental conditions regularly to maintain users’ psychological stability and comfort, thereby directly supporting cognitive functions such as attention that are essential in learning and work settings. Moreover, the study’s framework can be applied to other psychological indicators, such as stress, attention, and engagement, or other environmental variables, which can contribute to more effective environmental management.

To further illustrate the applicability of this approach, a potential application scenario is presented in Figure 9. In this example, light levels are adjusted regularly to enhance concentration (En, At) and reduce stress (St). These adjustments are made every five minutes to meet the corresponding targets and continue to be adjusted afterwards by re-collecting data. This enables the implementation of customized comfort management strategies in many settings, which may ultimately lead to enhanced productivity and other beneficial outcomes.

The system operates through four sequential steps. In the initial phase of “Real-time Monitoring”, EEG signals are recorded using a headset. During the subsequent “Data Analysis” phase, the program processes key metrics such as stress, engagement, and attention. In the “Illumination Adjustment” phase, the program dynamically regulates the light level based on the analysis to enhance attention and engagement while simultaneously reducing stress. Finally, in the “Feedback Loop” phase, illumination is continuously adjusted in real time according to changes in EEG signals to maintain optimal conditions for personalized comfort and performance.

The “Getting Started” panel in the center of the screen provides guidance on system initialization. In this process, the user wears the EEG headset and establishes a connection with the program, which then begins monitoring EEG signals. The interface also informs the user that the surrounding environment can be adjusted according to the recommended light values displayed. The left graph depicts the attention, engagement, and stress levels, which are measured at 10 s intervals in real-time. The dotted line from the present moment to the right illustrates the projected levels after modifying the light to align with the recommended parameters. The graph on the right displays the changes in illumination over a five-minute interval, along with the recommended illumination five minutes hence. Additionally, it shows the current cognitive score and the predicted score, which would be altered by adjusting the recommended illumination.

The images on the right depict the illumination situation five minutes ago (A), the current illumination situation (B), and the change according to the recommended illumination (C). These configurations clearly illustrate how the system uses real-time data to provide an optimal light environment that aligns with the user’s psychological and physiological state. This enhances user concentration and stress reduction, enabling customized comfort management in various indoor environments.

6. Conclusions

This study presents an EEG-driven personal comfort model designed to predict and optimize the indoor environment’s effect on an individual’s psychological state and productivity in real time. The data were collected using three environmental variables: illuminance, temperature, and humidity. A real-time prediction and recommendation system was constructed using the GRU model. This illustrates the value of personal comfort models. The personalized comfort model provides optimal environmental conditions that reflect each individual’s preferences and psychological state, integrating various sensor data to construct a comprehensive comfort model.

The results of this study address two key hypotheses. First, while the model does not claim to demonstrate direct causal effects, it was designed to identify illumination conditions correlating with EEG-indicated reductions in stress and improvements in attention and engagement, thus supporting psychological comfort. Second, the GRU-based model effectively integrated EEG and illumination conditions to enable real-time prediction of comfort states and provide personalized environmental recommendations.

The study’s findings yield the following significant implications:

• A real-time system that integrates EEG and environmental sensor data can accurately monitor and optimize the psychological comfort of users. This provides more accurate and objective data than traditional subjective assessment methods, which can contribute to realizing personalized environments.
• A quantitative approach can be employed to assess the influence of environmental factors, such as indoor illumination, on an individual’s stress, attention, and engagement levels.
• The model can adapt effectively to changing data patterns by incorporating new data collected in real time. The findings indicate that real-time personal comfort models have the potential for practical application in various indoor settings.

These findings highlight the study’s methodological and applied contributions. Methodologically, the research advances comfort modeling by integrating EEG with multi-sensor environmental data in a GRU-based framework, enabling objective and real-time assessment of psychological states. Practically, the model demonstrates value by providing personalized recommendations for indoor environmental conditions, offering direct implications for human-centered environmental design. Overall, the study establishes the feasibility of real-time EEG-driven personalized environmental control, laying the groundwork for future intelligent and user-centered built environments.