Developing an AI-Based Digital Biophilic Art Curation to Enhance Mental Health in Intelligent Buildings

Abstract

Biophilic design is a well-recognised discipline aimed at enhancing health and well-being, however, most buildings lack adequate representation of nature or nature-inspired art. Notable barriers exist such as wealth, education, and physical ability restricting people’s accessibility to nature and associated artworks. An AI-based Biophilic arts curation and personalised recommendation system were developed in this study to improve accessibility to biophilic arts. Existing Biophilic research mainly focuses on building design principles, limited research exists to examine biophilic arts and associated emotional responses. In this paper, an interdisciplinary study addresses this gap by developing metrics for Biophilic art attributes and potential emotional responses, drawing on existing Biophilic architecture attributes and PANAS items. A public survey of 200 participants was developed in this study. The survey collected art viewers’ ratings of Biophilic attributes and associated emotional responses to establish statistical correlations between Biophilic attributes and emotional responses. The statistical analysis established a positive correlation between Biophilic attributes and positive emotions. The public survey results show significant positive emotional impacts (p-value $<0.05$) after exposure to Biophilic images, supporting further research and development of the Biophilic art curation system. This digital curation system employs Computer Vision algorithms (ResNet50) to automate Biophilic art categorisation and generate personalised recommendations. This study emphasises the importance of integrating nature into built environments. It proposes that artificial intelligence could significantly enhance the categorisation and recommendation of Biophilic art, advocating for expanding Biophilic art databases for emotionally responsive art display systems, benefiting mental health, and making art more accessible.

1. Introduction

1.1. Nature Disconnectedness and Intelligent Biophilic Buildings

Current extensive urbanisation in the developing world and re-urbanisation in the developed world have resulted in uneven transformations of urban landscapes and unequal distribution of green infrastructure across socio-demographic groups. The world population is increasingly urbanised and disconnected from nature. By 2050, an estimated 70% of the world’s population will live in urban areas; urban retrofitting to reconnect with nature is an extremely challenging task [1,2,3,4]. On a global scale, the Human Spaces Survey [1] found that 58% of 7600 office workers in 16 countries had no plants in their offices. Urban dwellers typically spend over 90% of their daily time indoors. However, a recent survey reveals that one in four UK office workers reported the belief that their working environment did not support their well-being, and they were particularly unhappy with the lack of colour (80%), greenery (64%), and art (61%) in their workplaces [1]. To address the issues of nature disconnectedness, the concept of Biophilic design [5] is explored as an innovative approach that seeks to integrate natural elements into built environments to enhance human health, well-being, and productivity. It involves incorporating nature directly (e.g., plants, water, natural light) and indirectly (e.g., natural materials, colours, shapes, or arts) to create a connection between humans and nature within urban and architectural spaces. The visual aspects of spaces play a crucial role in influencing the comfort, aesthetics, and overall well-being of building users. Key visual elements include colour, lights and views of nature, and art, and natural elements may have a direct impact on mood and perception; for example, cool tones like blue and green promote calmness and relaxation, while warm tones such as red and yellow energise and stimulate activity [6], demonstrating that the right colour scheme can enhance the functionality of a space, making it more conducive to the activities intended for that environment. Researchers revealed that views of nature can improve cognitive function and mental health [6,7,8,9], indicating that views of real outdoor vistas or nature-themed artwork significantly enhance well-being by reducing stress and promoting a sense of tranquility.

This design philosophy aims to address the innate human affinity for nature, known as Biophilia [10]. The term “Biophilia” was originally introduced by Erich Fromm [10] to delineate the psychological inclination towards the preservation of life. The word Biophilia originates from the Greek word philia, which means love of. The concept of Biophilia was subsequently adopted by Edward O. Wilson and Stephen Kellert [5,7]. From an architectural design perspective, the “Patterns of Biophilic Design” as seen in Nature [7] articulates the relationships between nature, the human biology, and the design of the built environment so that one may experience the human benefits of Biophilia in the building design applications. Numerous studies in the field of environmental psychology have explored the comparative effects of natural settings versus urban environments in facilitating attention restoration, eliciting positive emotions, mitigating stress, and fostering additional health benefits [8,9,11]. It elucidated two fundamental constructs within Biophilia: an inclination towards a fascination with and affiliation to life. The efficacy of Biophilia is contingent upon the innate ability to effortlessly direct attention toward natural stimuli, fostering a profound fascination with the intricacies of the natural world. Living in Biophilic buildings can be regarded as a therapy with potential benefits for addressing health concerns and promoting various physical and mental well-being advantages [9,11]. The theory of Biophilia has been established for its ability to alleviate stress, foster creativity, and improve overall well-being. Biophilic design can reduce stress, improve creativity and clarity of thought, and improve well-being; with urbanisation an ongoing trend, these positive impacts on the natural environment are increasingly lost [12].

Biophilic building design is a well-studied area [1,5,6,7,8,9,11]. However, there is a lack of research to explore how digital technologies can enable intelligent buildings to promote effective engagement with nature through digital displays of nature [13,14]. One key feature of future intelligent biophilic buildings is the creation of multiple functions of building envelopes [12]. For example, building walls can provide interactive displays. In the last few decades, more and more digital devices have been installed in buildings. Digital displays (e.g., wall-mounted, stand-alone, or PC monitors) are very common features of modern buildings [14]. However, the potential of intelligent Biophilic buildings that utilise digital displays of nature remains largely unexplored and often overlooked in sustainable architecture design practices.

Nature can be a central focus of art, and it can be recreated through artistic expression. Human and natural interactions are complicated phenomena [8,9,11,12]. Much art inherently incorporates natural elements [15]; however, art is rarely studied in a Biophilic framework [14]. Researchers argued that art not only personalises a space but also acts as a visual focal point, providing mental relaxation and interest. Artwork featuring natural scenes can reduce anxiety and foster a calming environment [6]. Traditional aesthetics is often associated with the appreciation of art, but philosophical investigation has demonstrated that much of the aesthetic experience encompasses nature [7]. The knowledge of what one appreciates is essential for an appropriate aesthetic experience, and a scientific understanding of nature can enhance appreciation [7]. The relationship between nature and the arts is profound and symbiotic. Nature inspires art through landscapes, wildlife, and natural phenomena, while art reflects and celebrates the beauty, complexity, and essence of the natural world [15]. However, there is a significant lack of research exploring emotional responses to artwork within the framework of Biophilic theories [14]. This study introduces the concept of “Biophilic Art”, defined by two key perspectives: the Biophilic attributes embedded in the artwork and the emotional responses evoked in viewers when engaging with these pieces.

1.2. Inequality to Access Arts

Inequality in access to both arts and nature significantly affects marginalised communities, limiting their opportunities for cultural and environmental enrichment [3]. The United Nations Educational, Scientific and Cultural Organization (UNESCO) report “Culture: Urban Future” highlights that marginalised communities often face significant barriers to cultural participation [16]. It is well documented that access to the arts is unequal, with ethnic and income disparities leading to more affluent and highly educated demographics enjoying greater access [17]. Inequality of wealth and opportunity, social isolation, and mental ill-health form two of the three pillars of the Arts Council of England’s ‘Let’s Create’ strategy [18]. This is part of a broader move towards social prescription, a healthcare approach that links individuals with non-medical support in their communities to improve their overall well-being and address social, emotional, or practical needs, such as access to art. The review conducted on behalf of the National Health Service England (NHSE) [19] found compelling evidence to support the notion that engaging in various forms of artistic activities can result in many positive salutogenic effects. These benefits encompass enhanced social connections, reduced stress levels, adoption of healthier habits, and improved outcomes such as the development of skills and increased employability.

The evidence that access to the arts benefits health makes it imperative that structural barriers be removed. These barriers are not restricted to wealth and education but include physical ability. For many people, the cultural aura of art galleries is off-putting, creating a sense of intimidation and a feeling that these spaces are not for them [3]. While the expense of purchasing art for the home is prohibitive for many, others may simply lack the time to engage with art [17]. Moreover, art is a huge body of information that can be daunting to access and difficult to curate, even for experts. For these reasons, this study addresses access and organisation barriers. With the recent development of digital technologies, this study will explore an automated Biophilic arts curation system that can reduce costs and remove these accessibility barriers based on artificial intelligence technologies.

1.3. Developing AI-Based Automated Classification and Recommendation Systems for Biophilic Arts Curation

The classification of paintings is a complex and multifaceted task that can be done in various ways, often depending on the context and criteria used. Common ways of organising art are a theme, subject, iconography, medium, artistic movement, time, patronage, biography of key individuals such as artists or collectors, and geography. However, there is no existing categorisation based on Biophilic attributes, or indeed, any other health or salutogenic quality. Indeed, this aspect of the arts has been largely ignored by art historians and art academics and left to practitioners. Reference to art history or visual arts readers such as [20,21], which still reflect the core of the discipline, shows no reference to the move towards social prescription or indeed any kind of utility function for arts. However, even within this literature, there is an intimation of the power of art to make the viewer act or behave differently, or even to feel as though they are in the presence of another agency [22,23,24,25], and in a historical context, how art can be through this agency related to health [26]. It is only in more recent literature, for instance [27], tellingly a publication in a series that looks both at practice and theory and critique, that the connection between arts and health is gaining mainstream curatorial or academic attention, though clearly, emotion and the arts are intimately related [27]. A pivotal milestone in the contemporary examination of emotional reactions to art is attributed to Daniel Berlyne’s formulation of the “new experimental aesthetics”. In the early 1970s [28]. That theory explained the hedonic qualities of art by referring to arousal-modifying “collative properties” of art, such as complexity, novelty, uncertainty, and conflict. Different approaches are also proposed as appraisal theory [29]. However, there is a lack of validated scales to measure the emotional responses of the Biophilic arts.

Categorising artworks is a challenging task that involves the manual intervention of expert artists and scholars, which can not only prove to be time-consuming but also extremely costly. The advantage of using digital technology is that it can automate the process of digital arts classification and recommendation tasks with considerably less time and more efficacy. This study uses computer vision algorithms for categorisation and recommendation tasks. Computer vision is a field of artificial intelligence that tries to analyse visual data. It utilises image processing, pattern recognition, and machine learning techniques for performing tasks like image and video recognition, object detection, and scene reconstruction [30]. It is commonly accepted that the father of computer vision is Larry Roberts, who in his Ph.D. thesis [31] at MIT discussed the possibilities of extracting 3D geometrical information from 2D perspective views of blocks (polyhedra) [32]. Future researchers followed this work and studied computer vision in the context of the geometric constraints of 3D. Eventually, researchers realised it was necessary to tackle images from the real world, leading to much-needed research on low-level vision tasks like edge detection and segmentation. David Marr [33] proposed a framework that outlined a three-level approach—comprising the computational theory, algorithms and representation, and hardware implementation—to explore how the visual system interprets and represents visual information, which laid the groundwork for future research in both human visual perception and the development of computer vision algorithms. Marr’s work [33] is recognised for its interdisciplinary approach, integrating psychology, neuroscience, and computer science, and remains a foundational reference in the study of vision. Recently, deep-neural networks [20] have been used for automating the classification of artworks based on styles, genres, schools of art, and brushstrokes [21,22,34] with an accuracy ranging from 70% to 80%. Whereas categorising emotions from paintings has been attempted by very few studies [23,24,26,27], where models were developed to predict a very narrow range (three items) of emotions, e.g., whether the artwork invokes positive, negative, or neutral sentiments.

In this study, the focus group method and the state-of-the-art computer vision algorithms are utilised to identify Biophilic attributes and emotional responses of the visual arts. This study aimed to automate Biophilic and emotional categorisation and recommendation processes based on machine learning and artificial intelligence algorithms. The ultimate goal of this digital curation system is to automate the curation and display of art likely to improve mental health and well-being. A public survey was carried out to explore the AI-based therapeutic Biophilic display system. Unique contributions from combined Biophilia and arts in a built environment are explored in this study. This study explores the unique integrations of Biophilic principles and art within intelligent built environments, highlighting the potential of future intelligent health buildings.

2. Materials and Methods

This section provides a brief overview of the methodology framework as illustrated in Figure 1. The research tasks follow three stages:

• The first stage is to develop metrics for Biophilic arts attributes and emotional responses.
• The second stage involves the collation of the art dataset from a public repository and then conducting a public survey to assign labels to the painting.
• The third stage is to develop an ML algorithm to automate the categorisation process and then develop a recommendation system for the users.

2.1. Stage 1: Developing Metrics to Categorise Biophilic Attributes and Emotional Responses

Bespoke Biophilic attributes of the arts and associated emotional responses will be developed based on an interdisciplinary approach. The authors of this paper, as a multidisciplinary team of artists, architects, engineers, psychologists, and computer scientists, will identify bespoke metrics based on existing Biophilic design principles [7,8], and The Positive and Negative Affect Schedule (PANAS) [35].

Key Biophilic building design patterns have been developed by Browning [7]. Although these patterns are effective in architectural and design contexts, their direct translation to artwork analysis may not accurately capture the nuances and artistic expression inherent in Biophilic art. Novel Biophilic attributes of the arts will be developed based on the existing Biophilic design literature [1,7,8]. For the Biophilic buildings, three main categories of Biophilic attributes were developed [8]: Nature in the Space, Natural Analogues, and Nature of the Space. “Nature in the Space” involves direct interactions with elements like plants, natural light, and water, creating sensory experiences that mimic the natural world. “Natural Analogues” use materials, patterns, and colours inspired by nature, evoking a sense of the outdoors without direct contact. “Nature of the Space” explores spatial qualities, balancing open views (prospect) with sheltered areas (refuge), and stimulating curiosity through Mystery and a mild thrill through Risk/Peril. Together, these categories foster environments that enhance well-being, creativity, and comfort by engaging our innate connection to nature. The emotional responses will be assessed using a psychological scale—Positive and Negative Affect Schedule (PANAS) [35] will be adopted to assess emotional responses.

PANAS [35] is a psychological tool that measures two primary dimensions of emotional experience: Positive Affect (PA) and Negative Affect (NA). Each dimension provides insight into a person’s emotional state, capturing a range of moods and feelings. The PANAS scale is a valuable tool in research and clinical settings, providing a reliable measure of how emotional states fluctuate over time or in response to specific circumstances. It helps researchers understand the impact of emotions on behaviour, decision-making, and mental health, offering insights into individual well-being and the balance between positive and negative emotional experiences. An interdisciplinary approach is adopted to develop a dedicated set of PANAS items that will be selected for this study.

2.2. Stage 2: Develop Data and Survey to Establish Relationships Between Biophilic Arts and Emotional Responses

As there is a lack of prior research employing machine learning to categorise artworks based on Biophilic characteristics, no publicly available dataset exists for this purpose. A survey featuring carefully chosen images to initiate the data creation process was conducted. These selections were made in collaboration with experts from the fields of art, psychology, computer science, and Biophilic architectural design. The dataset was meticulously compiled from scratch, sourcing artwork from the public domain and administering multiple surveys. For investigation, a total of 872 artwork images were utilised, predominantly paintings, along with photographs of sculptures, all sourced from the public gallery of the Art Institute of Chicago with a Creative Commons Zero (CC0) licence. Creative Commons Zero (CC0) licences are referred to as the “no rights reserved” option. Content with a CC0 licence is equivalent to content in the public domain.

Utilising the Biophilic and emotional metrics developed in the previous stage 1, participants were recruited to rate the Biophilic attributes and emotional responses associated with a random subset of 20 artworks with cc0 licenses. Fifty distinct surveys were designed where participants were presented with sets of 20 images of artworks and asked to identify the predominant Biophilic characteristic along with their emotional reaction. Participants encountered different artworks for each question, while the response options remained consistent across all surveys. A total of 200 participants engaged in these surveys; their responses to each image were documented, and the relationship between Biophilic attributes and emotional reactions was analysed. Utilising this correlation, a straightforward recommendation system capable of suggesting Biophilic artworks to users based on various criteria, such as emotional state and time of day, was developed.

2.3. Stage 3: Machine Learning Techniques for Classification and Recommendation

To train a machine learning classifier effectively, having a diverse dataset is paramount. To increase the size of the dataset, a technique was formulated in which a pre-trained ResNet50 [36] model was used to extract features from images that garnered the highest responses in the surveys for each of the 14 Biophilic categories. Subsequently, several artworks from the public gallery of the Art Institute of Chicago were used for the next stage of annotation. Using the same ResNet model, features from these new images were extracted, and a cosine similarity was applied to measure their likeness to images representing the 14 Biophilic categories. The new images were then assigned the label of the Biophilic category corresponding to the most similar image. Through this process, a fully annotated synthetic dataset of 10,000 images was created, where the annotations represented the predominant Biophilic feature.

To train a Biophilic classifier, the dataset was split into training and validation sets in a ratio of 7:3 to be used for training and validating the efficacy of the classifier model. To further diversify the data, several data augmentation techniques were used, like normalisation, random clipping, the addition of noise, flipping, and shearing. These techniques help in improving the model’s generalisation capability. Next, a machine learning model was trained for the classification of images based on their dominant Biophilic features. Several popular classification models were used, such as the pre-trained ResNet50 [36], Swin Transformers [37], and DEIT [38] model. A ResNet50 model is one of the most popular deep learning architectures for classification problems that utilises several residual blocks of convolutions, where the input of the block is combined with its output, a technique called ‘skip connections,’ which improves the model’s generalisation capability. Swin and DEIT are transformer-based architectures. Transformers were initially developed for Natural Language Processing but found their utility in vision tasks due to their proven efficiency in general-purpose vision tasks.

Furthermore, a machine learning model was developed that is capable of predicting the percentage of each of the 15 emotional responses (ER), as discussed in the Section 2.1, present in each image. To achieve this, a new dataset was created utilising the 872 images from the survey. Each image was annotated with the percentage vote assigned to the 15 emotion labels. The next task was to train a classification model capable of predicting the distribution of emotional responses within each image. A CNN-based regression model was developed that analyses the visual features of the images to predict the percentage of each emotion present in an image. The model has two components: an image encoder and a decoder. The task of an image encoder is to generate a lower-dimensional representation of an image (image embeddings) whilst extracting all useful information. The OpenAI CLIP [39] image encoder was utilised to generate image embeddings. The function of a decoder is to utilise the embeddings to predict the label of the image. The CLIP [39] image encoder is a state-of-the-art model; it is trained on 400 million image–text pairs to generate image representations. Image representations, or embeddings, are representations of an image in a lower vector space. Embedding models are designed to transform complex visual data into a summarised representation in a lower-dimensional vector space while retaining the unique patterns and features of the image. The decoder model is relatively simple, consisting of three fully connected networks with batch normalisation and ReLU layers inserted in between them. Figure 2 illustrates the summary of the Decoder model.

To develop a basic recommendation system, first, the correlation between Biophilic and emotional labels was established from the survey results by following the below steps:

• Dominant niophilic labels: Identify images where there is a clear consensus among the participants for the dominant Biophilic characteristics.
• Proportion of emotions: For the selected images in the previous step, the proportion of each emotion expressed by participants is computed.
• Aggregated emotional proportions: Next, the proportion of emotions expressed for each Biophilic class is aggregated.

By following the above steps, a correlation table was generated depicting the correlation between Biophilic traits and emotional responses, which is the basis of the recommendation system to recommend artworks based on emotional responses. For instance, if someone is feeling ‘Sad’ or ‘Downhearted’ and seeks uplifting content, the system would recommend artworks associated with emotions like ‘Happy’ or ‘Cheerful’. Similarly, for someone seeking relaxation, the system would recommend artworks linked with emotions such as ‘Relaxed’ and ‘Calm’.

Participants provided ratings for their emotional state by assessing each of the 15 PANAS emotions on a scale ranging from 1 (indicating minimal intensity) to 5 (representing maximum intensity). Leveraging the correlation computed, the algorithm identifies suitable Biophilic characteristics and subsequently presents 20 images from these identified classes. For instance, if the predominant emotion is ‘Sad’, which is considered negative, the algorithm selects the Biophilic attribute where the lowest percentage of participants reported ‘Sad’. Conversely, if the primary emotion is ‘Relaxed’, being positive, the algorithm suggests images associated with Biophilic traits exhibiting a higher prevalence of other positive emotions. Furthermore, the recommendation system aims to enhance the least positive emotion by showcasing images from Biophilic classes known to evoke that specific emotion. To evaluate the recommendation system, a survey was devised where participants were requested to share their emotional responses after exposure to the Biophilic artworks recommended by the algorithm. Figure 3 depicts the high-level diagram of the recommendation system. This recommendation system is designed to suggest Biophilic artworks to users based on their current emotional state, aiming to support emotional well-being through nature-inspired art. Initially, artworks are analysed and tagged using two classifiers: a Biophilic classifier that assigns tags based on natural elements, an emotional classifier, and a correlation matrix that categorises artworks by the Biophilic features and emotional responses they evoke. Thus creating a database of artworks with both Biophilic and emotional attributes. When a user interacts with the system, they input their emotional state by rating the 15 emotions. The system identifies the dominant emotion from these ratings; if the dominant emotion is negative, the system recommends artworks associated with positive emotions to potentially uplift the user’s mood. Conversely, if the user’s dominant emotion is positive, it suggests other positively tagged artworks to maintain or enhance their positive emotional state. This approach utilises both Biophilic and emotional qualities in the art to offer tailored recommendations that align with and support the user’s emotional needs.

3. Results and Discussion

This section discusses the results in detail.

3.1. Metrics of Biophilic Attributes and Emotional Responses

In this study, the authors of this paper, as a multidisciplinary team of artists, architects, engineers, psychologists, and computer scientists, identified a set of 15 emotional responses and 14 Biophilic categories. Based on the most established Biophilic feature for architectural design [7], the “Biophilic attributes of arts” (BAA) have been defined:

Nature in the paintings: Connection with nature: seeing an element, organism, or process that is natural. Natural organisation: lack of human design or rigid order Presence of water: water in any form, ice, sea, pond, river, rain, etc. Presence of animals: animals of any sort. Presence of plants or fungi: plants and or fungi of any sort. Varying light: the varying intensity and variety of light effects that exist in nature, from the dynamic to the diffuse.

Natural analogues in paintings: Biomorphic shapes: shapes that are reminiscent of animals, plants, or natural forms. Natural materials: materials are either nature in its raw form or that retain a sense of nature after having been processed, i.e., timber, stone, etc. Complexity in order: natural detail across the macro and micro scales with a sense of how these scales relate to each other.

Nature of the paintings: Unimpeded views: a prospect, for example, from a mountain or through forest clearing. Refuge: a place of safety, like a cave, glade, or clearing. Mystery: a quality that entices a closer look or the desire to discover something new. Risk: an identifiable threat, but one that has a reliable safeguard; a sense of challenge. Awe: something that defies your understanding and that could lead to a change in perception.

For potential emotional responses, the existing validated Positive Affect Negative Affect Scale (PANAS) Items [18] were used for defining the “emotional attributes of Biophilic arts” (EABA). Positive emotions: ‘Relaxed, Calm’, ‘Proud, Grand’, ‘Inspired, Amazed’, ‘Happy, Cheerful’, ‘Determined, Confident’, ‘Safe, Cosy’, ‘Energized, Excited’, ‘Nourished, Fulfilled’, and ‘Attentive, Concentrating’. Negative emotions: ‘Upset, Distressed’, ‘Shy, Bashful’, ‘Sad, Downhearted’, ‘Hostile, Angry’, ‘Ashamed, Guilty’, and ‘Afraid, Frightened’.

3.2. Data Repository and Public Survey Results Analysis

To gain valuable insights from the survey data, statistical analyses were performed. At first, only those images were selected for which there was a consensus by the participants on the dominant Biophilic category; a table was created by recording each image and the responses for each emotion in numbers. A MANOVA was performed with the Biophilic traits as the independent variable and the emotional tags as the dependent variable.

Table 1. Multivariate-ANNOVA to establish a correlation between Biophilic attributes and emotion classes.

Intercept l	Value	Num DF	Den DF	F Value	Pr > F
Wilks’ lambda	0.00	15.00	844.00	4,445,658,579,532,924.50	0.00
Pillai’s trace	1.00	15.00	844.00	4,445,658,579,532,924.50	0.00
Hotelling–Lawley trace	79,010,519,778,428.75	15.00	844.00	4,445,658,579,532,924.50	0.00
Roy’s greatest root	79,010,519,778,428.75	15.00	844.00	4,445,658,579,532,924.50	0.00

records the results of the four methods used. The output shows the analysis using different test statistics. The second one, Pillai’s trace, is known to be relatively conservative: it gives a significant result less easily (the differences must be bigger to obtain significant output). The Pillai’s trace test statistics for the Biophilic data are statistically significant [Pillai’s trace = 1.0000, F(14, 6) = 224,336.37, p < 0.001], and the emotional data show [Pillai’s trace = 1.0000, F(15, 5) = 737,078,599.19, p < 0.001] and indicate that images of the questions have a statistically significant association with all the categories for both emotional data and Biophilic data. The Wilks’ Lambda is another often-used test statistic. Hotelling–Lawley trace and Roy’s greatest root are also alternative options. There is no absolute consensus in the statistical literature as to which test statistic should be preferred. The p-values are shown in the right column and are all inferior to 0.05, which confirms that images in the questionnaire have an impact on the categories for both Biophilic and emotion. The p-values for all the methods are inferior to 0.05, indicating that the Biophilic and emotional labels correlate with the images in the survey. See details in Table 1.

From the survey, only those images were selected for which there was a consensus among the participants on the dominant Biophilic category. A table was formulated, recording each image and the responses for each emotion in numbers. Based on the survey results, the distribution of emotions is shown in

Table 2. Prevalence of emotions by responses to the survey.

Emotional Label	Prevalence (%)
Hostile, Angry	0.8
Shy, Bashful	0.5
Ashamed, Guilty	0.2
Safe, Cosy	1.3
Determined, Confident	1.1
Energised, Excited	2.1
Happy, Cheerful	3.5
Nourished, Fulfilled	3.5
Afraid, Frightened	4.5
Upset, Distressed	5.3
Sad, Downhearted	7.7
Proud, Grand	8.8
Inspired, Amazed	9.8
Attentive, Concentrating	24.5
Relaxed, Calm	26.3

, which shows the prevalence of positive emotions. The emotions ‘Relaxed, Calm’ with 26.3%, ‘Attentive, Concentrating’ with 24.5%, and ’Inspired, Amazed’ with 9.8% are the most occurring emotions in the survey.

3.3. ML Classification Results and Recommendation Systems

The accuracies of the different models on the validation dataset are presented in

Table 3. Accuracy of the classification algorithms.

Models	Accuracy (%)
ResNet50	69.30
DEIT Transformers	68.40
Swin Transformers	68.80

. For this study, ResNet50, Swin Transformers, and DEIT were used; they represent state-of-the-art algorithms for image classification. The accuracies exhibited by the models are notably modest compared with traditional image classification studies based on styles, genres, schools of art, and brushstrokes [21,22,34]. Upon analysis, several shortcomings of the model were identified. Firstly, the artwork dataset comprised only 872 images, encompassing all 14 Biophilic categories. This dataset size is considerably inadequate for training machine learning models effectively. To increase the size of the dataset, a cosine similarity technique was used; this metric enables consideration of images that visually resonate most with participant responses in mind. Employing cosine similarity to assign labels to images presents several shortcomings and may not be considered an ideal method. The technique using cosine similarity fails to capture the spatial features because it concentrates only on the angle between the vector representation of the images. This method is also insensitive to any noise and is unable to extract complex features from the images. Thus, this method mislabels a lot of artworks to incorrect Biophilic categories. In the future, the dataset will be expanded by gathering labels from additional surveys or through manual annotations by experts. This will help diversify the dataset by reducing class imbalances. If Biophilic traits are underrepresented, then the model fails to learn the characteristics of the underrepresented traits. Secondly, survey responses introduce bias in the dataset; thus, data pre-processing techniques are to be used in the future to minimise the biases. Certain Biophilic tags, such as ‘Risk’, ‘Awe’, and ’Mystery’, are inherently subjective and open to interpretation. To mitigate the ambiguity, Biophilic traits with strong emotional aspects are to be discarded and discrete, objectively defined Biophilic labels are to be focused on. Moving forward, a multi-label classification algorithm capable of predicting multiple Biophilic labels present in an image rather than just one will be explored. Some of the best results for Biophilic attributes have been illustrated in Figure 4.

For evaluating the emotional classifier, the R-squared metric was used. R-squared is a statistical metric used in regression models to quantify the extent to which the independent variable can explain the variation in the dependent variable. An R-squared score of 0.82 was achieved on the training dataset and an R-squared score of 0.063 on the validation dataset. This indicates the model is suffering from overfitting. Training a machine learning model to predict emotions from artworks is very tough; to achieve this task, more data are needed, for which more surveys will be planned.

The next task is developing a recommendation system that takes a user’s emotional state as input and recommends Biophilic artworks. To address this challenge, only the images from the survey where the participants reached a consensus on the most dominant Biophilic trait were used. The percentage of each of the emotions for each Biophilic class was computed. As evident from the data, the majority of participants expressed positive emotions across all other classes. Specifically, many participants reported feeling ‘Relaxed’ or ‘Calm’, while others indicated they felt ‘Attentive’ or ‘Concentrating’.

Next, a table was created where the fraction of emotional labels for each Biophilic trait is computed with the help of the survey data; see

Table 4. Fraction of emotions per Biophilic category as per the survey. * Most expressed emotion. ^† Least expressed emotion.

Biophilic Label	Emotional Label
	Relaxed, Calm	Proud, Grand	Nourishing, Fullfilled	Attentive, Concentrating	Sad, Downhearted	Afraid, Frightened	Upset, Distressed	Inspired, Amazed	Energised, Excited	Happy, Cheerful	Determined, Confident	Safe, Cosy	Ashamed, Guilty	Shy, Bashful	Hostile, Angry
Mystery	0.144	0.095	0.023	0.225 *	0.107	0.062	0.055	0.081	0.022	0.039	0.033	0.034	0.014 †	0.028	0.04
Varying light	0.192 *	0.092	0.029	0.163	0.092	0.061	0.047	0.113	0.037	0.053	0.037	0.021	0.018 †	0.024	0.021
Presence of water	0.349 *	0.069	0.069	0.102	0.043	0.026	0.013	0.117	0.061	0.066	0.02	0.041	0.005 †	0.008	0.01
Connection with nature	0.189 *	0.047	0.134	0.103	0.037	0.037	0.045	0.137	0.047	0.063	0.026	0.066	0.024	0.026	0.018 †
Presence of animals	0.116	0.092	0.048	0.126 *	0.072	0.085	0.116	0.075	0.072	0.061	0.031	0.031	0.017	0.014 †	0.044
Biomorphic shapes	0.121	0.13	0.037	0.195 *	0.098	0.07	0.065	0.126	0.028	0.019	0.023	0.023	0.009 †	0.028	0.028
Natural materials	0.112	0.158 *	0.102	0.153	0.092	0.066	0.026	0.138	0.015	0.026	0.026	0.056	0 †	0.015	0.015
Unimpeded views	0.349 *	0.038	0.123	0.057	0.047	0 †	0.019	0.132	0.094	0.047	0 †	0.047	0.019	0.009	0.019
Awe	0.106	0.068	0.061	0.129	0.129	0.068	0.053	0.167 *	0.015 †	0.045	0.015	0.038	0.03	0.015 †	0.061
Refuge	0.065	0.087	0.022	0.217 *	0.174	0.13	0 †	0.087	0.065	0.022	0.022	0.043	0.022	0.022	0.022
Natural organisation	0.055	0.073	0.055	0.127	0.127	0.055	0.073	0.182 *	0.091	0.036	0 †	0.055	0 †	0.018	0.055
Presence of plants or fungi	0.278 *	0.022	0.183	0.106	0.044	0.006	0.006	0.039	0.044	0.128	0 †	0.133	0.006	0.006	0 †
Risk	0.037	0.074	0.007 †	0.074	0.132	0.206 *	0.162	0.059	0.022	0.015	0.044	0.015	0.022	0.015	0.118
Complexity in order	0.078	0.167	0.011 †	0.333 *	0.089	0.022	0.067	0.089	0.033	0.022	0.033	0.022	0.011 †	0.011 †	0.011 †

. The cells highlighted in blue represent the most dominant emotion, and the grey represents the least expressed emotion per Biophilic category. This correlation table was the basis of the recommendation system. To evaluate the effectiveness of the recommendation system, a straightforward survey was devised wherein the participants were asked to assess their emotional state by assigning ratings to 15 emotions on a scale from 1 to 5, with 1 representing the lowest and 5 the highest intensity. Subsequently, these ratings serve as input for the recommendation system to suggest 20 Biophilic artworks, each sequentially displayed for 5 s. Following the exhibition of all 20 images, participants are prompted once more to gauge their emotional state. In total, 50 surveys were conducted, analysing the participants’ emotional responses before and after exposure to the artworks.

The responses for each of the 15 emotions both before and after participants were exposed to the Biophilic artworks from 50 surveys were recorded.

Table 5. Recommendation System feedbacks ratings. * Increase in average ratings. ^† Decrease in average ratings.

Emotional Label	Recommendation System		Random Recommendations
	Before	After	Before	After
Relaxed, Calm	2.44 *	2.69 *	2.4 †	2.36 †
Proud, Grand	1.42 †	1.38 †	1.23	1.23
Nourished, Fulfilled	1.87 †	1.81 †	1.80 †	1.73 †
Attentive, Concentrating	2.87 †	2.75 †	2.53 *	2.63 *
Inspired, Amazed	1.48 *	2.02 *	0.96 *	1.56 *
Energised, Excited	1.57 *	1.81 *	1.06 *	1.26 *
Happy, Cheerful	1.89 *	2.18 *	1.6 †	1.53 †
Determined, Confident	2.02 †	1.87 †	1.8 †	1.63 †
Safe, Cosy	2.40 †	2.38 †	2.3 †	2.13 †
Sad, Downhearted	0.81 †	0.48 †	0.9 †	0.73 †
Afraid, Frightened	0.55 †	0.34 †	0.4	0.4
Upset, Distressed	0.44 †	0.34 †	0.3 *	0.4 *
Ashamed, Guilty	0.35 *	0.44 *	0.43 *	0.46 *
Shy, Bashful	0.71 †	0.34 †	0.7 †	0.53 †
Hostile, Angry	0.34 †	0.20 †	0.26 *	0.33 *

records the average before and after emotional ratings of all participants; blue signifies an increase in the emotion rating after exposure to artworks, and grey implies a decrease in the emotion rating post-exposure. Interesting emerging trends were observed, and the prototype was successful in improving a range of positive emotions like ‘Relaxed, Calm’, ‘Inspired, Amazed’, ‘Energised, Excited’, and ‘Happy, Cheerful.’ Meanwhile, negative emotions like ‘Sad, Downhearted’, ‘Afraid, Frightened’, ‘Upset, Distressed’, ‘Shy, Bashful’, and ‘Hostile, Angry’ were considerably diluted. The findings suggest that being exposed to Biophilic artworks and images has the potential to enhance positive emotions while mitigating negative ones. The average emotional responses before and after exposure to Biophilic artwork were compared with those obtained from random recommendations, where the system suggested artworks that were more neutral and less significantly Biophilic. This comparison was conducted to assess the performance of the recommendation system. As seen in Table 5, there is no underlying emotional pattern for randomly recommended artworks.

To further analyse the results, the Wilcoxon signed-rank test was performed. The Wilcoxon signed-rank test is a non-parametric univariate test. This method is used to examine whether there is any significant difference in scores ‘before’ and ‘after’ a condition, intervention, or treatment. The Wilcoxon signed-rank test was performed on the individual emotions, where the hypothesis is as follows:

• Null hypothesis ($H_0$): The difference between the observations is not statistically significant, i.e., the observations are random coincidences.
• Alternative hypothesis ($H_A$): The difference between the observations is statistically significant, i.e., the observations are not random coincidences.

If the p-value is <0.05, typically the Null Hypothesis ($H_0$) is rejected, as the probability of observing the data under the Null Hypothesis ($H_0$), i.e., random observations, is <5%.

In this study, Python’s ‘scipy.stats’ package was used to perform the Wilcoxon signed-rank test for each emotion based on the survey data. A table was created for each emotion with two columns, before and after ratings, capturing participants’ responses after exposure to Biophilic images recommended by this Biophilic curation system. In this study, the Wilcoxon signed-rank test was applied for each emotion using the code shown in Figure 5. Additionally, similar statistical analyses were performed for each emotion with random recommendations.

Table 6. Wilcoxon signed-rank test per emotion. * p-value < 0.05.

Emotional Label	p-Value (Recommendation System)	p-Value (Random Recommendation)
Relaxed, Calm	0.039 *	0.796
Proud, Grand	0.673	0.902
Nourished, Fulfilled	0.532	0.648
Attentive, Concentrating	0.201	0.405
Sad, Downhearted	0.007 *	0.268
Afraid, Frightened	0.108	1.0
Upset, Distressed	0.346	0.47
Inspired, Amazed	0.0002 *	0.013 *
Energised, Excited	0.053	0.318
Happy, Cheerful	0.022 *	0.694
Determined, Confident	0.252	0.272
Safe, Cosy	0.957	0.368
Ashamed, Guilty	0.601	0.748
Shy, Bashful	0.011 *	0.371
Hostile, Angry	0.114	0.414

records the p-value for each emotion. As per the observation in the case of the proposed recommendation system, the emotions ‘Relaxed, Calm’, ‘Sad, Downhearted’, ‘Inspired, Amazed’, ‘Happy, Cheerful’, and ‘Shy, Bashful’ (highlighted in blue in Table 6) have p-values < 0.05, which implies the responses post-exposure to Biophilic artworks were statistically significant from initial responses. For the emotions ‘Sad, Downhearted’ and ‘Inspired, Amazed’, it is observed that the p-values are <0.01, which indicates that it is highly unlikely (<1%) that the observed effect is due to random chance. Whereas for random recommendations the p-value is <0.05 only for the emotion ‘Inspired, Amazed’, thus for random recommendations the pre- and post-observations are not statistically significant. Thus, from the survey results, it can be inferred that the Biophilic artwork recommendation system effectively induces positive emotions. Further research is needed to study the effects on emotion of prolonged exposure to Biophilic artworks.

4. Conclusions

4.1. Summary of Key Findings

This study discusses the disparity in accessing art among economically and mobility disadvantaged groups. To bridge this gap, a novel low-cost AI-powered Biophilic arts curation and display system is needed to promote mental health and well-being. A novel Biophilic arts curation framework that offers a holistic approach to categorising Biophilic attributes and emotional responses in artworks. This curation framework aims to automate the categorisation and recommendation of Biophilic and emotional elements using machine learning and artificial intelligence algorithms, streamlining the analysis and enhancing accessibility to the arts and the precision of recommendations. Key Biophilic design patterns have been developed by Browning [7], but they can’t be directly translated to categorise and analyse Biophilic artworks. Based on focus group discussion, a set of 14 Biophilic attributes and 15 emotional metrics were developed. A public survey was carried out, and results show a positive correlation between positive emotional responses and Biophilic attributes. The top four most significant emotional responses are relaxation, attentiveness, pride, and inspiration, with a strong correlation with the presence of plants, natural organisation, and natural materials. Respectively, negative emotions like ‘Afraid, Frightened’, ‘Upset, Distressed’, and ‘Sad, Downhearted’ have a strong correlation with ‘Risk’. All the images used for the survey demonstrated therapeutic value with varied levels of significance. The statistical correlation was utilised to inform the design of the digital curation and personalised recommendation system.

This digital curation system integrates artificial intelligence and machine learning algorithms to produce an intelligent and personalised recommendation system to improve users’ mental health. This study aims to fill in the existing gap for developing AI-based digital Biophilic therapeutic systems. This study is devoted to building a Biophilic classification algorithm that categorises artworks into the corresponding Biophilic characteristics with the help of state-of-the-art deep-learning techniques. This artificially intelligent model can help artists sort their artwork by the various Biophilic characteristics and create a curated collection of nature-inspired art. The focus of this project is to improve the health and well-being of the occupants of the built environment. The simple recommendation system was able to boost positive emotions like ‘Relaxed, Calm’, ‘Inspired, Amazed’, ‘Energised, Excited’, and ‘Happy, Cheerful’ and reduce most of the negative emotions. With the p-value of emotions ‘Relaxed, calm’, ‘Sad, Downhearted’, ‘Inspired, Amazed’, ‘Happy, Cheerful’, and ‘Shy, Bashful’ $<0.05$ indicating the recommendation system had a statistical impact on the emotions of the participants.

4.2. Implications of the Findings

This study offers a preliminary framework and future research trends for developing an AI-based automated Biophilic art curation system to promote health and wellbeing. This framework is developed based on interdisciplinary methodologies that include computer science, arts and design, psychology, and intelligent buildings, heavily emphasising digital displays of the Biophilic arts. In the interconnected worlds of the Digital Age, digital Biophilic arts can help the urban population to engage with nature more effectively and meaningfully. Furthermore, digital automated curation enhances this access by efficiently categorising artworks based on Biophilic themes and emotions, allowing users to discover relevant pieces tailored to their interests. It also facilitates the creation of virtual galleries and interactive platforms, broadening engagement with diverse artistic expressions. By streamlining the organisation and showcasing of artworks, digital curation reduces barriers for artists and curators, amplifying the arts’ role in promoting sustainable practices and deepening connections with nature. Engagement with nature and the arts improves cognitive abilities and emotional well-being [1,3], fosters cultural awareness [16], encourages environmental stewardship [11], and stimulates dialogue on critical issues [5,6,8,10].

4.3. Limitation of the Study and Future Research

Biophilic artwork recommendations can be developed based on a myriad of factors like the weather conditions, primarily temperature, time of day, e.g., morning, noon, or evening, personal preferences, circadian rhythms, and emotional status of the user. Future research activities will also be explored to develop AI-based classification models to predict emotions from artworks and build a personalised recommendation system that recommends artworks based on Biophilic traits, emotional state, personal circadian cycles, environmental conditions, etc. On the other hand, in order to improve the classification accuracy, a bigger and better-annotated dataset is required to train a machine learning model successfully.