Impact of Architecture Façade Design on Neurophysiological Stress Using Functional Near-Infrared Spectroscopy and Heart Rate Variability

Abstract

Within industrialised and emerging industrialised economies people typically spend over 95% in industrialised and emerging industrialised economies typically spend over 95% of their time in built environments, yet the neurophysiological impact of architectural design remains poorly understood. While previous studies link visual patterning to cortical activity, the cortical-to-autonomic stress pathway remains largely unexplored—a key omission given that chronic stress contributes to allostatic overload. This study examined how architectural façade design influences neurophysiological stress through a multimodal approach combining functional near-infrared spectroscopy (fNIRS) to monitor occipital cortical activity with heart rate variability (HRV) as an index of autonomic regulation. Eighteen participants provided HRV data and subjective ratings for nine systematically varied façade images characterised by their deviation with respect to natural statistics, while a subset of twelve completed fNIRS recording due to signal acquisition constraints. Façade identity significantly affected discomfort, complexity, and interest ratings ($p<0.001$), and deviation from natural statistics predicted all three measures ($p<0.01$). Façade type also showed a small but significant effect on HRV ($p=0.003$), although variance was dominated by individual differences. No stimulus-specific occipital fNIRS differences were observed. However, due to the limited sample size, further research is needed to verify this observed result. Whilst global generalisations cannot be drawn due to the small sample size, these pilot research findings indicate that façades deviating from natural image statistics influence perceptual comfort and may modestly modulate autonomic balance. However, the present data does not provide clear evidence of stimulus-specific cortical effects, which, if present, likely remain below the detection thresholds of the current protocol given its methodological constraints. This study highlights methodological hurdles and establishes a scalable framework for linking computational visual metrics to physiological responses, informing future investigations into how architectural features influence human health.

1. Introduction

Research reveals research reveals that people in industrialised nations now spend more than 95% of their waking hours inside built environments [1]. This near-constant exposure to the built environment has prompted growing scientific interest in how the design of our surroundings shapes both psychological and physiological well-being [2]. Unlike the visual environments found in nature, which, despite their diversity, share core statistical properties that the human visual system evolved to process efficiently [3,4,5], modern architectural settings often present visual patterns that diverge considerably from these natural forms.

Natural scenes typically exhibit in their luminance a $1/f$ relationship between the Fourier amplitude and spatial frequency, a structure that aligns with human perceptual mechanisms and supports efficient neural processing by minimising unnecessary neural activation [3,4,5]. When viewing such scenes, neurons in the primary visual cortex respond sparsely and selectively, thereby reducing metabolic demand, an evolutionary adaptation that promotes neural efficiency.

However, mounting evidence suggests that the visual properties of many contemporary built environments do not support efficient neural processes. Regular, high-contrast visual patterns that are a common feature of many architectural façades, especially those with spatial frequencies near three cycles per degree, are now recognised as significant sources of visual discomfort [6]. Such patterns are not merely an aesthetic concern; for approximately 10% of the general population, and an even higher proportion of neurodiverse individuals, these visual features can trigger a range of adverse responses that vary from mild discomfort to severe neurophysiological effects such as migraines and seizures [6,7].

Contemporary built environments frequently diverge from these ‘natural’ visual parameters through repetitive high-contrast visual motifs, such as geometric tilings and patterns, striped façades, or high-contrast architectural features [8,9,10]. Evidence from perceptual studies indicates that façades whose luminance patterns deviate from natural image statistics—particularly by concentrating contrast energy around mid-spatial frequencies near ∼3 cycles per degree—are associated with increased visual discomfort [10,11,12]. Complementing this, computational modelling suggests that such ùncomfortable’ images elicit dense, non-sparse responses in the primary visual cortex, consistent with mechanistic accounts of visual discomfort [13,14].

Recent studies have demonstrated that exposure to such patterns impacts the primary visual cortex. Based on a series of small-scale research studies, Le et al. [15] found that participants who were exposed to images of buildings that deviated from natural statistical properties exhibited an increased haemodynamic response and reported greater subjective discomfort. However, critical limitations persist in this research domain in relation to the impact of built environment design on human neurophysiological responses. First, most studies present visual stimuli on standard computer screens at sizes that are not able to replicate the immersive visual nature of real-world built environments. The angle subtended by built environments differs substantially between small-screen presentation and real-world viewing, potentially affecting the spatial frequencies processed and the resulting neural responses. Second, while the relationship between visual patterns and cortical activity is increasingly well-documented, a crucial question remains unanswered: does this heightened cortical activity translate into systemic stress responses?

This question carries particular significance given that chronic stress can lead to allostatic overload, a state where the repeated activation of stress response systems begins to cause wear and tear on the body, potentially triggering systemic inflammation and associated health complications [16]. From a neurobiological perspective, visual stressors may induce prolonged neural or neuroendocrine responses that disrupt physiological systems and impact homeostasis. However, this pathway from localised cortical hyperactivity to broader physiological stress responses remains under-investigated.

Study Innovation and Approach

This pilot study addressed existing knowledge gaps in relation to physiological stress responses to the built environment through an innovative methodological approach that bridges experimental control with ecological validity. This research utilised architectural façade renderings, previously generated through artificial intelligence (Midjourney) by Valentine et al. [9], that feature systematic discreet variations in the façade design. For each façade image, the predicted visual stress was quantified using ViStA (Visual Stress Analysis), a tool that estimates how an image deviates from the statistical regularities of the luminance content of natural scenes based on Penacchio and Wilkins [10].

To enhance ecological validity, façade images were projected at scale in a large meeting room, approximating the visual angle and immersive experience of viewing actual buildings. Participants’ responses were measured using a multimodal approach combining:

• Functional near-infrared spectroscopy (fNIRS) to monitor haemodynamic activity in the primary visual cortex.
• Heart rate variability (HRV) metrics derived from fNIRS signals as a proxy for allostatic responses.
• Self-reported Likert scale ratings assessing visual discomfort, complexity, and interest for each façade.

This triangulated approach enables the examination of the critical link between subjective experience, localised cortical activation, and broader autonomic nervous system responses, addressing the fundamental question of whether architectural visual stress propagates from neural processing centres to systemic physiological responses.

The implications for such research are potentially significant in relation to human health and well-being. As urbanisation continues and people spend more time in built environments, understanding how architectural design affects human physiology becomes essential for developing evidence-based design practices that minimise adverse health outcomes while promoting well-being in built environments.

2. Methodology

2.1. Participants

This study initially recruited 18 participants as a convenience sample from university networks. Participants ranged in age from 25 to 59 years ($M=32.4$; $SD=9.8$), with both male ($n=6$) and female ($n=12$) participants represented. All research participants self-reported normal or corrected-to-normal vision and had no known history of neurological issues. Participants were screened to exclude individuals with diagnosed visual or auditory impairments, current treatment for migraines or seizures, and contraindications for fNIRS or HRV monitoring.

Participants completed an intake questionnaire covering: sex, date of birth, corrective lens use, screen-related headaches (“Do you get headaches when you read on screen?”), text movement perception (“When you’ve been reading for a long time, does the text ever seem to move?”), reading aid usage, epilepsy history, severe brain injury history, substance use affecting brain function (“Do you regularly take illicit substances or medications which are known to affect brain function (such as benzodiazepine, antipsychotic, antidepressant, sedative, or hypnotic drugs)?”), native language status (“Is English your first language?”), and migraine history. Among the participants ($n=18$), 44% ($n=8$) wore glasses or contact lenses, 17% ($n=3$) reported a history of migraines, and 61% ($n=11$) were native English speakers.

While all 18 participants completed the subjective Likert scale rankings, the fNIRS data from five participants (28% of the sample) had to be discarded due to signal quality issues and data corruption. This data loss highlights significant methodological limitations inherent to the fNIRS technology, particularly when measuring activity in the primary visual cortex region at the occipital lobe [17]. Signal acquisition is notably challenging with participants who have thick hair, dark hair, or darker skin tones, as these characteristics can increase the attenuation of near-infrared light, particularly in longer channels required for the deeper probing of cortical activity [18]. These technical limitations raise important diversity, equity, and inclusion (DEI) considerations that warrant critical reflection within the neuroimaging field. The systematic exclusion of data from participants with certain physical characteristics may introduce sampling biases that affect both the generalisability of the findings and the equitable representation of diverse populations in neuroscientific research. This study acknowledges these limitations as part of a broader call for technological improvements and methodological innovations to address these challenges in future research. Within this study, detailed demographic data were anonymised to protect participant confidentiality, and as such, they are not shown in disaggregated formats. This research obtained ethics approval from the University of Cambridge Department of Engineering Ethics Committee, Division C (Application #358).

Eighteen participants completed the experimental protocol and subjective ratings; however, usable fNIRS data were obtained from twelve participants due to signal quality issues and data corruption, corresponding to an attrition rate of 28%. This reduction in effective sample size substantially constrained statistical power for detecting stimulus-specific haemodynamic effects and limited the feasibility of subgroup or moderator analyses. As this study was designed as a pilot investigation, no a priori power analysis was conducted; instead, the sample size was determined by practical constraints associated with exploratory multimodal neurophysiological measurement. The implications of sample size and attrition for inference and generalisability are addressed in the Limitations Section.

2.2. Development of Visual Stimuli

Creating Façade Variations Utilising Generative AI (Midjourney)

The architectural façade stimulus set and computational visual stress metrics employed in this study were adopted from prior work [9]. Specifically, the nine façade variations were generated using Midjourney (v6.1) through the systematic prompt-based manipulation of discrete architectural parameters—principally contrast, repetition, and spatial frequency structure—while holding base façade geometry, viewpoint, and scale constant. These stimuli were originally developed to operationalise architectural visual stress in a controlled yet architecturally plausible manner and are reused here to ensure methodological continuity and cross-study comparability (see

Table 1. Overview of nine façade configurations, illustrating progressive variations in contemporary architectural articulation.

Number	Description	Image
Image 1	A minimalist white facade shows a minimalist white façade with a horizontal composition. The ground floor features a large glass window with black frames, divided into five sections with varying widths. Interior light creates a warm glow. The left side contains a recessed entrance with a black door and a small black panel mounted on the white wall. No upper-level windows are visible in this view.
Image 2	This image introduces introduces a wooden slat screen on the upper level, creating a horizontal rectangular feature above the storefront. The wooden slats provide texture and warm tones, contrasting with the white walls. The ground floor remains consistently designed across all variations.
Image 3	A variation with presents a variation with vertical metal slats in the upper rectangular opening. These white-coloured vertical elements are placed above a dark, unlit window, creating a high degree of contrast. The ground floor remains consistently designed across all variations.
Image 4	Presents a wooden slat presents a wooden slat motif but shows it with light passing through and reflecting off the window behind. The light behind the slats adds depth and dimension to the façade, creating a variation in contrast. The ground floor remains consistently designed across all variations.
Image 5	A window arrangement shows a window arrangement with three evenly spaced rectangular windows on the upper level, contrasting with the modern storefront below.
Image 6	This image introduces introduces recessed windows and balconies with galvanised metal, vertical bar railings. The vertical elements of the balcony railings create a subtle grid pattern.
Image 7	This image features features three large arched windows on the upper level. These curves provide a contrast to the rectangular geometry below.
Image 8	Vertical metal screening shows vertical metal screening elements on both the upper-level windows and a recessed balcony. The consistent use of vertical lines creates a cohesive rhythm across the façade, while the varying depths and arrangements add visual depth to the façade. The warm lighting behind the screens creates an atmospheric effect.
Image 9	This image combines arched combines arched windows with a dramatic larger arch, creating a bold geometric statement. The larger arch encompasses a lit interior space, adding depth to the façade. This variation most strongly juxtaposes curved and rectangular forms.

).

2.3. Visual Stress Analysis (ViStA) Methodology

Predicted visual stress for each façade was quantified using the same ViStA (Visual Stress Analysis) framework reported in Valentine et al. [9], which extends the method introduced by Penacchio and Wilkins [10]. ViStA estimates the extent to which an image’s luminance structure deviates from the statistical regularities of natural scenes by analysing its spatial frequency content relative to a natural image reference distribution. The method decomposes images into overlapping fovea-sized tiles, computes Fourier-domain residuals relative to natural image statistics, and weights these residuals according to human contrast sensitivity, yielding a scalar image-level metric that has been shown to predict visual discomfort. In the present study, ViStA was applied using the same parameters, pipeline, and implementation as in Valentine et al. [9] (see Figure 1); readers are referred to their work for a full technical description and validation of the method. Reusing this established computational framework allowed the present study to focus on extending prior computational and perceptual findings to physiological outcomes measured via fNIRS and HRV, rather than re-evaluating the visual metric itself.

2.4. Experimental Procedure

This study employed a multimodal experimental design to investigate both physiological and subjective responses to architectural façades. The protocols utilised balanced methodological rigour with ecological validity while capturing both neurological and cardiovascular reactions to the visual stimuli examined within this study.

2.4.1. Measurement Instrumentation

An Artinis Brite23 functional near-infrared spectroscopy (fNIRS) system (Artinis Medical Systems, Elst, The Netherlands) served as the primary physiological measurement device. This system utilises 10 transmitters and 8 receivers, creating 27 measurement channels to record the spatiotemporal characteristics of oxygenated and deoxygenated haemoglobin concentrations in the visual cortex (see Figure 2).

Heart rate variability (HRV) data was initially intended to be gathered using a Polar10 wristband. However, due to recording software issues encountered during data collection, the HRV data from the Polar10 device had to be discarded. Instead, HRV was derived using metrics from the pulsatile component of the fNIRS signal for the two primary stimuli (high- and low-visual-stress façades). This approach is outlined in the section on data processing below.

2.4.2. Presentation of Visual Stimuli

To enhance ecological validity while maintaining experimental control, façade images were projected at scale in a large meeting room using a high-resolution projector positioned 200 cm from the participant’s eye line. Each image measured 239.5 cm in width and 135.5 cm in height, subtending a horizontal visual angle of approximately 59.6° and a vertical angle of 36.8°. This ensured that each stimulus filled a substantial portion of the visual field, closely approximating the perceptual experience of viewing an actual façade from about 8–10 m away, depending on the building’s scale.

This large-format projection represents a key methodological advancement in neuroarchitectural research, which has often relied on VR or small-screen presentations that restrict the field of view and distort spatial frequency relationships. By preserving naturalistic visual angles and contrast distributions, the present approach captures the mid-range spatial frequencies (approximately 24cycles per degree) most relevant to visual discomfort and cortical hyperexcitability. In doing so, it strengthens ecological validity while maintaining experimental precision, bridging computational, physiological, and architectural domains of analysis.

2.4.3. Photometric Characteristics of Stimuli

The projected stimuli were not photometrically calibrated using a luminance meter or spectroradiometer. While stimulus geometry, viewing distance, projected image size, and relative visual angle relationships were tightly controlled across conditions, absolute luminance, contrast ratios, and colorimetric properties were determined by the projector’s native output and room lighting conditions. All façade images were presented using identical hardware settings, projection parameters, and environmental conditions to ensure internal consistency across stimuli; however, absolute photometric values were not measured or normalised. Consequently, the present study examines relative differences between stimuli rather than absolute luminance- or contrast-specific thresholds.

2.4.4. Experimental Protocol

The experimental session lasted approximately 20–25 min (excluding setup time) and followed a structured two-phase protocol (see Figure 3):

Phase 1: Baseline Measurement + Extended HRV Assessment

After fitting the physiological measurement equipment (approximately 15 min total), participants viewed a standardised 18% neutral grey background for 5 min to establish baseline physiological measures. We adopted 18% reflectance grey, a standard mid-tone in imaging, to anchor luminance near the perceptual midpoint and thus attenuate adaptation-driven bias. Participants then completed the Pattern Glare Test for 15 s to assess their sensitivity to visual stimuli, followed by a 60-s re-exposure to the 18% grey background to re-establish the baseline state.

Participants were then exposed to two architectural façade images, one with high predicted visual stress and one with low predicted visual stress. These images were presented to research participants in a pseudo-randomised order so as to control for potential sequence effects. Each façade was displayed for five minutes, during which fNIRS data were recorded continuously, though only the first 20 s was processed for analysis. Heart rate variability (HRV) was derived continuously across the full 5 min exposure period. A 60-s neutral grey background viewing period separated the two initial façade image exposures to minimise any carryover physiological effects impacting the magnitude of observed responses to each image.

Phase 2: Extended fNIRS Assessment

The protocol continued with the sequential presentation of eight additional architectural façade images, with fNIRS recorded for haemodynamics; HRV was also derivable from the fNIRS pulsatile signal but was not analysed for these brief 20-s exposures due to insufficient duration for reliable HRV estimation. Each façade image was displayed for 20 s, followed by a 30-s exposure to the 18% grey background. This careful sequencing allowed haemodynamic responses to return to baseline between stimuli while maintaining participant engagement throughout the session. The order of the visual stimulus presentation was pseudo-randomised so as to control for any potential image sequence presentation effects.

2.4.5. Subjective Assessment

Following exposure to each façade image, participants were exposed to each image a second time while verbally asked to complete a self-reported structured building evaluation questionnaire with three distinct dimensions, each rated on a 7-point Likert scale:

• Discomfort: Responding to the statement “I feel discomfort when looking at this building” on a scale ranging from 1 (No Discomfort: “I feel no discomfort at all”) to 7 (Extreme Discomfort: “I feel a significant and overwhelming level of discomfort”).
• Complexity: Assessing the statement “This building is visually complex” on a scale ranging from 1 (Extremely Low Complexity: “The building is exceptionally simple”) to 7 (Extremely High Complexity: “The building is exceptionally complex”).
• Interest: Rating the statement “This building is boring” on a scale ranging from 1 (Not Boring at All: “I find the building very engaging”) to 7 (Extremely Boring: “I find the building extremely boring”).

This comprehensive subjective assessment component allowed for triangulation between objective physiological measures and perceived experiences of each image in this study’s façade image set, providing insight into potential relationships between neurophysiological responses and the conscious appraisal of the architectural design elements examined. The experimental design’s strength lies in its multi-method approach, combining objective neurophysiological measurements with subjective experiential data while employing stimuli that balanced ecological validity with controlled variation in architectural parameters known to influence visual processing.

2.5. Preprocessing

Physiological datasets were preprocessed and analysed to quantify stimulus-evoked haemodynamic responses (fNIRS) and derive heart rate variability (HRV) as indices of cortical activity and autonomic regulation, respectively.

2.5.1. fNIRS Data Processing

The fNIRS data was processed to quantify and visualise the haemodynamic response to a series of visual stimuli presented to each participant in a random order. For each participant, channels with poor signal quality were first rejected using a Scalp Coupling Index (SCI) with a threshold of 0.8. Utilising the modified Beer–Lambert Law, the raw intensity data was then converted to changes in oxyhaemoglobin (HbO) and deoxyhaemoglobin (HbR) concentration. A fifth-order Butterworth filter was then applied with a high pass of 0.01 Hz and a low pass of 0.1 Hz to exclude signal drifts and physiological noise including the respiration rate and heart rate. Only oxygenated haemoglobin was used for further analysis as it generally has a higher signal-to-noise ratio and statistical power than deoxygenated haemoglobin [19]. To compare the magnitude of responses across conditions, the average of the last 10 s of the baseline and the subsequent stimulus was calculated for each condition and participant. Analysis was restricted to the first 10 $\mathrm{s}$ of the 5 min stimulus to focus on the initial haemodynamic response dynamics. Haemodynamic responses were baseline-corrected using a pre-stimulus baseline of 10 s and then averaged across participants for each condition. A general linear model (GLM)-based approach was not employed because the reliable estimation of trial-wise or condition-wise responses requires multiple trials per condition or a large number of total trials, both of which were not available in the present study.

Analysis was intentionally restricted to the initial haemodynamic response following stimulus onset. Although each primary façade stimulus was presented for five minutes, fNIRS analysis focused on the first 10–20 s to capture the early stimulus-locked haemodynamic response, which is the most directly attributable to visual processing demands. Longer analysis windows are more susceptible to baseline drift, slow systemic physiological fluctuations, attentional disengagement, and habituation effects, particularly in occipital recordings. To mitigate these confounds, data were high-pass-filtered and baseline-corrected using a pre-stimulus grey screen interval, and stimulus order was pseudo-randomised to minimise order and fatigue effects. The present approach reflects a conservative strategy appropriate for a pilot study with single-stimulus presentations; future work with repeated trials and hardware-synchronised timing will enable the modelling of both early and sustained haemodynamic dynamic shifts using GLM-based approaches.

2.5.2. HRV Data Processing

To calculate HRV, the fNIRS data from each participant was first filtered to a frequency range of 0.8–2.0 Hz (48–120BPM). Following this, the data was visually inspected to identify multiple channels with a clear heartbeat signal. For each participant, a peak-finding algorithm was then applied to one randomly selected channel. A second visual inspection verified that the algorithm had correctly identified the heartbeat peaks. Finally, the inter-peak intervals were extracted to a CSV file, along with the anonymised participant ID and event markers. The 10 Hz sampling rate of the device yielded a temporal accuracy of 0.1 $\mathrm{s}$ for these intervals. HRV estimates derived from the pulsatile component of the fNIRS signal should be considered as exploratory, as the 10 Hz sampling rate and optical PPG extraction do not provide clinical-grade temporal precision comparable to ECG-based measurement.

Heart rate variability was quantified using a single time-domain metric derived from inter-beat (RR) intervals extracted from the pulsatile component of the fNIRS signal. Specifically, variability was indexed as the standard deviation of normal-to-normal intervals (SDRR/SDNN) computed over the full 5 min exposure period for each of the two primary façade stimuli. Short-term HRV metrics sensitive to high-frequency variability (e.g., RMSSD) and frequency-domain indices were not computed, as the 10 Hz sampling rate and optical PPG-derived signal do not provide sufficient temporal precision for the reliable estimation of these measures. Accordingly, the HRV results are interpreted as coarse, exploratory indicators of peripheral physiological variability rather than comprehensive autonomic profiling.

3. Results

3.1. ViStA Results

Residuals were computed using the ViStA Tool (see

Table 2. This table presents a comparative overview of nine architectural façade images.

Image Number	Primary Feature Altered	Pattern Type	Average Residuals	Cycles/Degree	Greater or Less than 3 cpd
1	Blank upper wall	None	3.50 × 108	n/a
2	Horizontal slat screen	Horizontal repetition	1.70 × 109	3	equal
3	Vertical metal slats	Vertical repetition	8.30 × 108	2	less
4	Wooden slat screen	Vertical repetition	8.50 × 108	2.1	less
5	Traditional windows	Regular spacing	4.30 × 108	1.8	less
6	French balconies	Grid pattern	7.50 × 108	4.3	more
7	Arched windows	Curved repetition	5.60 × 108	n/a	n/a
8	Vertical screening	Layered vertical	8.70 × 108	7.8	more
9	Large arch	Curved hierarchy	5.20 × 108	n/a	n/a

), which breaks each image into overlapping squares and averages the deviations from natural image statistics across those tiles, following an established method [9].

3.2. Likert Scale Ratings

To assess the effect of the stimulus on discomfort, a mixed-effects model with random intercepts for participants was compared against a null model containing only the random effects. The likelihood ratio test revealed a significant improvement in model fit when stimulus was included, ${\chi }^2\left(8\right)=45.23$, $p=3.3\times {10}^{-7}$. This indicates that discomfort ratings differed significantly between stimuli. The mixed-effects model explained 20% of the variance in discomfort through the fixed effect of stimulus identity (marginal $R_m^2=0.20$) and 36% when both fixed and random effects were considered (conditional $R_c^2=0.36$), suggesting that inter-individual variability accounted for a notable proportion of the total variance. The same statistical procedure was applied to analyse the effect of stimulus identity on complexity and boredom (see Figure 4), both showing a clear effect of stimulus identity (complexity, ${\chi }^2\left(8\right)=64.74$, $p=5.4\times {10}^{-11}$, $R_m^2=0.26$, $R_c^2=0.46$; boredom, ${\chi }^2\left(8\right)=48.26$, $p=8.8\times {10}^{-8}$, $R_m^2=0.23$, $R_c^2=0.31$).

Residual correlations between the three dependent variables were estimated using a Bayesian multivariate model including participant and stimulus as random effects. Discomfort and complexity were positively correlated (posterior mean $r=0.24$, 95% confidence interval $[0.07,0.39]$), as were discomfort and boredom ($r=0.36$, 95% CI $[0.21,0.50]$), but complexity and boredom were not ($r=-0.03$, 95% CI $[-0.20,0.14]$). These estimates reflect the associations between variables after accounting for individual and stimulus-level variability.

We next assessed whether deviation from natural statistics predicted self-reported discomfort, as would be expected from the literature [10,12,14,20]. To this end, we compared a mixed-effects model with the ViStA output as fixed effect and random intercepts for participants and stimulus to a null model containing only the random effects. We found that deviation with respect to natural statistics indeed improved model fit (${\chi }^2\left(8\right)=23.89$; $p=0.0024$; $R_m^2=0.16$; $R_c^2=0.48$). Deviation from natural statistics also improved the prediction of reported façade complexity (${\chi }^2\left(8\right)=27.45$; $p=0.00060$; $R_m^2=0.18$; $R_c^2=0.62$) and boredom (${\chi }^2\left(8\right)=24.81$; $p=0.0017$; $R_m^2=0.18$; $R_c^2=0.47$).

3.3. Heart Rate Variability

To check whether there was an effect of façade on HRV, we compared a mixed-effects model with stimulus as a fixed factor and random intercepts for participants with a model only containing the random effect. The likelihood ratio test revealed a significant improvement in model fit when stimulus was included, ${\chi }^2\left(9\right)=25.40$, $p=0.0026$, even if most of the variance was accounted for by inter-individual differences ($R_m^2=0.0006$; $R_c^2=0.72$). Accordingly, façade type had a significant but small effect on HRV, with most variance driven by individual differences.

3.4. Functional Near-Infrared Spectroscopy

The effect of stimulus identity on the mean HbO and peak HbO was assessed by comparing a mixed-effects model with random intercepts for participants to a null model containing only random effects. No effect for the mean HbO (${\chi }^2\left(9\right)=7.47$; $p=0.59$) or peak HbO (${\chi }^2\left(9\right)=6.54$; $p=0.69$) was found. Furthermore, this research did not find an effect of the mean or peak HbO on self-reported discomfort (mean: $t=-0.53$, $p=0.60$; peak: $t=0.51$, $p=0.61$), boredom (mean: $t=0.18$, $p=0.86$; peak: $t=-0.515$, $p=0.61$), or complexity (mean: $t=0.74$, $p=0.46$; peak: $t=-0.85$, $p=0.40$). Taken together, these analyses did not reveal clear stimulus-specific differences in occipital haemodynamic responses; however, this absence of effect should be viewed cautiously given the limited sample size, brief exposure durations, and technical challenges associated with fNIRS measurement in the occipital region. Further research is required to assess the degree of stimulus-specific responses among diverse populations.

4. Discussion

4.1. Principal Findings and Interpretation

This pilot investigation examined the degree to which specific architectural façade features evoke measurable neurophysiological stress responses. Three principal patterns emerged from the data. These should be interpreted with the consideration of the methodological constraints and limitations of this study.

First, façade identity reliably explained the variance in subjective perceptual ratings across all three evaluated dimensions. Likelihood ratio tests demonstrated that stimulus identity significantly improved model fit for discomfort (${\chi }^2\left(8\right)=45.23$; $p=3.3\times {10}^{-7}$), complexity (${\chi }^2\left(8\right)=64.74$; $p=5.4\times {10}^{-11}$), and boredom (${\chi }^2\left(8\right)=48.26$; $p=8.8\times {10}^{-8}$). Critically, ViStA residuals, quantifying the stimulus image computational deviation from natural image statistics, predicted self-reported discomfort (${\chi }^2\left(8\right)=23.89$; $p=0.0024$), complexity (${\chi }^2\left(8\right)=27.45$; $p=0.00060$), and boredom (${\chi }^2\left(8\right)=24.81$; $p=0.0017$). This convergence between objective computational metrics and subjective phenomenological experience provides strong support for the theoretical framework linking visual statistical properties to perceptual processing efficiency. Façades with greater deviations from natural statistics, particularly those featuring concentrated energy at mid-spatial frequencies of around three cycles per degree with high contrast and regular repetition, were consistently appraised as more uncomfortable.

Second, façade identity demonstrated a very small but nonetheless statistically significant effect on heart rate variability (${\chi }^2\left(9\right)=25.40$; $p=0.0026$). However, the marginal $R^2$ was notably low ($R_m^2=0.0006$), while the conditional $R^2$ was substantial ($R_c^2=0.72$), indicating that façade type modestly influenced autonomic activity and most variance reflected between-person differences rather than stimulus-specific responses. Importantly, the exposure windows are short relative to real-world occupancy, so any autonomic impact may be underestimated, and potential cumulative or dose-dependent effects under prolonged exposure remain an open question for future work. This pattern suggests that autonomic responses to architectural visual stress may be highly individualised, potentially moderated by factors such as baseline autonomic tone, pattern glare susceptibility, migraine history, or trait anxiety. Caution should be taken in viewing these results—given the small observed effect size, in combination with the small pilot sample size, these observations require further testing and validation.

Third, contrary to our a priori hypotheses, we did not detect stimulus-specific differences in occipital haemodynamic responses as measured by fNIRS. Neither the mean HbO nor peak HbO amplitude varied systematically across façade conditions (mean HbO: ${\chi }^2\left(9\right)=7.47$, $p=0.59$; peak HbO: ${\chi }^2\left(9\right)=6.54$, $p=0.69$). Furthermore, haemodynamic measures did not predict subjective ratings of discomfort, boredom, or complexity. This null finding represents a departure from the prior literature documenting increased primary visual cortex activity in response to images deviating from natural statistics [15]. Given that HRV and subjective ratings followed the expected pattern, this anomalous null fNIRS result potentially reflects limited power and instrumentation constraints, underscoring the need for a larger sample and more sensitive instrumentation to examine potential subtle cortical effects.

Taken together, these findings support a necessarily conservative interpretation of the findings. Computational deviations from natural image statistics reliably track the conscious perceptual appraisal of architectural façades, and preliminary, though statistically weak, evidence suggests that façade visual characteristics may subtly impact autonomic balance at the temporal scales examined. However, within this protocol, this research did not establish a robust cortical haemodynamic signature of visual stress, nor did it demonstrate a clear cortical-to-autonomic propagation pathway linking early visual processing load to systemic physiological responses. This pattern argues for methodological refinement including larger, stratified samples; longer exposures; short-separation fNIRS with improved optode coupling; ECG-grade HRV; and expanded coverage of appraisal and autonomic integration regions to determine whether subtle or individual-specific cortical effects are being missed.

4.2. The Cortical–Autonomic Coupling Hypothesis Revisited

The theoretical motivation for this study is informed by emerging models in architectural neuroscience and stress physiology that posit multilevel interactions between sensory processing, cognitive–affective appraisal, and autonomic regulation. Within this context, the notion of a cortical–autonomic cascade—whereby sustained visual processing demands may, under certain conditions, influence downstream autonomic balance and contribute cumulatively to allostatic load—provides a useful conceptual lens. However, the present study does not empirically test a causal cortical–autonomic pathway, and care is required to distinguish conceptual motivation from that which the data can directly support.

The experimental design was structured to examine whether architectural façade characteristics associated with visual discomfort are accompanied by measurable cortical (occipital fNIRS) and autonomic (HRV) responses within the same experimental context. As such, this study probes co-occurring physiological responses rather than mechanistic propagation between systems. Observed HRV differences should therefore be interpreted as associated autonomic responses to architectural stimuli, not as downstream consequences of visual cortical activity per se. Demonstrating a true cortical–autonomic cascade would require designs capable of establishing temporal precedents, mediation, or directed connectivity between neural and autonomic signals, such as time-lagged analyses, causal modelling, or simultaneous measurement of intermediate regulatory regions (e.g., insula, anterior cingulate cortex). These requirements exceed the scope of the present pilot investigation.

The absence of stimulus-specific occipital haemodynamic effects further underscores this distinction. As discussed in Section 4.3, the effective fNIRS sample size ($n=12$), brief exposure durations, and known signal-to-noise limitations in occipital fNIRS substantially constrain sensitivity to small-to-moderate effects. Consequently, null fNIRS findings cannot be interpreted as evidence against cortical involvement, and they also do not provide support for a tested cortical–autonomic pathway. Within this context, this study’s contribution lies in delineating the empirical boundaries of what can be detected under current methodological constraints, rather than validating a multistage physiological cascade.

Second, any potential coupling between visual cortical processing and autonomic regulation may hinge on individual-level moderators that were measured but could not be examined systematically given this study’s limited statistical power. Susceptibility to pattern glare and visually induced discomfort varies substantially across individuals [7], with approximately 10% of the general population exhibiting pronounced sensitivity to specific spatial patterns, a proportion that is higher among individuals with migraine or photosensitive epilepsy. Although the present sample included participants with a history of migraine (17%), this study was not powered to support subgroup or moderation analyses. The substantial proportion of variance attributable to inter-individual differences in heart rate variability ($R_c^2=0.72$ compared to a marginal $R^2$ of 0.0006) further suggests that averaging across participants may obscure meaningful stimulus–response relationships present among susceptible subpopulations. Importantly, this pattern supports an interpretation of differential susceptibility and associated autonomic responses, rather than evidence for a causal cortical–autonomic cascade.

Taken together, the cortical–autonomic framework is advanced here as a conceptual and heuristic structure that motivates experimental design and future hypothesis testing, rather than as an empirically validated pathway. The present findings indicate that architectural visual statistics reliably shape subjective appraisal and are accompanied by modest peripheral physiological signals consistent with autonomic engagement under certain conditions, while cortical haemodynamic signatures remain unresolved at the group level under current constraints. Establishing whether and how visual cortical load propagates to autonomic regulation will require future studies with larger, stratified samples; extended exposure durations; calibrated stimuli; higher-resolution multimodal measurements; and analytic approaches explicitly designed to test temporal and causal relationships. Within these limits, the present study contributes by clarifying what can—and cannot—be inferred at this stage while providing a disciplined empirical foundation for subsequent investigations into the impact of architecture on stress physiology.

4.3. Limitations

4.3.1. Sample Size and Statistical Power

This exploratory pilot study obtained usable fNIRS data from 12 participants (28% attrition due to signal quality issues), substantially limiting statistical power. Given the typical effect sizes in visual discomfort research (Cohen’s $d\approx 0.3–0.5$), this sample could only detect large effects ($d\ge 0.8$). The absence of significant occipital haemodynamic differences should not be interpreted as evidence of no effect but rather as reflecting limited sensitivity. Future studies require substantially larger samples ($n\approx 80–100$), formal a priori power analyses, and stratified recruitment to enable moderator testing and establish generalisable findings.

fNIRS signal quality issues raise equity concerns, as acquisition is particularly challenging with thick or dark hair and darker skin tones, creating sampling bias that threatens generalisability and equitable representation [18,21,22]. Future work should recruit diverse populations, report participant characteristics, employ improved optode designs, and adopt processing algorithms that minimise performance disparities.

4.3.2. Technical and Methodological Constraints

Several technical limitations constrained measurement precision. We did not employ short-separation channels to remove extracerebral contamination, potentially confounding neural signals with scalp blood flow. HRV was derived from 10 Hz fNIRS-PPG signals rather than clinical-grade ECG (≥500 Hz), severely limiting temporal precision and precluding reliable frequency-domain analyses; façade identity explained negligible HRV variance (marginal $R^2=0.0006$). The 5 min exposure windows may have been too brief to capture sustained metabolic changes and approached the minimum acceptable duration for HRV estimation. Non-optimal synchronisation between stimulus and data streams complicated temporal attribution.

Photometric calibration was not performed. While identical projection settings preserved relative differences between stimuli, absolute luminance levels, contrast ratios, and spectral characteristics were not measured, constraining mechanistic interpretation. Future studies should incorporate full photometric calibration to disentangle the contributions of spatial statistics, luminance contrast, and chromatic content.

4.3.3. Construct Validity and Ecological Generalisation

Our occipital montage targeted the early visual cortex, but subjective discomfort may depend critically on unmeasured higher-order regions: the prefrontal cortex for attentional control, anterior cingulate for conflict detection, and insula for autonomic integration [23]. HRV is a coarse index influenced by factors including but not limited to respiration, posture, fitness, age, and medication. Triangulation with electrodermal activity, respiratory sinus arrhythmia from ECG-grade recording, and salivary biomarkers would strengthen inferences.

The findings pertain to static images viewed while seated in controlled laboratory conditions, substantially different from real-world architectural exposure involving movement, multisensory experience, and extended duration. Generative AI stimuli, while enabling systematic control, may attenuate physiological responses compared to in situ exposure. Large-scale projection may have diluted effect sizes if peripheral regions reduced the impact of focal high-stress features. Generalisation requires field-based validation using mobile physiological monitoring in actual urban environments.

5. Conclusions

This pilot study integrated computational visual analysis, ecologically scaled stimulus presentation, subjective perceptual ratings, and dual-modality physiological monitoring to investigate whether architectural façade patterns that deviate from natural image statistics propagate from visual processing to systemic stress responses. Three key findings emerged: First, façades with greater statistical deviation were reliably judged as more uncomfortable, more complex, and more boring, with ViStA metrics significantly predicting all three subjective dimensions. Second, façade identity showed a small but significant effect on heart rate variability, though most variance was attributable to individual differences rather than stimulus-specific responses. Third, no stimulus-specific occipital cortical haemodynamic responses were observed under the present protocol. This result, however, should be viewed in relation to the small sample size and methodological limitations.

Within the current study format, the limited sample size and technical issues encountered did not enable a statistically valid assessment of a cortical-to-autonomic stress pathway. Caution should be taken in viewing the fNIRS results within this study as the null fNIRS findings potentially reflect this study’s methodological constraints rather than evidence against cortical effects. The small effective sample size ($n=12$ after 28% participant attrition due to signal quality issues) severely limited statistical power to detect the small-to-moderate effect sizes typical of visual processing studies. Additional limitations included technical challenges in occipital signal acquisition, brief exposure durations insufficient to capture sustained responses, HRV derivation from low-sampling-rate fNIRS signals rather than clinical-grade ECG, and construct validity concerns regarding cortical regions of interest and peripheral autonomic indices.

This study’s principal contribution is methodological, demonstrating a feasible pipeline for testing architectural stimuli at scale while surfacing concrete technical barriers that must be overcome to obtain decisive physiological evidence. The limitations encountered directly inform next-generation protocols requiring upgraded sensors with short-separation channels and improved diversity-inclusive performance, larger stratified samples adequately powered for individual moderator testing, extended exposure durations with hardware-synchronised timing, multi-system physiological measurement including gold standard ECG and complementary autonomic indices, and expanded cortical coverage including higher-order appraisal regions. Three critical questions remain unresolved and warrant full-scale investigation: whether cortical hyperexcitability responses to façades with higher predicted visual stress are reliably detectable with adequate methodological rigour; whether and when such cortical responses couple to autonomic changes (particularly in susceptible individuals or under prolonged exposure); and which specific façade characteristics, viewing contexts, and individual factors most strongly drive these effects.

While this study does not yet provide conclusive evidence of a cortical-to-autonomic stress pathway, it contributes an important methodological foundation for future work at the interface of architecture, neuroscience, and public health. By demonstrating both the feasibility and current technical limits of measuring physiological responses to architectural form, this research delineates a clear roadmap for the next generation of studies capable of testing these relationships with greater precision. The findings underscore the importance of advancing neuroarchitectural research through larger, more diverse samples and through the integration of higher-resolution multimodal measurement approaches. The limited sample size notwithstanding, the pilot study results provide preliminary support for the assessment that architectural visual statistics can shape perceptual and physiological experience while emphasising that the field is still in the early stages in establishing the causal mechanisms through which built form influences human health.