Predicting Architectural Space Preferences Using EEG-Based Emotion Analysis: A CNN-LSTM Approach

Abstract

Architectural elements—such as shapes, colors, and lighting—significantly influence how users emotionally respond to spaces. This study addresses the challenge of capturing unconscious and rapid emotional responses by employing a 32-channel electroencephalography (EEG) approach with 40 participants, who viewed multiple images of architectural spaces while real-time brain activity was recorded. Event-related potential (ERP) analysis focusing on N100, N200, P300, and late positive potential confirmed reliable differences in neural signals between preferred and non-preferred stimuli. Two convolutional neural network long short-term memory deep learning models were trained on the EEG data: one using all the ERP segments, and the other focusing on statistically significant ERP features. The first model achieved a high recall but a relatively lower precision, while the second improved accuracy and precision at the expense of recall. These findings suggest real-time, objective measures of users’ emotional responses can inform early-stage architectural design and reduce reliance on subjective evaluations. By integrating EEG-based insights into smart architecture or virtual reality simulations, designers may optimize building features to align with user preferences and well-being, contributing to the development of effective and user-centric built environments.

1. Introduction

Spending prolonged periods in indoor environments can have an effect on people’s emotional stability, creativity, and stress levels, which are heavily dependent on the physical and environmental elements of architectural spaces [1,2,3,4]. However, traditional design processes often rely on intuition and subjective judgments to gauge users’ potential emotional experiences; thus, they cannot accurately capture unconscious and immediate emotional states [4,5]. Methods such as interviews and surveys commonly used in prior research focus mainly on conscious and verbally expressed emotions but do not fully capture subtle, instantaneous affective responses elicited by spatial stimuli [6,7]. To overcome these challenges, neuroscientific approaches have been recently introduced. In particular, electroencephalography (EEG) offers high temporal resolution and noninvasive measurement capabilities, enabling the real-time evaluation of users’ unconscious affective and cognitive states [8,9].

Among the various EEG analysis techniques, event-related potential (ERP) analysis is a precise and powerful tool for measuring immediate emotional responses to specific stimuli [7,10,11,12]. According to neuroscientific research, ERP components N200, P300, and late positive potential (LPP) are closely associated with emotional valence to stimuli (e.g., preference, positivity, or negativity). When users like or dislike a given spatial stimulus, measurable differences in these ERP components can be observed. Further, recent artificial intelligence research shows that convolutional neural network (CNN) long short-term memory (LSTM) models, which capture both the spatial and the temporal properties of EEG data, can effectively learn subtle neural patterns to improve the accuracy of emotion classification [13]. Therefore, we hypothesize the following:

• EEG patterns, particularly ERP components (e.g., N200, P300, and LPP), will significantly differ between preferred and non-preferred stimuli.
• A deep learning model trained on EEG signals can effectively detect these distinctions in ERP components, thereby achieving a high accuracy in classifying affective responses (i.e., preferred vs non-preferred) to architectural spaces.

By quantitatively measuring users’ affective experiences in architectural environments using EEG and ERP analysis and integrating these findings with deep learning techniques, we propose a methodology for personalized design. This approach is expected to overcome the limitations of traditional architectural design methods while contributing to a more scientific and objective process for affective architectural design.

This paper briefly reviews the previous research on the relationships between architectural spaces, emotions, and preferences and the theoretical background of EEG-based emotion analysis. It then details the procedures used to preprocess and analyze the EEG data and design and evaluate the model. Through these steps, we aim to establish a process for quantitatively understanding users’ emotional responses to architectural spaces and building a foundational dataset for personalized space design. Ultimately, we show the potential of neurophysiological approaches in architectural practice for incorporating user experiences more deeply, contributing to developments in affective architecture and real-time adaptive space design.

2. Literature Review

2.1. Physical Characteristics of Architectural Spaces and Their Influence on Emotions and Preferences

Several studies have confirmed that the physical attributes of an architectural space significantly affect users’ psychological responses and spatial preferences (see

Table 1. Key architectural elements and their emotional and preference effects.

Architectural Element	Characteristic	Emotional Effect	Preference Effect
Spatial form [15,16,17,18,19,20]	Curved	Provides psychological stability	Increases positive evaluation
	Excessive curves	Can cause stress	Reduces preference
	Straight lines	Conveys an efficient and organized impression	Preferred in certain contexts
Lighting [21]	Natural lighting	Relieves stress and fosters positive emotions	Increases preference
	Artificial lighting	May create emotional distance	Sometimes lowers preference
Ceiling height [22]	High ceiling	Engenders openness and creativity	Enhances space preference
	Low ceiling	Promotes concentration	Preferred for certain tasks
Color [14,15]	Warm tones	Provide psychological comfort	Boost preference generally
	Cool tones	Convey a sense of distance	Preference depends on context
Materials [14]	Natural materials	Impart warmth and stability	Yield a higher preference
	Artificial materials	Create emotional distance	Tend to lower preference
Sound [14,22]	Proper acoustic design	Offers a comfortable, stable atmosphere	Improves user experience
	Excessive noise	Generates unpleasant feelings	Decreases preference

) [14,15]. Regarding spatial form, curved structures can provide psychological comfort and lead to more positive evaluations, while an overabundance of curves sometimes creates visual confusion or stress [16,17,18,19,20]. As for lighting, natural daylight tends to reduce stress and foster positive emotional states, thereby enhancing preference, whereas artificial lighting can elicit a sense of emotional distance in certain cases [21]. High ceilings can increase openness and creativity, whereas lower ceilings may enhance concentration but risk inducing a sense of confinement [22]. Other design elements—such as color, materials, and acoustics—also deliver diverse psychological experiences to users and serve as key factors in shaping overall spatial preference [14]. For example, warm colors and natural materials often foster psychological comfort and increase preference, while cooler tones or synthetic materials may create a sense of detachment.

Such emotional responses triggered by various physical elements have a direct impact on the assessment of buildings and act more broadly as a crucial indicator of how well a space might be utilized or succeed.

These studies have relied on subjective surveys to evaluate emotions or preferences regarding space. The self-reporting approach can be limited because of biases in cognition or memory distortion, underscoring the need for a more objective, quantifiable approach to understanding how architectural environments relate to user emotions.

To address this limitation, this study proposes leveraging EEG-based bio-signal analysis. EEG captures real-time neural responses to specific architectural stimuli—such as images of spaces or virtual reality (VR) environments—thereby quantifying the user’s neurophysiological reactions. By examining ERP components, one can derive meaningful differences in emotional responses. This approach overcomes some of the shortcomings of traditional subjective evaluation methods and may help interpret systematically the causal relationship between emotional responses and the physical properties of architectural spaces.

This study aims to contribute to user-centered architectural design by measuring emotional and cognitive responses objectively and quantitatively via EEG, employing it for preference classification in architectural contexts. An approach like this can effectively integrate emotional factors—often overlooked in traditional design processes—and ultimately enhance user satisfaction and the potential success of a space.

2.2. Theoretical Background of EEG-Based Emotion Analysis

“EEG” is a noninvasive method for measuring the brain’s electrical activity in real time through electrodes placed on the scalp [23]. It offers a high temporal resolution on the scale of milliseconds, capturing momentary emotional and cognitive processes. EEG signals span multiple frequency bands, from delta (0.5–4 Hz) to gamma (30–100 Hz or higher), each reflecting various cognitive states [24,25]. Because of these features, EEG is widely applied in fields such as learning, decision-making, and emotion recognition. Among EEG-based methods, ERP analysis focuses on time-locking the EEG signal to the onset of specific stimuli, thereby isolating and quantifying the brain’s rapid responses to those stimuli [26]. By observing ERP components such as N100, N200, P300, and LPP following spatial stimuli, one can quantitatively infer how users perceive and evaluate those stimuli [27].

EEG data can be analyzed from different perspectives—time, frequency, or spatial domain—as summarized in

Table 2. Principal electroencephalography analysis methods.

Method Category	Description	Key Strength	Typical Applications
Time domain [28,29,30]	Direct examination of signal amplitude, event-related potential (ERP) peaks, or statistical features	Allows for an intuitive detection of real-time changes (e.g., ERP onset)	ERP analysis, capturing stimulus response timing
Frequency domain [31,32,33,34,35,36]	Decomposes signal into delta, theta, alpha, beta, and gamma bands	Facilitates linking specific frequency bands to cognitive/emotional states	Monitoring workload, attention, relaxation, and mental fatigue
Time–frequency domain [37,38]	Combines time and frequency analysis simultaneously	Identifies temporal evolution of abnormal signals or events (e.g., seizures)	Detailed tracking of brain oscillations in tasks or events
Spatial domain [39,40]	Examines spatial distribution across multiple electrodes (e.g., frontal and parietal)	Pinpoints which brain regions respond strongly to specific stimuli or tasks	Generating topographical maps, localizing task-specific activation
Nonlinear dynamics [41,42]	Assesses signal complexity (e.g., entropy and fractal dimension)	Useful for uncovering unpredictable/subtle emotional or cognitive patterns	Advanced stress or affective state detection
Functional connectivity [43,44,45]	Investigates inter-regional interactions via graph theory or correlations	Clarifies how multiple cortical areas cooperate under specific conditions	Understanding large-scale network activation in tasks

. For instance, a time-domain analysis is well suited for ERP assessments, while a frequency-domain analysis uses specific bands (e.g., alpha and beta) to gauge the cognitive load or emotional state [28,29,30,31,32,33,34,35,36].

In the architectural domain, ERP analysis has emerged as a powerful technique for quantifying psychological responses to specific spatial elements. For example, N100 and N200 reflect the recognition of stimuli or the degree of attentional shift, P300 is associated with positive emotions and decision-making, and LPP represents sustained emotional processing over time.

Table 3. Major event-related potential components.

ERP Component	Typical Time Window (ms)	Main Cognitive Function	Architectural Design Relevance
N100	~100	Initial attention, stimulus detection	Evaluates whether design elements effectively capture attention [46]
N200	~200	Negative emotion, conflict detection	Helps refine designs to avoid excessive/ conflicting visual stimuli [47]
P300	~300	Positive emotion, decision-making	Assesses user satisfaction and positive valuation of design aspects [48]
Late positive potential	>400	Sustained positive emotional response	Identifies design elements that enhance prolonged well-being/comfort [49]
Positive slow wave	>500	Prolonged emotional engagement	Determines whether emotionally significant elements maintain interest [50]

summarizes the major ERP components, their typical time ranges, main cognitive functions, and their relevance to architectural design [46,47,48,49,50,51].

Because EEG provides real-time emotional analysis and more objective data than self-reported measures, its potential has been recognized in diverse applications such as human–computer interactions and brain–computer interfaces [52,53]. For instance, EEG in gaming or VR environments can detect real-time changes in user state, adjusting interfaces or content dynamically [54,55,56]. The wide range of frequency information (delta to gamma) also enables the simultaneous interpretation of various states like stress, relaxation, or concentration [24,27,57,58].

However, EEG is sensitive to noise from eye blinks, electromyographic signals, and electrical interference, and its relatively low spatial resolution complicates pinpointing localized activations in the brain [52,53]. Moreover, even when exposed to identical stimuli, EEG patterns can vary greatly among individuals, making it challenging to construct generalized models [59]. To overcome these problems, sophisticated preprocessing methods such as independent component analysis or empirical mode decomposition are needed, possibly in tandem with multi-modality approaches (EEG plus functional magnetic resonance imaging) and large-scale data training using deep learning [60,61].

Table 4. The advantages and disadvantages of electroencephalography-based emotional analysis.

Aspect	Advantage	Limitation
Temporal resolution	Millisecond-level measurement of neural activity	Highly sensitive to artifacts from blinking, muscle movements, and electrical interference
Spatial resolution	Noninvasive scalp electrodes are relatively easy to set up	Difficult to precisely localize deeper brain regions
Objectivity	Less prone to self-report biases since they use direct brain signals	Intersubject variability can reduce generalizability
Practical application	Real-time emotion monitoring feasible (e.g., human–computer interaction, and virtual reality)	Requires complex preprocessing (e.g., independent component analysis) to ensure clean signals

summarizes the key advantages and limitations of EEG-based emotion analysis.

Recently, an increasing number of empirical studies in the architectural domain have applied EEG.

Table 5. Empirical EEG-based studies in the architectural context.

Study (Year)	Main Topic	EEG Setup	Primary Conclusion
Li et al. (2020) [62]	Natural elements and user perception/ satisfaction	Beta ratio at F7, CP2, FC6, and P3	Beta ratio linked to task performance; natural settings improved comfort
Banaei et al. (2017) [19]	Emotional effects of curved interiors	Anterior cingulate cortex, theta, and 128 channels	Curved design induced a positive emotional response, though relatively short-lived
Erkan (2018) [63]	Ceiling height and its effect on cognition/ wayfinding	F3/F4, Fz (alpha, beta, delta, and theta)	Higher ceilings correlated with better cognition, reduced beta activity
Marín-Morales et al. (2019) [64]	Comparing emotional responses to virtual versus real environments	Nine channels (alpha, beta, gamma, and theta)	VR elicited emotions similar to real settings; SVM used for classification
Zou et al. (2019) [65]	Office design elements and motivation	14 channels	Artificial neural network achieved ~85% accuracy for personal experience classification
Vaquero-Blasco et al. (2020) [66]	VR chromotherapy and stress relief	Eight channels (gamma/alpha and theta)	VR produced stress relief akin to real therapy; gamma waves served as key biomarker
Cruz-Garza et al. (2022) [67]	Room size and window placement for cognitive effects	Alpha, beta, delta, and theta (57 channels)	Larger rooms increased alpha power; SVM predicted user responses
Yeom et al. (2021) [68]	Indoor green walls for stress reduction	Relative alpha, theta, and beta (O, P)	Green walls mitigated stress; alpha wave levels rose with relaxation
Zou and Ergan (2021) [69]	Aesthetic satisfaction, resilience, anxiety	Frontal/occipital (14 channels)	EEG effectively classified emotional states; frontal area central for evaluation
Kalantari et al. (2022) [70]	Cognitive responses in VR versus identical real spaces	Delta, theta, alpha, beta, and gamma	VR elicited cognitive patterns similar to those in real environments
Liu et al. (2022) [71]	Influence of wall color on attention and memory	C3, Cz, F3, and Fz (beta, high beta)	Cool-colored classrooms boosted arousal, attention, and memory
Kalantari et al. (2021) [72]	Impact of design elements on wayfinding	Delta–gamma	Improved features enhanced spatial orientation and cognitive engagement
Bower et al. (2023) [73]	Room size and color in brain connectivity	Theta and gamma (right frontal/temporal)	Smaller rooms plus blue color, shifted gamma/theta connectivity patterns

presents representative examples of this research and their key findings.

As shown in Table 5, EEG can quantitatively measure users’ emotional and cognitive reactions in real time, providing a systematic interpretation of how design factors in an architectural environment affect human response [74]. When combined with VR or augmented reality technologies, EEG also facilitates experiments on design scenarios during early phases of construction [64,72]. Despite the field of neuro-architecture being at a relatively early stage—requiring larger sample sizes and studies across various environmental conditions—integrating EEG throughout the design process can effectively incorporate user emotional factors. This may ultimately advance architectural environments that enhance quality of life [58,75,76].

2.3. Deep Learning-Based EEG Data Analysis

“Deep learning” is a machine learning technique based on multilayer artificial neural networks, modeled after the structure of the human brain. It excels at learning complex, nonlinear data patterns and is particularly effective for unstructured data such as images, audio, and time series [77]. Since EEG is a complex biological signal, deep learning has recently garnered attention for its capacity to automatically extract meaningful features from such data. Deep learning models comprise an input layer, multiple hidden layers, and an output layer, typically trained via backpropagation algorithms. Optimizers such as stochastic gradient descent, adaptive moments of gradient (Adam), and root-mean-square propagation and regularization methods like dropout and batch normalization are used to prevent overfitting and enhance generalizability [78,79].

As outlined in

Table 6. Types and characteristics of activation functions.

Function	Equation	Output Range	Characteristics	Common Usage	Limitations
Sigmoid	$\sigma \left(x\right)={\displaystyle \frac{1}{1+e^{-x}}}$	(0, 1)	Good for binary classification; may saturate gradients	Output layer (binary)	Highly sensitive to artifacts from blinking, muscle movements, and electrical interference
Tanh	$tanh\left(x\right)=2\sigma \left(2x\right)-1$	(−1, 1)	Zero-centered; faster convergence than sigmoid	Hidden layers	Hard to precisely localize deeper brain regions
Rectified linear unit	$ReLU\left(x\right)=max\left(0,x\right)$	[0, ∞)	Simple computation, rapid convergence; can produce ‘dead’ neurons	Hidden layers	Intersubject variability can reduce generalizability
Softmax	$softmax\left(x_i\right)={\displaystyle \frac{e^{\left(x_i\right)}}{\mathit{\Sigma }\left(e^{\left(x_j\right)}\right)}}$	(0, 1) with sum = 1	Normalizes outputs into multi-class probabilities	Output layer (multi-class)	Requires complex preprocessing (e.g., independent component analysis) to ensure clean signals

, activation functions provide nonlinear transformations that help learn the complex spatiotemporal patterns in multichannel EEG data. EEG’s high temporal resolution and inter-channel spatial correlation increase the difficulty of analysis [80,81]. Through its layered architecture, deep learning can jointly learn these spatiotemporal characteristics, positioning the method as effective for EEG-based emotion recognition or cognitive state classification [61,82].

Different deep learning models are applied depending on data characteristics and analytical objectives. CNNs excel at extracting spatial features in multichannel EEG data but may be limited in modeling temporal dependencies. In contrast, recurrent neural networks and variants like LSTM are better at handling time-series continuity but can be computationally expensive [77,83,84,85]. Gated recurrent units offer a simpler structure than LSTM, reducing computational load, whereas the transformer model leverages parallel processing and attention mechanisms for large-scale data learning but demands significant computational resources [78,86].

Originally praised for their performance in image processing, CNNs have shown notable success in detecting spatial patterns across multiple EEG channels [87,88,89]. The convolutional layers in a CNN capture local features in multichannel EEG data, while pooling layers reduce spatial dimensions to mitigate noise and improve computational efficiency [77,88]. Meanwhile, LSTM, a specialized recurrent neural network architecture, addresses the challenge of long-term dependencies in sequential EEG data [85,90,91]. LSTM alleviates vanishing or exploding gradients by incorporating input, forget, and output gates, making it particularly effective for ERP interval analysis or irregular EEG pattern detection [92,93,94,95].

Recently, hybrid CNN-LSTM approaches, which combine the ability of CNNs to extract spatial features with the ability of LSTM to capture temporal dependencies, have shown a superior accuracy and generalization over single-model alternatives [96].

Table 7. A comparison of key deep learning models.

Model	Key Feature	Advantage	Limitation
Convolutional neural network (CNN)	Spatial pattern recognition	Efficient for spatial electroencephalographic data	Limited temporal pattern learning
Long short-term memory (LSTM)	Temporal sequence modeling	Captures long-term dependencies in signals	Computationally expensive
CNN-LSTM	Combined spatial–temporal learning	High accuracy and generalization	Complex architecture

briefly compares CNN, LSTM, and CNN-LSTM models, while Figure 1 visualizes an example CNN-LSTM architecture [96,97].

Empirical studies show EEG-based deep learning models excel in various tasks, including emotion recognition, motor imagery classification, seizure detection, and sleep stage classification, among others [92,99,100,101,102].

Table 8. Studies using electroencephalography data for classification tasks.

Field	Study/Studies	Model	Accuracy
Emotion recognition	Yang et al. (2017) [91]	Hierarchical network (CNN)	~88.00%
	Alhagry et al. (2017) [108]	LSTM	~85.00%
	Chao et al. (2018) [99]	Deep belief network–graph convolutional network + LSTM + support vector machine	~87.00%
	Li et al. (2017) [105]	CNN-LSTM	~90.00%
Motor imagery classification	Zhang et al. (2019) [102]	CNN-LSTM	>90.00%
	Wang et al. (2018) [94]	LSTM	82.52%
	Xu et al. (2018) [107]	Wavelet transform + CNN	85.59%
Seizure detection	Acharya et al. (2018) [109]	CNN	88.67%
	Affes et al. (2019) [110]	CNN–recurrent neural network	>97.00%
Sleep stage classification	Supratak et al. (2017) [111]	CNN + bidirectional-LSTM	74.80%
	Craik et al. (2019) [61]	CNN	~80.00%
Event-related potential detection	Lawhern et al. (2018) [88]	EEGNet (lightweight CNN)	>90.00%

summarizes key studies that achieved a circa 85–90% or higher accuracy in these tasks [91,93,99,103,104,105,106]. CNN-LSTM hybrid models have recorded over 90% accuracy in emotion recognition, significantly expanding the potential of EEG-based brain–computer interface applications [102,107].

Building on these previous works, the present study applies a CNN-LSTM hybrid model to EEG emotion analysis for classifying users’ emotional preference toward architectural stimuli. Our goal is to achieve a high classification accuracy and robust generalization, while simplifying complex preprocessing steps.

3. Materials and Methods

3.1. Experimental Setup

EEG data were used to quantitatively analyze user preferences and emotional responses to architectural stimuli. We employed an Enobio 32 (Neuroelectrics, Barcelona, Spain)—a 32-channel wireless EEG system—and NIC software (NIC v2.1.3.11) for precise, reproducible data collection. Figure 2 shows the EEG device with channel positions and the NIC software platform.

The Enobio 32 noninvasively measures high-density neural signals at a 24-bit resolution [112,113]. It supports a 500 Hz sampling rate and covers key EEG frequency bands (delta, theta, alpha, beta, and gamma) [113,114]. Its electrode placement follows the international 10–20 standard, which includes frontal (Fz, F3, and F4), parietal (Pz, P3, and P4), and occipital regions, all closely tied to emotion and cognition [7,13,112,115].

The NIC software provides real-time EEG visualization, time–frequency analysis, and 3D brainwave mapping [112,114,116,117]. Data are stored in multiple formats (e.g., CSV and EDF). A 60 Hz line filter and visual bandpass filtering from 1 to 50 Hz were applied to minimize noise, focusing especially on alpha and beta bands, which are strongly associated with emotional responses [26,118,119,120]. Additionally, the wireless setup lets participants move more naturally, reducing external signal interference and user anxiety, thus improving data reliability [112,121].

For architectural stimuli, we used Python 3.9 and the google-images-search library (“https://pypi.org/project/Google-Images-Search/ (accessed on 25 January 2025)”) to collect images of 25 architectural styles (e.g., modern, contemporary, and rustic) [122]. All images had a minimum resolution of 1920 × 1080 pixels and were standardized for brightness, color, and contrast to minimize visual discrepancies [123,124].

During the experiment, each image was presented in a random order on a monitor for 5 s, followed by a 2 s rest interval to reduce participant fatigue and clearly capture EEG responses (ERP) [13,125]. After viewing each image, participants indicated their response—like (1), neutral (2), or dislike (3)—via a keyboard input, which was synchronized with the EEG signal [126].

The NIC software’s event markers align image-presentation timing and participant-response timing with the EEG signal [127,128]. This enables precise mapping between each image’s preference rating and the corresponding EEG pattern. EEG and event data were stored in various file formats (.easy, .edf, and .nedf), ensuring a broad compatibility with different analytical tools. The header files contain participant IDs, the experiment date and time, channel information, and event information that precisely matches stimulus onsets and responses.

3.2. Pilot Experiment

Preliminary Validation

Before conducting the main experiment, we performed a pilot experiment with eight architecture majors to verify the suitability of stimulus presentation and data analysis procedures. Participants watched multiple architectural images while EEG was recorded, focusing in particular on ERP components (N100/N200, P300, and LPP) [129,130]. Figure 3 illustrates example ERP patterns in parietal and occipital channels, confirming the experimental design was appropriate for measuring emotional responses.

From the pilot experiment, parietal channels showed more negative amplitudes (N100/N200 range) and sustained positive amplitude in the LPP range for ‘like’ (preferred) stimuli, indicating robust initial sensory processing and prolonged emotional engagement. In contrast, ‘dislike’ (non-preferred) stimuli elicited a higher positive amplitude near the P300 range, suggesting a deeper cognitive processing of negative stimuli. Similar trends were observed in occipital channels, where preferred stimuli showed strong negative amplitudes in the 0.1–0.2 s interval, and non-preferred stimuli displayed positive amplitude increases around P300. These ERP differences reflect variations in visual and cognitive processing triggered by the emotional value of the stimuli.

Table 9. Event-related potential (ERP) responses by stimulus type and brain region.

Stimulus Type	ERP Window	Brain Region	Amplitude Tendency	Interpretation
Like	N100/N200	Parietal	Strong negative amplitude	Reflects a rapid and robust neural reaction to the initial stimulus
		Occipital	Strong negative amplitude	Denotes intense visual processing for a preferred stimulus
	P300	Parietal	Low positive amplitude	Indicates minimal cognitive resource allocation
		Occipital	Low positive amplitude	Suggests for reducing cognitive load for positive stimuli
	Late positive potential (LPP)	Parietal	Sustained positive amplitude increase	Demonstrates prolonged maintenance of a positive emotional response
		Occipital	Sustained positive amplitude increase	Shows an enduring reinforcement of a favorable emotional state
Neutral	N100/N200	Parietal	Weak negative amplitude	Suggests only a modest initial response to the sensory stimulus
		Occipital	Weak negative amplitude	Indicates a lower-level response in early visual processing
	P300	Parietal	Moderate positive amplitude	Reflects limited cognitive engagement
		Occipital	Moderate positive amplitude increases	Indicates moderate cognitive processing
	LPP	Parietal	Moderate positive amplitude increases	Represents a tempered level of emotional maintenance
		Occipital	Moderate positive amplitude increases	Suggests a moderate degree of emotional carryover
Dislike	N100/N200	Parietal	Low negative amplitude	Implies reduced priority in initial stimulus handling
		Occipital	Low negative amplitude	Suggests minimal resource use in early visual processing for disliked stimuli
	P300	Parietal	High positive amplitude	Indicates elevated cognitive effort and resource allocation for negative stimuli
		Occipital	High positive amplitude increases	Points to an intensified cognitive processing of unfavorable stimuli
	LPP	Parietal	Negative amplitude	Shows a rapid diminution of a negative emotional reaction
		Occipital	Negative amplitude	Demonstrates a reduction or disappearance of an adverse emotional response

summarizes the observed ERP features and interpretations for each stimulus type (like, neutral, and dislike).

The pilot results indicate that this research protocol is indeed suitable for exploring emotional responses; stable EEG patterns were obtained [131,132,133,134]. Building on these insights, we further investigate the potential to predict user preference and analyze architectural elements that induce affective responses in the main experiment.

3.3. Experiment Design and Procedure

3.3.1. Participants

The main experiment involved 40 participants (20 males and 20 females, average age 23.98 years), including architecture students and practitioners. A G*Power analysis indicated an optimal sample size of around 32–34, but 40 participants were recruited to accommodate potential dropouts [135,136,137,138,139,140,141,142,143].

Table 10. The age distribution of the participants.

Age Range (Years)	No. Participants
18–20	8
21–30	30
31–35	1
≥36	1
Total	40

summarizes the age distribution of the participants. Selection criteria accounted for the influence of architectural knowledge on spatial emotion evaluation, and participants reported no neurological or psychiatric conditions. All participants had normal or corrected-to-normal vision and lifestyle factors that might affect EEG signals (e.g., excessive alcohol/caffeine consumption and smoking) were also controlled [26,135,136,137,138,139]. This study was approved by the Institutional Review Board(IRB) at Hanyang University (HYUIRB-202504-006), and all participants provided written informed consent [144].

3.3.2. Stimuli Selection and Categorization

To evaluate neural responses to various architectural styles, we initially referred to 25 distinct design styles proposed in previous research [145]. These styles were chosen based on (1) diversity (covering cultural, regional, and temporal variations), (2) comparability (visually distinctive characteristics enabling comparative analysis), and (3) popularity and recognition. For each of the 25 styles, we used Google Images API (SerpApi) to crawl approximately 10 candidate images per style, resulting in an initial pool of 250 images.

To ensure the visual consistency and reliability of EEG data, we applied the following screening procedures:

• Only high-resolution images of at least 2 megapixels were selected to maintain visual clarity.
• Brightness, luminance, and contrast were standardized to eliminate unnecessary visual biases.
• Images were presented in a random order to mitigate learning effects.
• Duplicate or low-quality images were excluded to enhance data reliability.

Additionally, each image was verified by two architecture experts to ensure it accurately represented its intended category (e.g., visible eco-friendly elements for Sustainability, deteriorated facades for Urban Condition). Any image that did not clearly exhibit category-specific attributes or that showed irrelevant content was removed.

As a result, we finalized 100 images (four images per style on average) for the EEG experiment. However, for illustrative clarity, Figure 4 shows eight representative styles—Sustainability, Urban Condition, Luxury, Minimalist, Materiality, Historical Style, Decrepit Buildings, and Modern and Contemporary—along with typical features characteristic of each. By covering a broad spectrum of aesthetic and functional attributes, these styles can elicit varying degrees of liking or disliking among users, making them well suited for EEG-based affective analysis.

3.3.3. Experimental Setup and Environment

Experiments took place in Room 410 of the Design Computing Lab at Hanyang University’s Science and Technology building. The lab space measures approximately 7620 × 10,430 × 2640 mm. A booth measuring 800 × 1200 × 2000 mm was placed at the center, with interior surfaces finished in aluminum-coated film to provide uniform lighting and block external electromagnetic interference [53,146,147]. The monitor (1920 × 1080 pixels resolution) was positioned 60 cm from participants (Figure 5).

3.3.4. Experimental Procedure

Participants were seated in a noise-controlled environment. Prior to the experiment, participants were instructed to, as follows:

• Avoid excessive movement to minimize EEG artifacts.
• Maintain a visual fixation on the screen during the stimulus presentation.
• Rate each image based on their intuitive emotional response.

To establish baseline resting-state brain activity, participants first gazed at a black screen for 60 s while EEG signals were recorded. Subsequently, architectural images were presented sequentially, and EEG responses were continuously recorded. Each trial proceeded as follows (see Figure 6):

• Stimulus Presentation (t = 0): an architectural image was presented for 5 s.
• Participants rated the image using a keyboard, selecting one of three options (1: like; 2: neutral; or 3: dislike).
• An interval of 2 s was introduced between each stimulus to prevent neural carryover effects.
• The trial sequence was repeated 100 times for a total duration of approximately 12 min.

The EEG data were stored in multiple formats (.easy, .edf, .nedf, and .info). Event files recorded the timestamp for image presentation, response inputs, and the preference code (i.e., 1: like, 2: neutral, and 3: dislike).

4. Data Analysis and Model Development

4.1. EEG Data Preprocessing

To improve signal quality and ensure accurate analysis, we conducted a multistep preprocessing procedure on the EEG data [148,149]. This included (1) filtering and artifact removal, (2) channel selection and re-referencing, and (3) final dataset construction, as shown in Figure 7.

The initial dataset comprised 4000 EEG trials collected from 40 participants, with each participant evaluating 100 architectural images. EEG signals were recorded at a sampling rate of 500 Hz using the 10–20 electrode system, with a focus on the parietal (Pz, P3, and P4) and occipital (Oz, O1, and O2) regions. Each trial included a 1-s baseline (−200 ms to 0 ms) followed by a 5-s stimulus response window (0–5 s).

Firstly, we applied a 1–30 Hz bandpass filter to preserve frequency bands particularly relevant to emotional and cognitive processing (notably, alpha [8–12 Hz] and beta [12–30 Hz]), while removing low- and high-frequency noise [150,151].

A finite impulse response filter (order = 2048, Hamming window) was first applied to band pass the EEG signals within the 1–30 Hz range. This was aimed at preserving the frequency bands most relevant to emotional and cognitive processing—namely, the alpha (8–12 Hz) and beta (12–30 Hz) bands—while attenuating low- and high-frequency noise. To eliminate power line interference, a fourth-order Butterworth notch filter was applied at 60 Hz.

Next, we conducted an independent component analysis to identify and remove physiological artifacts such as eye blinks, muscle movements, and cardiac signals. Next, the ‘pop_runica’ function in the EEGLAB toolbox was used to apply the extended infomax algorithm (‘extended’, 1), with default EEGLAB parameters including a maximum iteration count of 512 and an initial learning rate of 0.001 [116,152].

Following the independent component analysis decomposition, the ICLabel plugin was used to automatically classify each independent component into categories such as ‘brain’, ‘eye’, and ‘muscle’. Components with a probability of 0.9 or higher in the ‘eye’ or ‘muscle’ categories were marked for removal. This criterion was implemented using the ‘pop_icflag’ function, with the threshold set as [NaN NaN; 0.9 1; 0.9 1; NaN NaN; …], ensuring the preservation of neural signals while selectively removing artifact-related components.

In our analysis, we focused on parietal (Pz, P3, and P4) and occipital (e.g., Oz and O2) channels. Parietal channels are closely linked to spatial cognition, attention, and emotion processing—particularly useful for examining P300—while occipital channels are central to visual information processing [26,153,154,155].

We then applied a common average reference across channels to reduce variance and enhance dataset consistency [156,157,158]. This yielded an improved interpretability in ERP analyses [158,159,160].

We performed common average referencing by subtracting the average signal across all channels from each individual channel. This method reduces the common-mode noise shared among electrodes and mitigates reference bias between channels [161,162]. Specifically, common average referencing suppresses unnecessary global fluctuations (e.g., non-neural noise uniformly distributed across electrodes), thereby enhancing the relative differences between channels and improving the interpretability of ERP and other signal analyses.

EEG data are originally recorded with an initial reference electrode (typically represented as a zero-filled channel). By including this initial reference channel in the averaging process and subsequently removing it after re-referencing, the full-rank status of the data can be preserved. This procedure is consistent with recommendations from Makoto’s preprocessing pipeline and the EEGLAB tutorial and is crucial for preventing rank deficiency issues such as the generation of ghost independent components [156,163,164].

Following preprocessing, artifacts such as motion-induced noise and excessive eye blinks were removed, resulting in 3200 usable trials (80% retention rate). We then segmented the EEG signals into 5-s epochs corresponding to each stimulus and labeled them either “preference” (preferred), “neutral,” or “non-preference” (disliked) before saving the data in the MATLAB (.mat) format.

Given that our primary focus was on differentiating clear preference responses, trials labeled “neutral” were excluded from the final dataset, ensuring a balanced classification between preference and non-preference categories.

Table 11. The final dataset.

Data Type	No. Channels	Time Interval	No. Stimuli
Raw electroencephalograph trials	10	5 s	4000
Post artifact removal	10	5 s	3200
Preference data	10	5 s	1121
Non-preference data	10	5 s	814

summarizes the final dataset configuration.

These preprocessing steps formed the foundation for exploring differences in neural signals between preference and non-preference stimuli, enabling a robust analysis of how users’ emotional responses varied with architectural images. The final dataset was central to the subsequent deep learning model development.

4.2. ERP Component Analysis

4.2.1. Heatmap Visualization

Before building our model, we visualized brainwave changes in the parietal and occipital regions for preference and non-preference stimuli as heatmaps [165]. Because the parietal region is crucial for emotion evaluation and attention, and the occipital region is key for the initial processing of visual stimuli, previous research has shown that observing ERP changes in these areas can provide insights into the neural correlates of emotional responses to architectural stimuli [166,167,168].

Figure 8 presents a heatmap of the average brain activation across all participants, illustrating the differences between preference (a) and non-preference (b) conditions. EEG signals were extracted from selected parietal (Pz, P3, and P4) and occipital (Oz, O1, and O2) channels. In the occipital area, preferred stimuli exhibited a strong positive potential, whereas non-preferred stimuli presented a broader negative potential, indicating differing neural responses to architectural stimuli.

According to Davidson [120] and related research, this pattern reflects approach tendencies for preferred stimuli (positive emotion) versus withdrawal tendencies for disliked stimuli (negative evaluation). In the occipital area, non-preference stimuli yield a pronounced negative potential, suggesting a more intensified selective attention or visual rejection response [129,168].

Table 12. Parietal and occipital patterns.

Region	Preferred Stimulus	Non-Preferred Stimulus	Interpretation
Left frontal/parietal	Strong positive activation	Widespread negativity	Reflects positivity versus aversion/avoidance
Occipital	Mild positive amplitude	Prominent negative amplitude	Efficient visual acceptance versus rejection

summarizes the observed activation patterns and corresponding interpretations for parietal and occipital regions.

4.2.2. ERP Components (N100, N200, P300, LPP)

We used MATLAB R2024b’s EEGLAB and FastICA for preprocessing, analyzing windows from −0.2 s before stimulus onset to 1.0 s after [26,169]. This interval encompasses early sensory responses (N100 and N200), resource allocation (P300), and sustained emotional processing (LPP). To reduce noise, we applied a 40-point moving average filter.

Table 13. Event-related potential component analysis windows.

Component	Time Window(s)	Key Domain	Meaning
N100	0.08–0.12	Early sensory stage	Immediate perceptual response
N200	0.18–0.35	Attention shift/mismatch detection	The reallocation of resources for unexpected stimuli
P300	0.25–0.50	Cognitive evaluation	Assigning significance to stimuli
Late positive potential	0.30–0.70	Emotional maintenance	Sustained emotional processing or valence

outlines each ERP component’s typical time range.

Figure 9 shows the average waveforms for each ERP component, derived from the EEG signals recorded by the parietal (Pz, P3, and P4) and occipital (Oz, O1, and O2) electrodes. These waveforms illustrate the neural responses under preference and non-preference conditions. Specifically, the following points summarize the main characteristics and implications of each ERP component’s waveform:

• N100 (80–120 ms): after stimulus onset, preferred stimuli exhibit an initial negative deflection (N100) followed by a positive rebound, indicating a heightened sensory attention, whereas non-preferred stimuli show a more stable negative amplitude (see Figure 9a).
• N200 (180–350 ms): preferred stimuli display a sharp positive-to-negative transition, suggesting a strong recognition of relevant stimuli, while non-preferred stimuli show a delayed negative amplitude, indicating mismatch detection (see Figure 9b).
• P300 (250–500 ms): for preferred stimuli, a robust positive peak around 300 ms reflects the processing of significant stimuli, whereas non-preferred stimuli show an initial positive deflection followed by a negative shift (see Figure 9c).
• LPP (300–700 ms): preferred stimuli maintain a continuous positive amplitude, signifying a high emotional engagement, while non-preferred stimuli remain at a negative level, indicating a weaker emotional processing (see Figure 9d).

Collectively, these results show preferred stimuli invoke stronger responses across early sensory processing, resource allocation, and emotional evaluation, whereas non-preferred stimuli elicit relatively weaker and sometimes delayed responses. This confirms that architectural images can reliably differentiate user emotion at the neural level, offering crucial insights for future model training and affective design.

To examine whether ERPs were elicited in response to architectural stimuli, paired t-tests were conducted on data collected from 40 participants, with a focus on the ERP components N100, N200, P300, and LPP. The results revealed statistically significant differences between preferred and non-preferred conditions in the early ERP components (N100 and N200). These findings suggest that early attentional responses and the detection of aversive stimuli may vary depending on the type of architectural space presented.

In contrast, no significant differences were observed in the later ERP components (P300 and LPP), possibly because of the relatively short stimulus duration or individual variability. This outcome implies that late-stage cognitive and affective processing may have been less pronounced under the experimental conditions. Future studies incorporating prolonged stimulus exposure or additional environmental factors may help clarify these late ERP responses.

In summary, the findings indicate that architectural stimuli can evoke significant differences in early ERP components, suggesting their potential utility as neural indicators of the initial cognitive processing of visual and emotional aspects of architectural environments.

4.3. Model 1: CNN-LSTM with ERP Segments

To predict emotional responses to architectural images, we built a CNN-LSTM hybrid model capable of simultaneously learning both the spatial and temporal characteristics of EEG data. This model extracts spatial features through 2D convolutional layers and learns temporal dependencies via LSTM layers.

Table 14. The convolutional neural network long short-term memory configuration.

Model	Parameter	Value	Description
Convolutional neural network (CNN)	Convolution filters	50 (Conv1), 100 (Conv2)	50 filters in the first convolutional layer and 100 filters in the second layer for spatial feature extraction
	Kernel size	3 × 3	Defines the receptive field size of each filter Smaller sizes allow for finer spatial feature learning
	Activation function	Rectified linear unit	Adds nonlinearity to the model, enabling it to learn complex features effectively
	Batch normalization	Applied	Normalizes the output of each layer to improve training stability and speed
	Pooling size	2 × 2	Specifies the filter size for max pooling, reducing spatial dimensions while preserving critical information
	Strides	2	Sets the step size of the pooling filter, reducing computational load
Long short-term memory (LSTM)	Units	40	The number of neurons in the LSTM layer, defining the dimensionality of the output
	Return sequences	True	Maintains the sequence of outputs for passing to subsequent layers (e.g., attention)
	Activation	Built-in gate functions (sigmoid, tanh)	Internal functions in the LSTM for managing and updating cell states
	Input shape	(input_shape[0] // 4, (input_shape[1] // 4) × conv_filters × 2)	Dimensions of the sequential data generated from the CNN layer’s output
	Optimizer	Adaptive moments of gradient (learning rate = 0.000203)	Optimization algorithm for training; adjusts weights to minimize loss

describes the hyperparameters and roles of the CNN and LSTM components.

Figure 10 visually depicts the CNN-LSTM model architecture.

The final output of the LSTM layer had a hidden state dimension of 40, which was passed through a fully connected layer, leading to the final classification stage. The classification layer consisted of two output neurons with a Softmax activation function, designed to compute the probabilities for the two target classes: preference and non-preference. To address the issue of class imbalance, focal loss (α = 0.25, γ = 2.0) was employed as the loss function, and the Adam optimizer was used with a learning rate of 0.000203.

We utilized Optuna to optimize key hyperparameters such as LSTM unit count, CNN filter count, dropout rate, learning rate, and batch size. After 50 trials, the optimal combination was as shown in

Table 15. The best trial hyperparameter configuration.

Long Short-Term Memory Units	Conv1 Filters	Conv2 Filters	Learning Rate	Validation Loss	F1 Score
40	50	100	0.000203692	0.849423917	0.849423917

.

By incorporating focal loss, the model emphasized the contribution of difficult-to-classify samples (e.g., non-preferred stimuli) by assigning greater loss values to these instances. The training was conducted using a mini batch size of 16 over a total of 30 epochs. Early stopping was applied to terminate training when the validation loss failed to improve over a predefined number of iterations, thereby preventing overfitting and enhancing generalization.

To address data imbalance between preference (positive) and non-preference (negative) classes, we applied focal loss, which places a greater emphasis on challenging samples. We used α = 0.25 and γ = 2.0. After 30 training epochs, the model’s performance metrics were as follows:

• Recall: 98.76% (very high, rarely missing positive/preference cases);
• Precision: 72.11% (relatively low; some non-preference cases misclassified as preference);
• F1 score: 83.36%;
• Accuracy: 77.09%.

Although precision was somewhat modest, the model’s exceptionally high recall could be advantageous for real-time emotion monitoring applications.

4.4. Model 2: CNN-LSTM with Significant Features (t-Test)

The second model focuses on extracting only those features from the ERP intervals (e.g., N100, N200, P300, and LPP) that show statistically significant differences between preference and non-preference stimuli. By removing noise and superfluous information, we feed only key variables (e.g., mean amplitude, peak, area under the curve, and peak latency) that directly contribute to distinguishing emotional responses, aiming for improved performance and learning efficiency.

Firstly, we performed t-tests on each ERP interval to determine which variables exhibited significant differences between preferred and non-preferred stimuli (p < 0.05) (

Table 16. Merged event-related potential (ERP) feature definitions and segment-specific importance.

ERP Segment	Feature	Definition and Importance
N100	Mean	The average signal value in the ERP segment, reflecting the baseline response strength to a stimulus. Higher mean values indicate stronger emotional stimuli, aiding in assessing initial sensory responses
	Peak	Highest amplitude in the ERP segment, reflecting the immediate emotional or attentional intensity. Higher peak values for N100 denote stronger initial responses, useful for evaluating response intensity
	Area under the curve (AUC)	Cumulative strength of the response over time within the ERP window. At N100, it assesses the duration of early neural responses, capturing initial reactions to emotional stimuli
N200	Mean	The average ERP amplitude reflecting baseline response strength. At N200, it indicates an attention shift or mismatch detection—key for identifying stimuli causing cognitive conflict
	AUC	Represents sustained cognitive processing over time. For N200, it evaluates the strength and duration of an individual’s response to stimuli in conflict or requiring heightened attention
P300	Peak latency	The time from stimulus onset to maximum amplitude. For P300, shorter latencies imply a higher emotional salience and stronger cognitive evaluation of stimulus importance
	AUC	The total amplitude across the P300 window. Reflects how much cognitive resource is allocated, highlighting sustained engagement with an important stimulus
Late positive potential (LPP)	Peak latency	The time taken to reach peak amplitude within the LPP window. Used to gauge the speed of emotional response, helping clarify early emotional assessments
	AUC	Cumulative amplitude from stimulus onset through the LPP period, indicating how long and how intensely emotional processing is maintained over time

).

Table 17. Significant channels and features from the t-test results.

Channel	Event-Related Potential Segment	Feature	t-Statistic	p-Value
FP1	N100	Mean	−2.468	0.0137
FP1	N100	Peak	−2.487	0.0130
FP1	N100	Area under the curve (AUC)	−2.470	0.0136
FP1	N200	Mean	−2.444	0.0146
FP1	N200	AUC	−2.444	0.0146
F3	P300	Peak latency	−2.275	0.0230
F3	LPP	Peak latency	−2.334	0.0197
F8	N100	Variance	−2.051	0.0404
Fz	P300	Variance	−2.238	0.0253
Oz	N100	Mean	2.163	0.0307
Oz	N100	Peak	2.136	0.0328
Oz	N100	AUC	2.165	0.0305
Oz	N200	Peak latency	−2.656	0.0080
O2	P300	Mean	−2.422	0.0155
O2	P300	AUC	−2.423	0.0155
O2	LPP	Mean	−2.713	0.0067
O2	LPP	Peak	−2.288	0.0222
O2	LPP	AUC	−2.712	0.0067

summarizes the t test results by electrode, and

Table 18. Selected channels and features by event-related potential segments.

Channel	Channel Name	ERP Segment	Features
0	FP1	N100	Mean, peak, area under the curve (AUC)
0	FP1	N200	Mean, peak, AUC
1	FP2	N100	Mean, peak, AUC
1	FP2	N200	Mean, peak, AUC
2	F3	P300	Peak latency, AUC
2	F3	Late positive potential (LPP)	Peak latency, AUC
3	F4	P300	Peak latency, AUC
3	F4	LPP	Peak latency, AUC
6	Fz	P300	Peak latency, AUC
6	Fz	LPP	Peak latency, AUC
9	O2	P300	Peak latency, AUC
9	O2	LPP	Peak latency, AUC

lists the final channel–feature combinations chosen based on statistical significance and neuroscientific relevance.

These selected features were processed by a CNN (Conv1D) layer to capture partial spatial patterns, followed by an LSTM layer for temporal dependency (

Table 19. Integrated convolutional neural network (CNN) and long short-term memory (LSTM) settings.

Model	Parameter	Value	Description
CNN	Filters	64	Number of filters for spatial feature extraction
	Kernel size	3 × 3	Defines the size of the convolution window
	Activation function	Rectified linear unit	Enables the learning of complex spatial features
	Padding	Same	Preserves the input size
	Pooling size	2 × 2	Reduces the dimensionality of the input feature map
LSTM	Units	50	The number of neurons in each LSTM layer
	Return sequences	True	Keeps sequential outputs for the attention mechanism
	Dropout rate	0.3	Reduces overfitting by randomly disabling neurons during training

).

We also employed SMOTEENN to address minority-class imbalance (non-preferred samples) by combining an oversampling of challenging cases with noise filtering.

The final classification layer, consistent with Model 1, consisted of two output neurons with a Softmax activation function for binary classification between preference and non-preference. The model was trained using focal loss and the Adam optimizer.

Using an 80–20 train–validation split and EarlyStopping (stopping within 50 epochs), the final model achieved the following:

• Accuracy: 81.90%;
• Precision: 85.00%;
• Recall: 70.00%;
• F1 score: 73.00%.

By focusing on statistically significant ERP features, Model 2 achieved an improved accuracy and precision. However, its recall was somewhat lower, indicating subtle positive responses may have been missed. Still, these outcomes provide a robust neuroscientific foundation for designing affective architectural spaces.

4.5. Model Evaluation and Comparison

We then compared the performances of Model 1 (ERP Segment CNN-LSTM) and Model 2 (t-test-based CNN-LSTM). Model 1, which learns comprehensive spatiotemporal ERP features, and achieved a very high recall (98.76%), rarely missing positive/preferred cases. However, its precision (72.11%) was relatively low, leading to more false positives. Meanwhile, Model 2, using only statistically significant ERP features, improved accuracy (81.90%) and precision (85.00%) but had a lower recall (70.00%).

Table 20. A comparison of advantages and disadvantages between Model 1 and Model 2.

Model	Advantages	Disadvantages
Model 1: ERP Segment CNN-LSTM	Captures comprehensive spatiotemporal features of EEG signals across all ERP segments High recall (98.76%) ensures sensitivity to emotional stimuli, minimizing missed positive responses Suitable for real-time emotion monitoring or attention shift applications	Lower precision (72.11%) due to potential overfitting on irrelevant patterns Higher computational complexity due to processing all ERP segment features
Model 2: t Test–based CNN-LSTM	Uses statistically significant features (mean, peak, area under the curve, and peak latency), reducing noise and enhancing efficiency High accuracy (81.90%) and precision (85.00%) ensure reliable positive predictions Efficient for scenarios requiring trusted emotional predictions, such as risk management systems	Lower recall (70.00%) than Model 1, leading to a higher likelihood of missing subtle positive responses May miss out on valuable spatiotemporal patterns excluded during feature selection

summarizes each model’s strengths and weaknesses. In short, Model 1’s extensive coverage of ERP segments yields a higher recall, whereas Model 2’s targeted features lead to a better accuracy and precision.

To further evaluate the generalization performance of Models 1 and 2, a fivefold cross-validation was conducted. Given the relatively limited dataset and intersubject variability in ERP signals, some variation in accuracy was observed across folds. However, Model 2 consistently outperformed Model 1 in terms of classification accuracy. This result supports the notion that extracting meaningful ERP features and feeding them into the CNN-LSTM architecture can lead to a more stable and improved classification performance.

Further, a paired t-test was conducted to statistically compare the cross-validation accuracies of Models 1 and 2. The results revealed a significant difference (t = –10.4913, p = 0.0005), indicating that the ERP feature-based approach of Model 2 is more effective for EEG-based affective classification. These findings suggest that additional data collection and further feature optimization could further enhance model performance.

These findings validate the feasibility of EEG-based emotion analysis for distinguishing preference versus non-preference in response to architectural stimuli. In practical terms, designers may choose or combine these models based on whether they prioritize sensitivity (high recall) or precision in capturing user preferences. The next section discusses potential real-world applications of these models and their significance for personalized affective architecture.

5. Discussion

This study confirms that users’ emotional responses toward and preferences for architectural stimuli can be quantitatively evaluated using EEG. We proposed a methodology based on deep learning models to objectively classify these responses. We recruited a total of 40 participants and collected their neural responses to architectural images using a 32-channel EEG device. Thorough data preprocessing (filtering, noise removal, channel selection, and re-referencing) ensured high-quality EEG data. Subsequently, ERP components (N100, N200, P300, and LPP) were analyzed to identify spatiotemporal differences in brain responses between preferred and non-preferred stimuli. The results revealed significant differences in the P300 and LPP intervals, demonstrating that an ERP-based evaluation of architectural preferences is valid, supporting Hypothesis 1.

The two CNN-LSTM deep learning models proposed in this research performed differently depending on the ERP features used. Model 1 (CNN-LSTM using the entire ERP) achieved a high recall of about 98.76%, minimizing missed detections of emotional responses, but its precision was relatively lower (72.11%). Model 2 (CNN-LSTM using t-test-selected ERP features) had an improved accuracy (81.90%) and precision (85.00%) but exhibited a lower recall (70.00%). This indicates that the entire ERP dataset can provide a broader range of emotional information but may be sensitive to noise, whereas selectively using statistically significant ERP features enhances precision but may overlook subtle neurological changes. Hence, ERP feature selection plays a crucial role in determining model performance, partially supporting Hypothesis 2.

To further validate model performance and generalizability, a fivefold cross-validation was conducted. Although both models showed some variability across folds, Model 2 consistently demonstrated a superior and more stable performance. Moreover, a paired t-test revealed a statistically significant difference in accuracy between Model 1 and Model 2 (t = −10.4913, p = 0.0005), confirming that the ERP-based feature selection may be more effective for EEG-based emotion classification and model performance can vary according to how input data are constructed.

Several important limitations should be noted. Firstly, the sample size (n = 40) is relatively small, warranting caution in generalizing the results. Larger-scale follow-up studies are needed to improve reliability. Additionally, because EEG data may not fully conform to a normal distribution, we adopted a t-test for exploratory analysis. Future research should consider nonparametric approaches (e.g., Mann–Whitney U test) or multiple comparison corrections (e.g., false discovery rate) to enhance the robustness of statistical analyses. Finally, this study employed image-based stimuli. Therefore, long-term measures in real-world or VR spaces and the integration of other physiological signals (e.g., heart rate and galvanic skin response) are warranted to ensure ecological validity.

Nevertheless, the value of the methodology proposed here is that it offers an objective approach to incorporating users’ affective responses during the early stages of architectural design. Accordingly, the findings of this study should be considered in relation to recent EEG-based research in the field of neuroarchitecture. While this study highlights the feasibility of applying EEG-based emotion analysis in architectural design, several challenges must be addressed before it can be practically applied, as follows:

• Real-world data collection: to enhance ecological validity, future researchers should conduct longer term measurements in real-life environments such as hospitals, parks, commercial complexes, and transportation facilities, where diverse contextual factors come into play [1,5].
• Multimodal bio-signal integration: coupling EEG with other physiological signals (e.g., galvanic skin response or heart rate variability) could improve the precision and real-time capability of emotion recognition [13,170].
• Model optimization and distribution: CNN-LSTM models can be computationally heavy. Techniques such as field-programmable gate array deployment, mobile graphics processing unit acceleration, multiagent systems, and ensemble methods may be used to reduce computational load and enhance field applicability [88,171,172].
• Utilizing digital twins and VR/augmented reality: building an evaluation platform that integrates EEG data into virtual environments could incorporate user feedback in the earliest design phases, significantly boosting satisfaction and efficiency.

6. Conclusions

Based on EEG and ERP analysis, the findings of this research demonstrate that users’ emotional preferences for architectural stimuli can be quantified. Notably, distinct variations between preferred and non-preferred stimuli were identified in the P300 and LPP components, suggesting that EEG-based emotion classification is neurophysiologically valid. When evaluating the CNN-LSTM deep learning models, Model 1, which utilized the entire ERP dataset, achieved a high recall but a relatively lower precision. In contrast, Model 2, which selectively incorporated only statistically significant ERP features, demonstrated a high accuracy and precision but a lower recall. Results from the fivefold cross-validation further confirmed the more consistent and superior performance of Model 2. A paired t-test indicated a statistically significant difference in accuracy between the two models, supporting the effectiveness of selective ERP feature extraction. These findings underscore the potential of EEG-based emotion classification as an objective and neurophysiologically valid evaluation tool while emphasizing the importance of identifying relevant ERP components.

Nonetheless, the study is limited by a small sample size, assumptions of data normality, and concerns regarding ecological validity. Future research should address these limitations by expanding the participant pool, gathering data in real-world settings, and integrating multiple physiological signals.

Moreover, this work highlights a promising approach to embedding user emotions and preferences into the architectural design process. As real-world validation expands, user diversity increases, and technical methodologies become more refined, these neurophysiological strategies may significantly contribute to the development of truly human-centric, affective architectural environments.