Data-driven Mapping of Wellbeing Science: A Multi-objective Optimization Framework using BERTopic

1 Data-driven Mapping of Wellbeing Science: A Multi-objective Optimization Framework using BERTopic Yusuke Inoue Graduate School of Media and Governance, Keio University [email protected] Abstract This study presents a multi-objective optimization framework for BERTopic to analyze large-scale scholarly literature on wellbeing. Using the Non-dominated Sorting Genetic Algorithm II (NSGA-II), UMAP and HDBSCAN hyperparameters were optimized to maximize C_V coherence and topic diversity under constraints on topic count and noise ratio. Applying the optimized model to 80,040 PubMed articles (2020–2024) achieved a C_V coherence of 0.83, a topic diversity of 0.74, and identified 266 interpretable topics grouped into 22 semantically coherent communities. The optimization process also compressed the noise ratio to 34.75%, improving topic stability. The analysis revealed a structured landscape of wellbeing research spanning medical and public health, psychological and behavioral wellbeing, and social and environmental determinants, interconnected through bridging areas such as digital health, occupational wellbeing, and lifestyle interventions. Within this broader set of 22 communities, three major domains—public health and pandemic response, psychological and behavioral wellbeing, and social and environmental determinants—emerged as central hubs. Adopting a data-driven and inductive approach to large-scale literature analysis, the framework enables systematic reconstruction of wellbeing science. It provides a transparent and reproducible tool for evidence-based research mapping, policy design, and strategic foresight in complex interdisciplinary fields. Keywords: BERTopic, Multi-objective optimization, Scholarly literature analysis, Knowledge discovery, Decision support, wellbeing, Interdisciplinary mapping 1. Introduction Wellbeing research has rapidly expanded as an interdisciplinary field bridging the life sciences, social sciences, and behavioral sciences. Over the past two decades, the number of related publications has grown exponentially, encompassing diverse domains such as medicine, psychology, environment, and technology. As a result, wellbeing has evolved into a crosscutting academic domain pursued through multiple disciplinary lenses. However, precisely because of this diversity, the concept of “wllbeing” remains theoretically ambiguous. Different fields have developed their own definitions, operationalizations, and measurement frameworks, resulting in a fragmented knowledge system in which the underlying conceptual core has become diffuse. Systematically organizing and theoretically reconstructing this vast and heterogeneous body of research has thus become a central challenge for both academic and practical purposes. For researchers, policymakers, and practitioners, understanding the behavioral mechanisms that underlie wellbeing is essential for designing effective interventions, promoting sustainable lifestyles, and implementing evidence-based policies. Yet, most existing meta-analytical studies presuppose theoretical categories such as “subjective wellbeing,” “mental health,” or “social capital.” Few have empirically examined how these concepts are actually structured and interconnected across the diverse literature. 2 Recent advances in transformer-based topic modeling—particularly BERTopic—now enable the extraction of latent knowledge structures from large-scale textual corpora. Such methods function not merely as tools for bibliometric classification but as instruments for inductive theory reconstruction. They offer the potential to reconstruct the implicit conceptual systems through which scientific communities have defined, measured, and intervened upon wllbeing. This study analyzes approximately 80,040 PubMed-indexed articles related to wllbeing. Using BERTopic, we extract 266 topics and construct a topic network based on keyword cooccurrence and lexical similarity. Through this network, we aim to elucidate the behavioral knowledge structure embedded in contemporary wellbeing research. In contrast to previous studies that impose predefined theoretical frameworks, this study adopts a data-driven inductive approach to uncover the latent conceptual organization and interrelations among topics. The central research question is whether wellbeing research as a whole exhibits a set of shared behavioral principles—value formation, action generation, and environmental adaptation—that operate as a non-hierarchical and dynamic system. Subsequently, we examine how the resulting structure aligns with established frameworks in utility theory, decision theory, and behavioral economics, not as a priori premises but as a post hoc theoretical reconstruction. The main contributions of this study are as follows:. 1. A large-scale, data-driven, and inductive analysis of wellbeing research that identifies an interrelated behavioral knowledge structure composed of the three domains of value formation, action generation, and environmental adaptation. 2. A methodological framework integrating topic modeling and network analysis for inductive theory generation from heterogeneous scholarly corpora. 3. A theoretical synthesis demonstrating how data-driven knowledge structures can be reinterpreted within the conceptual scope of utility theory, decision theory, and behavioral economics, thereby ensuring both theoretical coherence and extensibility. 2. Related Work BERTopic has emerged as a prominent embedding-based topic modeling method for extracting latent thematic structures from academic literature, complementing traditional probabilistic and matrix factorization approaches. Classical models such as Latent Dirichlet Allocation (LDA), Hierarchical Dirichlet Process (HDP), and Dynamic Topic Modeling (DTM) have been widely used, alongside Non-negative Matrix Factorization (NMF) and more recent embedding-based approaches such as Top2Vec. BERTopic leverages transformer-based language models, notably BERT (Devlin et al., 2019), to generate document embeddings, followed by dimensionality reduction and density-based clustering. Several studies have compared BERTopic to conventional techniques. Egger and Yu (2022) found that both BERTopic and NMF achieved high classification accuracy in social media analysis. Pavithra and Savitha (2024), in a comparative study involving LDA, HDP, NMF, BERTopic, and DTM for academic literature, concluded that DTM provided the most effective results, with the others—including BERTopic—showing broadly comparable performance. Other applications of BERTopic in scholarly contexts include He et al. (2025) and Al Azher et al. (2025), which demonstrated its capacity to uncover domain-specific topic structures. Beyond these scholarly contexts, topic modeling has also been actively applied to map research landscapes and forecast emerging trends (Yu & Xiang, 2023; Park et al., 2024). Other studies have integrated word embeddings with probabilistic models for tasks such as economic policy uncertainty estimation and domain-specific forecasting (Huang, Zhang and Zhou, 2023). More broadly, recent works emphasize actionable analytics, interpretability, and explainability in applied decision-support systems (Abusitta et al., 2024; Panagoulias et al., 2024). Our study is aligned with this perspective but extends prior work by focusing not only on extracting topics 3 but also on systematically optimizing BERTopic’s hyperparameters through a multi-objective framework to ensure robustness and reproducibility in large-scale literature analysis. Despite these strengths, the interpretability and stability of BERTopic classifications are highly dependent on hyperparameter configurations in its dimensionality reduction (UMAP) and clustering (HDBSCAN) stages. Prior research has reported that both LDA and BERTopic often overestimate the number of topics, reducing interpretability, and that parameter sensitivity can undermine classification stability (Farea et al., 2024; Grootendorst, 2024). Similarly, Medvecki et al. (2024) emphasized the difficulty of ensuring consistent clustering quality in multilingual contexts. Recent work by Inoue (2025) further explored this issue by formulating a multi-objective optimization framework for BERTopic applied to large-scale scholarly literature on autoregressive models, demonstrating that multi-objective search algorithms such as NSGAII can effectively balance topic coherence and diversity under topic-count and noise constraints (Inoue, 2025). These findings collectively indicate that, while BERTopic and related models offer promising results, they still lack mechanisms to guarantee robust and stable topic clustering across large-scale corpora. To date, few studies have systematically addressed this challenge through multi-objective hyperparameter optimization. In particular, existing approaches have not provided a practical and quantitative framework that balances interpretability, accuracy, and stability in large-scale literature analysis. Nor have they explicitly demonstrated how such frameworks could support decision-making and knowledge discovery in applied domains. Addressing this gap, the present study develops and evaluates a multi-objective optimization framework for BERTopic that ensures stable clustering performance and provides actionable, interpretable insights for large-scale scholarly literature analysis, validated through an inductive approach based on large-scale data. 3. Method 3.1 Data collection The dataset employed in this study was retrieved from the PubMed database using the NCBI E-utilities API (ESearch / EFetch). Publications were included if the terms “wellbeing,” “wll eing,” or “well-being,” appeared in the title or abstract. The search period covered January 1, 2020, to December 31, 2024. Each publication was uniquely identified by its PubMed ID (PMID), and all duplicate entries were removed. For each record, metadata including title, DOI, authors, journal, publication year, and abstract text were extracted. In total, 80,040 unique articles were collected, distributed as follows: 10,439 (2020), 14,514 (2021), 17,064 (2022), 16,963 (2023), and 21,060 (2024). Duplicate records were properly processed and eliminated. This corpus of 80,040 publications served as the input for the subsequent embedding and topic-network analyses. 3.2 Proposed Multi-objective Optimization Framework for BERTopic In this study, BERTopic was employed to extract topics from a large-scale academic corpus. Each document was pre-embedded using Sentence-BERT (all-mpnet-base-v2, 768 dimensions). Dimensionality reduction was performed with UMAP (cosine metric), and clustering was conducted using HDBSCAN (cluster_selection_method = eom). In BERTopic, the parameter calculate_probabilities was set to False, and the key hyperparameters—top_n_words and min_topic_size—were included in the search space. The hyperparameter tuning was formulated as a bi-objective single-constraint optimization problem. The search space comprised n_neighbors, min_dist, and n_components for UMAP; min_cluster_size and min_samples for HDBSCAN; and top_n_words and min_topic_size for BERTopic. The optimization objectives were defined as follows (maximized in theory, minimized in implementation): 4 ◼ C_V Coherence, measuring the semantic consistency of top words within each topic. ◼ Topic Diversity, defined as the proportion of unique lexical items among top-ranked words across topics. A single feasibility constraint was imposed, requiring the number of valid topics to be no fewer than MIN_TOPICS = 80 (approximately 0.1% of the corpus, ensuring sufficient thematic granularity). The noise ratio (proportion of documents assigned to topic = –1) was also recorded as a supplementary indicator. The optimization employed NSGA-II, with a total of 1000 evaluations executed in multiple batches. The population size was set to 70, the number of offsprings to 30, and Latin Hypercube Sampling (LHS) was used for initialization with duplicate elimination. Early stopping was based on the absence of improvement in the hypervolume indicator (reference point [1.0, 1.0]) with patience = 15. Evaluation parallelism was automatically adjusted according to computational resources. The evaluation data were processed in two phases: the first using 30% of the corpus (subset) and the second using the full dataset. The subset was randomly sampled with a fixed SEED, and corresponding embeddings were extracted using identical indices. For each trial, normalized parameters, C_V coherence, diversity, noise ratio, topic count, evaluation stage (subset/full), and document count were recorded for reproducibility. In addition, a Pareto analysis was conducted on the feasible solution set using C_V and Diversity as dual objectives. Although the Pareto-optimal front was successfully identified, none of the Pareto-optimal solutions satisfied the noise-ratio constraint (≤ 35%). Therefore, the final model selection followed the conventional weighted aggregation approach for practical interpretability. The final selection criteria were as follows: (i) C_V score ≥ 0.5 (ii) Number of topics ≥ 80 (≈ 0.1% of the corpus) (iii) g_topics ≤ 0 Here, g_topics represents the constraint function defined as g_topics = MIN_TOPICS − Num Topics, which measures the degree of constraint violation in the optimization process (feasible if g_topics ≤ 0). Feasible solutions were normalized using min–max scaling, and a Combined Score was computed as 0.5 × C_V + 0.5 × Diversity. The top 20 candidates were retained, and among those with a noise ratio ≤ 35%, the solution with the highest Combined Score was designated as the optimal configuration. 4. Analysis Results The optimal parameter configuration derived from the above multi-objective optimization process achieved a C_V Coherence Score of 0.83 and a Topic Diversity Score of 0.74, yielding 266 interpretable topics, while the noise ratio was constrained to 34.75% despite the large scale of the dataset, demonstrating the framework’s ability to maintain clustering stability in extensive corpora. A summary of the topic frequency distribution is presented in Figure 1, and representative keywords for the most frequent topics are provided in Figure 2. Among the 266 interpretable topics, Topics 0–47 collectively accounted for 50.45% of all classified documents. Topic 0 alone represented 3.52% of the corpus, suggesting a more balanced distribution of thematic concentration compared to earlier results. The overall topic frequency distribution continues to exhibit a Zipf-like pattern, characterized by a few highfrequency topics dominating the corpus alongside a long tail of less frequent ones. To understand the overall topic structure, the analysis focused exclusively on topics representing more than 1% of the corpus, thereby identifying the dominant domains within wellbeing research. The results, excluding noise topics, are presented in Figure 2. An overview based on the representative keywords of each topic shows that the largest topic (3.52%) concerns COVID-19 and pandemic-related mental health (Topic 0), indicating 5 sustained scholarly attention to the psychological impacts of global crises from the perspective of wellbeing. This is followed by Refugees and Migration (Topic 1; 2.99%), Dementia and Caregiving (Topic 2; 2.48%), and Violence and Abuse (Topic 3; 2.30%), underscoring the continued importance of research on social vulnerability and care. Moderate-sized topics include Mindfulness and Stress Reduction (Topic 4; 1.87%), Parental and Child Wellbeing during the Pandemic (Topic 5; 1.77%), Oral Health (Topic 6; 1.76%), IndigenousWellbeing (Topic 7; 1.56%), Social Media Use and Addiction (Topic 8; 1.54%), and Gender and Sexual Minority Health (Topic 9; 1.39%). Additionally, smaller yet thematically distinct topics, each accounting for approximately 1% of the corpus, were identified: Sleep Quality (Topic 10; 1.35%), HIV and Stigma (Topic 11; 1.32%), Arts and Music Therapy (Topic 12; 1.24%), Nutrition and Diet (Topic 13; 1.19%), Autism Spectrum (Topic 14; 1.15%), and Maternal and Fetal Wellbeing (Topic 15; 1.08%). These findings objectively reaffirm that wellbeing research spans a diverse and interdisciplinary range of areas, including mental health, social inclusion, caregiving, lifestyle, development, maternal and child health, and the arts. They also demonstrate that the proposed multi-objective optimization framework enables systematic extraction and reproducible structuring of complex, multi-topic knowledge across varied social and health domains. Figure 1. Number of topics identified and their cumulative distribution across the wellbeing-related literature corpus. 0% 20% 40% 60% 80% 100% 120% 0 200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000 Count Cumulative 6 Figure 2. Representative keywords of the most frequent topics. Topic Count Proportion Representation_cleaned Topic Count Proportion Representation_cleaned 0 1836 3.52% covid19, pandemic, covid19 pandemic, mental, lockdown 70 206 0.39% energy, environmental, economic, ewp, pollution 1 1563 2.99% refugees, refugee, migrant, migrants, discrimination 71 204 0.39% physical activity, physical, activity, fitness, pa 2 1293 2.48% dementia, caregivers, care, caregiving, people dementia 72 202 0.39% nursing, nursing students, resilience, students, nurses 3 1200 2.30% violence, ipv, child, abuse, aces 73 197 0.38% nature, forest, naturebased, natural, forest therapy 4 979 1.87% mindfulness, yoga, meditation, mindfulnessbased, stress 74 196 0.38% welfare, animals, animal, captive, animal welfare 5 927 1.77% pandemic, covid19, children, parents, covid19 pandemic 75 195 0.37% epilepsy, seizure, pwe, seizures, patients epilepsy 6 919 1.76% oral, oral health, dental, caries, ohrqol 76 189 0.36% burnout, physicians, physician, physician burnout, professional 7 815 1.56% indigenous, aboriginal, strait, torres, torres strait 77 185 0.35% gaming, games, game, video, esports 8 803 1.54% media, social media, use, media use, addiction 78 180 0.34% sentiment, tweets, twitter, social media, media 9 724 1.39% transgender, lgbtq, gender, sexual, minority 79 179 0.34% vaccine, vaccination, immunization, vaccines, covid19 10 706 1.35% sleep, sleep quality, insomnia, sleep duration, sleep problems 80 176 0.34% medical students, medical, students, student, academic 11 690 1.32% hiv, living hiv, stigma, living, prep 81 175 0.34% acne, skin, vitiligo, cosmetic, scars 12 648 1.24% music, arts, art, music therapy, singing 82 175 0.34% stroke, stroke survivors, survivors, rehabilitation, poststroke 13 620 1.19% food, dietary, nutrition, diet, eating 83 174 0.33% digital, internet, older, older adults, adults 14 600 1.15% autistic, autism, asd, autism spectrum, spectrum 84 174 0.33% bullying, victimization, cyberbullying, bullying victimization, school 15 562 1.08% fetal, pregnancy, maternal, placental, fetal wellbeing 85 169 0.32% climate, climate change, change, heat, extreme 16 488 0.93% breast, reconstruction, breast reconstruction, breastq, mastectomy 86 166 0.32% couples, marital, relationship, romantic, divorce 17 488 0.93% cardiac, heart, heart failure, patients, failure 87 165 0.32% cancer, parents, children cancer, childhood cancer, survivors 18 481 0.92% postpartum, childbirth, birth, women, pregnancy 88 161 0.31% menopause, menopausal, women, testosterone, menopausal symptoms 19 475 0.91% diabetes, type diabetes, type, t1d, insulin 89 160 0.31% eating, weight, pandemic, covid19, lockdown 20 471 0.90% alcohol, use, substance, substance use, drug 90 160 0.31% pain, opioid, surgery, postoperative, delirium 21 464 0.89% teachers, teacher, teaching, school, job 91 159 0.30% china, older, older adults, adults, rural 22 432 0.83% disabilities, disability, intellectual, intellectual disability, children 92 158 0.30% mental health, school, schools, mental, schoolbased 23 425 0.81% app, digital, mental health, mental, apps 93 157 0.30% religious, religiosity, spirituality, religion, spiritual 24 404 0.77% schizophrenia, psychosis, recovery, stigma, mental illness 94 156 0.30% homelessness, housing, homeless, experiencing homelessness, experiencing 25 396 0.76% employees, job, employee, work, leadership 95 155 0.30% burnout, physicians, pandemic, covid19, healthcare workers 26 388 0.74% military, veterans, ptsd, veteran, personnel 96 154 0.29% cows, dairy, welfare, cattle, animal 27 382 0.73% nurses, nurse, burnout, nursing, work 97 154 0.29% eating, eating disorder, eating disorders, nervosa, anorexia 28 364 0.70% students, academic, university, university students, student 98 153 0.29% water, sanitation, households, household, water insecurity 29 359 0.69% pain, chronic pain, chronic, fibromyalgia, treatment 99 150 0.29% cancer, exercise, cancer survivors, survivors, physical activity 30 354 0.68% pandemic, covid19, trainees, medical, training 100 149 0.29% cam, complementary, medicine, cancer, integrative 31 332 0.64% hearing, hearing loss, tinnitus, cochlear, loss 101 148 0.28% endometriosis, pain, uterine, women endometriosis, women 32 331 0.63% workers, hcws, healthcare workers, covid19, healthcare 102 147 0.28% cancer, patients, cancer patients, survivorship, symptom 33 330 0.63% sport, athletes, sports, elite, athlete 103 147 0.28% maternal, antenatal, antenatal care, anc, maternal health 34 330 0.63% emotion, neural, regulation, emotion regulation, brain 104 145 0.28% police, officers, police officers, firefighters, personnel 35 330 0.63% pregnant, pregnant women, women, covid19, pandemic 105 144 0.28% media, social media, news, social, media use 36 303 0.58% dog, dogs, pet, owners, cats 106 144 0.28% moral, moral distress, nurses, distress, nursing 37 303 0.58% ecosystem, ecosystem services, land, services, ecological 107 142 0.27% stunting, malnutrition, nutrition, nutritional, dietary 38 302 0.58% school, adolescents, students, school climate, adolescent 108 141 0.27% spiritual, spiritual wellbeing, cancer, spirituality, cancer patients 39 295 0.56% ibd, bowel, inflammatory bowel, disease, bowel disease 109 141 0.27% indoor, thermal, air, heat, comfort 40 288 0.55% homes, residents, care, pandemic, covid19 110 141 0.27% obesity, weight, overweight, overweight obesity, childhood obesity 41 281 0.54% knee, hip, pain, foot, patients 111 140 0.27% nurses, covid19, nursing, pandemic, nurse 42 281 0.54% aging, older, age, older adults, ageism 112 140 0.27% older, older adults, adults, india, ageing 43 280 0.54% kidney, dialysis, hemodialysis, ckd, kidney disease 113 139 0.27% robots, robot, ai, social robots, social robot 44 277 0.53% long covid, covid, long, covid19, infection 114 136 0.26% disasters, disaster, radiation, earthquake, hurricane 45 277 0.53% prison, incarceration, correctional, incarcerated, prisoners 115 135 0.26% body, body image, image, appreciation, body appreciation 46 276 0.53% sexual, sexual health, dysfunction, sexuality, desire 116 135 0.26% climate, climate change, change, climate anxiety, ecoanxiety 47 275 0.53% residents, burnout, resident, wellness, residency 117 133 0.25% players, load, training, soccer, training load 48 273 0.52% loneliness, older, social, social isolation, isolation 118 133 0.25% smoking, tobacco, cessation, smoking cessation, cigarettes 49 263 0.50% apps, technology, technologies, dementia, older 119 133 0.25% ai, artificial intelligence, artificial, intelligence, intelligence ai 50 259 0.50% child, early, children, early childhood, development 120 131 0.25% incontinence, urinary, urinary incontinence, ui, pelvic 51 252 0.48% vision, eye, visual, glaucoma, vi 121 129 0.25% hsct, hematopoietic, transplantation, stem cell, hematopoietic stem 52 249 0.48% gut, microbiota, gut microbiota, microbiome, gut microbiome 122 129 0.25% psychedelic, psychedelics, psilocybin, ayahuasca, microdosing 53 248 0.47% physical activity, physical, activity, exercise, older 123 128 0.25% breast cancer, breast, cancer, cancer patients, survivors 54 245 0.47% ms, multiple sclerosis, sclerosis, multiple, pwms 124 128 0.25% tb, tuberculosis, leprosy, la, stigma 55 245 0.47% students, university, university students, covid19, pandemic 125 128 0.25% air, pollution, air pollution, air quality, pm 56 245 0.47% infertility, fertility, couples, reproductive, ivf 126 127 0.24% nanoparticles, detection, electrochemical, biosensors, sensors 57 243 0.47% palliative, palliative care, care, endoflife, death 127 126 0.24% brain injury, injury, tbi, brain, abi 58 239 0.46% work, wfh, employees, working, telework 128 125 0.24% learning, online, students, teaching, online learning 59 238 0.46% arthritis, ra, disease activity, jia, rheumatoid 129 124 0.24% adolescents, parenting, parental, adolescent, parentadolescent 60 233 0.45% violence, workplace, bullying, workplace violence, incivility 130 123 0.24% sleep, sleep quality, covid19, insomnia, pandemic 61 232 0.44% copd, asthma, pulmonary, chronic obstructive, obstructive pulmonary 131 122 0.23% prostate, prostate cancer, cancer, adt, pca 62 225 0.43% pd, parkinsons, parkinsons disease, disease, nonmotor 132 121 0.23% financial, financial wellbeing, financial literacy, fwb, debt 63 221 0.42% suicide, suicidal, ideation, suicidal ideation, selfharm 133 120 0.23% employees, covid19, work, pandemic, workers 64 220 0.42% vr, virtual, virtual reality, reality, immersive 134 120 0.23% selfcompassion, compassion, compassionate, cft, relationship selfcompassion 65 218 0.42% physical activity, activity, physical, pa, exercise 135 119 0.23% menstrual, menstruation, menstrual health, girls, premenstrual 66 216 0.41% contraceptive, marriage, women, reproductive, adolescent 136 118 0.23% local, community, policy, assets, implementation 67 215 0.41% compounds, antioxidant, bioactive, food, foods 137 118 0.23% icu, intensive care, intensive, family, family members 68 213 0.41% pharmacy, pharmacists, burnout, community pharmacists, pharmacy students 138 117 0.22% cancer, covid19, patients, pandemic, cancer patients 69 211 0.40% sleep, shift, sleep quality, night, fatigue 139 117 0.22% cp, cerebral palsy, palsy, cerebral, children cp 7 Topic Count Proportion Representation_cleaned Topic Count Proportion Representation_cleaned 140 117 0.22% design, architectural, architecture, spaces, environment 210 74 0.14% poultry, broiler, chickens, production, birds 141 116 0.22% cannabis, cbd, cannabis use, use, cannabidiol 211 73 0.14% migraine, headache, headaches, migraines, cgrp 142 115 0.22% tourism, travel, tourists, transport, cycling 212 73 0.14% light, lighting, circadian, light exposure, daylight 143 115 0.22% pregnancy, women, postpartum, weight, physical activity 213 73 0.14% gambling, problem gambling, gamblers, problem, harm 144 114 0.22% chd, congenital heart, congenital, heart disease, heart 214 72 0.14% exercise, physical activity, activity, physical, pa 145 114 0.22% dental, dental students, dentists, students, dentistry 215 72 0.14% metals, water, metal, heavy, arsenic 146 114 0.22% nurses, covid19, pandemic, frontline nurses, covid19 pandemic 216 71 0.14% dmd, muscular, sma, dystrophy, muscular dystrophy 147 113 0.22% older, older adults, agefriendly, housing, aging 217 71 0.14% amputation, limb, prosthesis, lower limb, amputees 148 112 0.21% wound, healing, tissue, hydrogel, wound healing 218 70 0.14% sexual, sexuality, cse, sexuality education, sexual wellbeing 149 111 0.21% psoriasis, psa, psoriatic, patients psoriasis, psoriatic arthritis 219 70 0.14% adsorption, wastewater, removal, water, remediation 150 111 0.21% genetic, traits, genomewide, polygenic, neuroticism 220 70 0.14% spiritual, clergy, spirituality, workplace spirituality, workplace 151 111 0.21% digital, parenting, mothers, postpartum, pregnancy 221 70 0.14% musculoskeletal, musculoskeletal disorders, workrelated musculoskeletal, ergonomics, ergonomic 152 111 0.21% penile, sexual, vaginal, hypospadias, vulvodynia 222 70 0.14% ties, friendship, network, networks, social 153 111 0.21% compassion, compassion fatigue, selfcompassion, fatigue, compassion satisfaction 223 70 0.14% euthymia, swemwbs, validity, mental, properties 154 110 0.21% adhd, adhd symptoms, disorder adhd, disorder, hyperactivity 224 69 0.14% colorectal, colorectal cancer, cancer, crc, patients 155 108 0.21% parasites, animals, animal, spp, antimicrobial 225 69 0.14% digital, digital health, technologies, healthcare, health 156 108 0.21% veterinary, veterinarians, veterinary students, veterinary medicine, profession 226 68 0.14% vaccine, vaccination, hesitancy, vaccine hesitancy, covid19 vaccine 157 107 0.20% aging, neurodegenerative, cellular, agerelated, neuronal 227 67 0.14% pesticides, fish, water, pesticide, aquatic 158 106 0.20% green, greenspace, green space, space, spaces 228 67 0.14% sexual, dating, covid19, sexual health, desire 159 105 0.20% disabilities, intellectual, people disabilities, pandemic, covid19 229 67 0.14% workplace, workplace health, miners, health promotion, wellness 160 104 0.20% financial, cancer, financial toxicity, toxicity, survivors 230 67 0.14% gardening, gardens, garden, community gardening, community 161 104 0.20% caregivers, caregiver, cancer, burden, caregiving 231 66 0.14% income, inequality, subjective, swb, subjective wellbeing 162 104 0.20% sci, spinal cord, spinal, cord, cord injury 232 66 0.14% concussion, injury, athletes, injuries, players 163 102 0.20% sexual, cancer, cervical, cervical cancer, sexual health 233 66 0.14% farmers, farming, agricultural, farm, farmer 164 101 0.19% transplant, kidney, recipients, transplantation, donors 234 65 0.14% scoliosis, lumbar, spine, surgery, spinal 165 101 0.19% machine, fetal, machine learning, learning, prediction 235 64 0.14% machine, machine learning, learning, prediction, accuracy 166 100 0.19% flourishing, wellbeing, happiness, swb, positive 236 64 0.14% personality, values, hedonic, eudaimonic, traits 167 100 0.19% breastfeeding, milk, mothers, exclusive breastfeeding, exclusive 237 64 0.14% tumor, drug, cells, microneedles, cancer 168 100 0.19% abortion, miscarriage, women, pregnancy, loss 238 64 0.14% medicaid, mothers, income, poverty, cash 169 100 0.19% medical, medical education, medicine, professional, professionalism 239 63 0.14% falls, fall, falling, older, fall prevention 170 98 0.19% head neck, neck, head, neck cancer, hnc 240 63 0.14% lifestyle, students, healthy, healthy lifestyle, habits 171 97 0.19% cells, cancer, cell, molecular, docking 241 63 0.14% driving, mobility, older, road, driving cessation 172 97 0.19% sdgs, global, health, global health, sustainable 242 63 0.14% chatbot, chatbots, conversational, cas, user 173 94 0.18% green, nature, parks, urban, park 243 62 0.14% cancer, survivors, app, mobile, digital 174 93 0.18% ad, atopic, dermatitis, atopic dermatitis, skin 244 62 0.14% workplace, mental health, employees, mental, smes 175 92 0.18% bariatric, bariatric surgery, surgery, weight, weight loss 245 62 0.14% fishing, fisheries, conservation, wildlife, fishers 176 92 0.18% gut, microbiota, microbiome, gut microbiota, axis 246 62 0.14% calves, milk, cows, colostrum, sows 177 92 0.18% icecapa, eq5d5l, eqhwb, items, eqhwbs 247 61 0.14% spiritual, spiritual care, spirituality, care, spiritual wellbeing 178 91 0.17% mice, animal, rats, animals, cage 248 60 0.14% climate, climate change, planetary, planetary health, environmental 179 90 0.17% social prescribing, prescribing, social, sp, link workers 249 60 0.14% food insecurity, insecurity, food, fi, food security 180 90 0.17% cmc, children, children medical, medical complexity, parents 250 60 0.14% housing, tenants, social housing, housing insecurity, housing tenants 181 90 0.17% sitting, sedentary, office, office workers, workplace 251 60 0.14% purpose, purpose life, sense purpose, sense, life 182 89 0.17% sarscov2, viral, antiviral, virus, viruses 252 59 0.14% fish, aquaculture, shrimp, microbiota, gut 183 88 0.17% wearable, stress, detection, accuracy, signals 253 59 0.14% brain, childhood, connectivity, neural, network 184 88 0.17% als, amyotrophic lateral, amyotrophic, lateral sclerosis, mnd 254 59 0.14% yoga, cancer, breast cancer, breast, cancer patients 185 88 0.17% dance, dancers, dancing, older adults, older 255 58 0.14% nutritional, cancer, cachexia, patients, malnutrition 186 86 0.16% marine, species, ecosystem, ocean, ecosystems 256 58 0.14% palliative, palliative care, hospice, care, burnout 187 85 0.16% athletes, sport, sports, elite, covid19 257 58 0.14% plastic, microplastics, plastics, mps, microplastic 188 84 0.16% neighborhood, cohesion, neighborhood social, social cohesion, neighbourhood 258 58 0.14% inpatient, staff, mental, psychiatric, mental health 189 84 0.16% bpd, psychotherapy, alliance, personality disorder, borderline personality 259 58 0.14% olfactory, smell, odors, odor, olfactory dysfunction 190 83 0.16% pcos, polycystic, ovary syndrome, polycystic ovary, ovary 260 57 0.14% gut, microbiota, microbiome, dogs, gut microbiome 191 83 0.16% equity, community, health equity, healthy people, health 261 57 0.14% plant, bee, soil, plants, bees 192 83 0.16% wound, ulcers, wounds, foot, ulcer 262 57 0.14% alopecia, hair, aa, aga, areata 193 81 0.16% thyroid, thyroid cancer, hypothyroidism, thyroidectomy, hypoparathyroidism 263 57 0.14% athletes, energy, intake, energy availability, performance 194 80 0.15% workers, resilience, healthcare, hcws, pandemic 264 57 0.14% twh, safety health, safety, worker, osh 195 80 0.15% gratitude, gratitude interventions, gratitude intervention, grateful, positive 265 57 0.14% subjective, subjective wellbeing, social capital, capital, china 196 80 0.15% ehr, electronic health, electronic, health record, record 197 79 0.15% pandemic, covid19, healthcare, covid19 pandemic, staff 198 78 0.15% epigenetic, methylation, dna methylation, dna, telomere 199 78 0.15% frailty, older, frail, older people, older adults 200 78 0.15% cancer, mindfulness, mindfulnessbased, cancer patients, patients 201 77 0.15% opioid, fasd, substance, substance use, use 202 76 0.15% religious, religiosity, religious coping, faith, covid19 203 75 0.14% aphasia, people aphasia, speech, pwa, communication 204 75 0.14% horses, horse, equine, welfare, equineassisted 205 74 0.14% gender, community, countries, health, chws 206 74 0.14% weight, weight stigma, obesity, stigma, weight bias 207 74 0.14% mistreatment, residents, harassment, leadership, sexual harassment 208 74 0.14% discrimination, asian, racism, black, racial 209 74 0.14% nicu, neonatal, infants, parents, neonatal intensive 8 To examine the structural relationships among topics derived from BERTopic modeling, we constructed a topic co-occurrence network based on topic probability matrices (Figure 3). For each document, topics with a probability of 0.02 or higher were selected; this empirically determined threshold balanced the inclusion of meaningful co-occurrences against excessive network density. Only documents containing at least five such topics were retained to ensure sufficient contextual overlap. Within these documents, all possible topic pairs were counted as co-occurrence candidates. Edge weights corresponded to the frequency of co-occurrence, and edges with fewer than 50 co-occurrences were removed to eliminate weak or spurious associations. Subsequently, isolated nodes were excluded, and betweenness centrality was computed to assess the relative structural influence of each topic. The network was visualized using the Kamada–Kawai layout, with node size scaled according to topic occurrence frequency and color intensity indicating the relative rank of centrality. Nodes with yellow–green hues denote higher centrality values, representing topics that serve as key intermediaries connecting multiple research domains within the wellbeing corpus. Figure 3. Topic co-occurrence network of wellbeing-related literature, visualizing the structural relationships among topics based on pointwise mutual information (PMI) weighting and betweenness centrality. The network reveals three major thematic domains—public health and pandemic response, psychological and behavioral wellbeing, and social and environmental determinants—interconnected through bridging topics such as digital health technologies and occupational wellbeing. Node size represents topic frequency, and edge thickness denotes cooccurrence strength, while network layout and cluster separation highlight both central and peripheral areas of research, ensuring interpretability even in grayscale reproduction. 9 The co-occurrence network derived from the wellbeing corpus revealed a structured and multi-layered organization linking biomedical, ecological, and social domains. At its core, a central cluster consisting of Topic 102 (cancer survivorship), Topic 123 (breast cancer), Topic 252 (fish, aquaculture, gut microbiota), and Topic 261 (plants, bees, soil) exhibited the highest betweenness centrality. These topics function as pivotal hubs bridging medical sciences, environmental research, and biological systems. Despite originating from distinct disciplinary contexts, they converge on a shared conceptual foundation—the maintenance, adaptation, and restoration of life and health. Topics 102 and 123 emphasize the processes of recovery and quality of life among patients and survivors, whereas Topics 252 and 261 focus on ecological resilience, biodiversity, and sustainable resource systems. Their central co-occurrence suggests an emerging synthesis between human health and ecological wellbeing within the broader conceptual landscape of wellbeing research. Surrounding this core is a cluster centered on women’s health, childbirth, and parenting, composed of Topics 168 (abortion and pregnancy loss), 18 (childbirth and postpartum), 209 (neonatal intensive care), 167 (breastfeeding), and 151 (digital parenting and postpartum support). Together, these topics delineate a continuous thematic spectrum from pregnancy to early childcare, capturing the physiological, emotional, and social dimensions of women’s wellbeing.