Our primary objective was to systematically map how research on the neural mechanisms of physical activity interventions for depressive disorder has evolved over time and to identify the evolution of key research themes through the use of a network of co-cited references and a network of co-occurring keywords, and secondary objectives were to provide clinicians and researchers with metrics of the research network (countries, organisations, authors, and journals) and to detect research richness, gaps, emerging trends, biases, and limitations.

We searched the Web of Science Core Collection (WOSCC) as the most comprehensive database for scientometric analyses (22).

A comprehensive search strategy was employed using both MeSH (Exercise, Depressive Disorders, and Neurobiology Subject Headings) terms and keywords. The MeSH terms were selected based on their relevance to the research topic, ensuring a broad coverage of relevant literature. The full search terms were provided in the study protocol ((TS=(Exercises) OR TS=(Exercise, Physical) OR TS=(Exercises, Physical) OR TS=(Physical Exercise) OR TS=(Physical Exercise) OR TS=(Physical Exercises) OR TS=(Physical Exercises) OR TS=(Physical Activity) OR TS=(Physical Activity) OR TS=(Physical Activity) OR TS=(Physical Activity) OR TS= (Physical Activity) OR TS=(Activities, Physical) OR TS=(Activity, Physical) OR TS=(Physical Activities) OR TS=(Physical Activities) OR TS=(Exercise, Aerobic) OR TS=(Aerobic Exercise) OR TS=(Aerobic Exercises) OR TS=(Exercises, Aerobic) OR TS=(Exercise, Isometric) OR TS=(Exercises, Isometric) OR TS=(Isometric) Exercises) OR TS=(Isometric Exercise) OR TS=(Acute Exercise) OR TS=(Acute Exercises) OR TS=(Exercise, Acute) OR TS=(Exercises, Acute) OR TS=(Exercises, Acute) OR TS=(Exercise Training) OR TS=(Exercise Trainings)) And ((TS=(Depressive Disorders) OR TS=(Disorder, Depressive) OR TS=(Disorders, Depressive) OR TS=(Disorders, Depressive) OR TS=(Neurosis, Depressive) OR TS=(Depressive Neuroses) OR TS=(Depressive Neurosis) OR TS=(Neuroses, Depressive) OR TS=(Depression, Endogenous) OR TS=(Depressions, Endogenous) OR TS=(Endogenous Depression) OR TS=(Endogenous Depressions) OR TS=(Melancholia) OR TS=(Melancholias) OR TS=(Unipolar Depression) OR TS=(Depression, Unipolar) OR TS=(Depressions, Unipolar) OR TS=(Unipolar Depressions) OR TS=(Depressive Syndrome) OR TS=(Depressive Syndromes) OR TS=(Syndrome, Depressive) OR TS=(Syndromes, Depressive) OR TS=(Depression, Neurotic) OR TS=(Depressions, Neurotic) OR TS=(Depressions, Neurotic) OR TS=(Neurotic Depression) OR TS=(Neurotic Depressions)) AND ((TS=(neurobiology) OR TS=(brain function) OR TS=(neurophysiology) OR TS=(neuroscience) OR TS=(biological mechanisms) OR TS=(neurotransmitters) OR TS=(brain plasticity) OR TS=(brain-derived neurotrophic factor) OR TS=(amygdala) OR TS = (prefrontal cortex)). The source of the database was restricted to WOS, the publication type was ‘article’, the language type was English, and WOSCC extracted complete records with cited references published up to 31 December 2024 and converted them into tab-delimited plain text files. Duplicates were eliminated using CiteSpace.

Once the search was conducted, the extraction and the reasons reported for excluding articles were processed in the flowchart ( Figure 1 ). The screening process was conducted by two independent reviewers. First, titles and abstracts of the identified articles were assessed to determine relevance. After an initial screening, 2 non-English articles were excluded, leaving 367 articles for further assessment. During the second phase, the full texts of these articles were reviewed, and 197 studies were excluded because they were not relevant to the topic of neural mechanisms of physical activity interventions in depression. Ultimately, 170 studies were eligible and included in the review. Discrepancies between reviewers were resolved through discussion and consensus.

The study used the bibliometric software CiteSpace (version 6.4.R1) (9) to perform a thorough analysis of scientific literature. The analysis examined various aspects, such as author contributions, journal impact, reference networks, international collaborations, institutional affiliations, and patterns of keyword co-occurrence. To create the scientific knowledge graph, the network topology was divided into two main types based on relationship structures: directed graphs, which include direct citation networks that show asymmetric citation relationships among documents, and undirected graphs, which consist of co-citation and co-occurrence networks that illustrate symmetric relationships between elements. A systematic comparative analysis by Boyack and Klavans (2010) emphasized the differences between direct citation and co-citation networks in how they represent knowledge structures (23).

The bibliometric results included citation counts, co-citation, and co-occurrence metrics. Co-citation count refers to how often two published works are cited together in later articles (24). In this study, three bibliometric indicators were used: citation frequency, which reflects the influence of literature; co-citation intensity, which indicates the strength of knowledge connections; and co-occurrence frequency, which shows the degree of conceptual relationship. The co-citation network has a unique advantage in systematic reviews due to its dynamic nature, which effectively tracks the evolution of knowledge communities (25). At the same time, the co-occurrence network clarifies relationships among variables using graph-theoretic methods.

The CiteSpace platform offers various dimensional metrics to effectively evaluate network characteristics. It initially employed temporal indicators, particularly Citation Burst, to track the chronological development of research focuses. Following this, structural indicators were used to evaluate how well nodes connect within the network. Nodes with high centrality often serve as vital knowledge hubs that link different clusters. Additionally, composite indicators such as Modularity Q and Silhouette S metrics (26) were integrated to assess the overall structure of the network. The Q-value, which ranges from 0 to 1, indicates network modularity, with values above 0.3 suggesting a significant clustering structure. In contrast, the S-value, which ranges from -1 to 1, measures the homogeneity of clusters; thresholds of 0.3, 0.5, and 0.7 reflect increasing confidence in the clustering. An S-value of 1 may indicate that a cluster is isolated, necessitating careful interpretation within the specific domain context.

Bibliometric techniques were utilized to analyze scientific literature from various perspectives. For cluster labeling, a system was created by extracting noun phrases from the keyword lists of each literature cluster, based on significance determined by likelihood ratio tests (p < 0.001). Data collection and analysis adhered to standardized protocols: CiteSpace version 6.4.R1 was employed for multidimensional knowledge graph analysis, concentrating on international collaboration networks at both country and institutional levels, co-citation networks that highlight author and literature clustering, and author keyword co-occurrence networks. Furthermore, journal impact factor data were sourced from the 2023 Journal Citation Reports (JCR) of the Web of Science, in terms of index calculation, we adopt the g-index (g-index) as the core index for author impact assessment, which, by integrating the citation distribution characteristics (27), effectively corrects the over-sensitivity of the h-index to highly cited papers, and ensures the weight of high-impact papers while taking into account the low/zero-citation literature’s CiteSpace’s time slicing function removes empty time intervals through an automatic optimisation algorithm to achieve intelligent compression of the analysis time domain.

To investigate research focal points and trends, we conducted an analysis of the most frequently cited keywords by generating a co-occurring keywords network diagram using CiteSpace, as illustrated in Figure 4A . This analysis revealed nine keyword clusters, with the most prominent ones being “major depression,” “aerobic exercise,” and “brain-derived neurotrophic factor” during the period from 2006 to 2014, referred to as cluster A. In the following period from 2015 to 2024, represented as cluster B, a similar network diagram shown in Figure 4B displayed eight clusters, which included “Major Depressive Disorder,” “Physical Activity,” “Aerobic Exercise,” “Adaptable Depression,” “Step-down Task,” “Hippocampal Neurogenesis,” and “Stress-induced Depression Model.” Both keyword networks from the two periods demonstrated significant silhouette scores greater than 0.6 and satisfactory modularity scores, indicating a robust structure.

Additionally, burst analysis depicted in Figure 5 identified the six most vigorous keywords at the onset of their citation surges: “physical exercise” with an intensity of 4.23, “stress” at 3.38, “major depression” at 2.62, “BDNF” at 2.43, “meta-analysis” at 2.57, and “functional connectivity” at 2.25. Furthermore, the keyword “augmentation treatment” showed a low burst intensity of 1.39 during the period from 2015 to 2018, suggesting a slow advancement or ongoing debates in this area, particularly regarding drug combination therapies. The term “cell proliferation” experienced a brief burst from 2011 to 2013 with an intensity of 1.75, likely due to technical constraints, such as challenges in in vivo tracking, which led to diminished interest afterward. Similarly, “cognitive impairment” had an early burst from 2007 to 2008 with an intensity of 1.17 that was not sustained, likely reflecting a shift in focus toward emotional symptoms like depression and anxiety.

An examination of the evolution in the field, as illustrated by the emergence of keywords, reveals that from 2006 to 2014, the primary focus was on exploring fundamental mechanisms and interventions, with “major depression” highlighting biological processes such as the monoamine hypothesis and the development of pharmaceuticals. By around 2015, “physical exercise” shifted from being a peripheral topic to a recognized treatment option, supported by increasing clinical evidence. Between 2015 and 2020, the field began to embrace precision and interdisciplinary approaches. The years 2015 to 2017 were characterized by the standardization of research methodologies through meta-analyses and a push for evidence-based medicine regarding exercise interventions.

By 2022, advancements in neuroimaging techniques, such as fMRI, started to reveal brain network irregularities associated with “adult neurogenesis.” More recently, from 2021 to 2024, the focus has shifted towards “associations,” reflecting the rise of multi-omics, including genomic and epigenomic studies, as well as phenotypic investigations. Together, these emerging keywords illustrate the field’s progression from investigating pathological mechanisms to developing precise interventions. Future research should emphasize technology-driven, multimodal studies that explore brain networks and biomarkers, alongside clinical translational applications like personalized exercise prescriptions, while also addressing unresolved challenges such as the variability in the effectiveness of augmentation therapies.

Our initial dataset included 337 references, but after applying the data filtering process described in the protocol, we excluded 50% of these references, resulting in a final dataset of 170 studies published between 2006 and 2024. These studies are all original articles, with an average of 6.36 authors per publication. Notably, the average number of co-authors increased from 5.8 during the period from 2006 to 2014 to 6.5 from 2015 to 2025. The earliest study in our dataset, conducted by Sarbadhikari (28), proposed and experimentally tested the hypothesis that moderate exercise could improve depressive symptoms by enhancing the activity of specific neurotransmitters.

Figures 6A, B depict the evolution of publication outputs over time. In the initial period from 2006 to 2009, the annual number of publications was relatively low, ranging from 1 to 3 per year. However, starting in 2010, the number of publications began to fluctuate and increase, reaching peaks of 13 articles in 2015, 16 articles in 2019, and 23 articles in 2023. This sustained growth, particularly after 2015, is likely attributed to technological advancements, increased financial support, and enhanced academic collaboration. Interestingly, the number of citations did not always align with the volume of publications; for example, in 2016, there were high citation counts despite a lack of a corresponding increase in articles, indicating the significant impact of certain studies. Conversely, the high number of articles published in 2023 was associated with relatively low citation counts.

Table 3 provides an analysis of the journals referenced in this study. The five most cited journals include Elife Sciences Publ Ltd, which leads with 498 citations, followed by Molecular Psychiatry with 348 citations, and both the Journal of Neuroscience and the Journal of Biological Chemistry, each with 316 citations. Additionally, Evidence-Based Complementary and Alternative Medicine has garnered 139 citations. In terms of publication volume, the Journal of Affective Disorders, established in 1979, and the Journal of Psychiatric Research, founded in 1966, are tied for the highest contribution, each accounting for 4.7% of the total articles, which amounts to 8 articles, with impact factors of 4.9 and 3.7, respectively.

Moreover, Frontiers in Psychiatry, which began publishing in 2010, has quickly gained prominence, contributing 7 articles, or 4.1% of the total, with an impact factor of 3.2. Among the high-impact journals, Translational Psychiatry, founded in 2011 and boasting an impact factor of 5.8, has a relatively low publication rate, representing only 2.3% of the total publications with 4 articles, highlighting its commitment to high-quality research. In contrast, Pharmacology Biochemistry and Behavior, with an impact factor of 3.3, achieves a higher impact factor than some journals with greater publication volumes by focusing on specialized niche areas.

As shown in Figure 7A , 43 countries have contributed to this field. Analysis of the frequency and centrality of these countries reveals that the United States (USA) has the highest frequency (45 occurrences) and centrality (0.31), underscoring its central role and high level of participation in the network. In contrast, although China and Germany show relatively high frequencies (22 and 19, respectively), their centrality values (0.14 and 0.05) are lower, suggesting that their contributions are less well connected within the overall network. Switzerland, with a frequency of 7 but a centrality of 0.09, plays a key role in connectivity, while Sweden, with 6 occurrences and a centrality of 0.04, has relatively weak participation and influence (see Table 4 ).

Figure 7B indicates that a total of 446 organizations have contributed articles in this field. The institutions with the highest citation counts include Tehran University of Medical Sciences with 9 citations, Islamic Azad University with 8, Harvard University with 6, and both Humboldt University of Berlin and Karolinska Institutet with 5 citations each, as detailed in Table 5 . An analysis of citation bursts over the past decade highlights that England exhibited the strongest citation burst (strength 2.03) from 2016 to 2019, followed by Iran (strength 1.55) from 2020 to 2024, Portugal (strength 1.57) from 2015 to 2016, and China (strength 1.22) from 2012 to 2021. These patterns suggest a consistent growth in research output, influenced by national initiatives like China’s “double first-class” program.

Among institutions, notable citation bursts were recorded for the University of Texas System (strength 1.61, 2011–2013) and the University of Texas at Urbana-Champaign (strength 2.03, 2016–2019), likely linked to significant advancements in cancer research and neuroscience, bolstered by NIH funding. Additionally, Karolinska Institutet (strength 1.57, 2021–2024) experienced a burst likely tied to research on COVID-19 vaccines and Nobel Prize-related studies, while King’s College London (strength 1.56, 2017–2019) has concentrated on psychiatry and neuroscience, particularly brain imaging studies on depression, supported by the UK Brain Programme. More comprehensive information can be found in Figures 8 and 9 .

In this study, we developed a collaborative network of authors based on their co-authorship of papers that explore the neural mechanisms underlying physical exercise interventions in depression, as illustrated in Figure 10 . We utilized an emergence detection technique to analyze these authors’ collaborative networks, their research activity, and their geographical distribution. The findings, presented in Figure 11 , highlight prominent authors in this field, such as Carneiro, Lara S F (2015, intensity 117) and Gongpoweis, Joanne (2017, intensity 121), whose collaborative efforts were notably concentrated between 2015 and 2018. Furthermore, authors Brand, Serge from the University of Basel and Gerber, Markus showed significant publication outputs in 2014, with 6 and 5 papers respectively, as shown in Table 6 . However, their citation burst intensity is relatively low, with Brand’s intensity at only 0.58. This suggests that their research may focus more on long-term academic contributions rather than generating immediate, short-term influence.

Since 2019, scholars such as Bahmani, Dona Sadeghi, and Agariz, Ingrid have demonstrated notable citation burst intensities, ranging from 1.03 to 1.12, while their collaborative networks have expanded across various institutions, including the Free University of Berlin, represented by researchers like Heinzel, and Islamic Azad University, showcasing a trend of diversified geographical cooperation. The cooperation pattern is predominantly European, with key institutions such as the University of Basel in Switzerland, the Free University of Berlin in Germany, and the University of the Basque Country in Spain acting as central hubs.

A significant aspect of this collaboration is the long-term partnership between Brand and Gerber, which has been ongoing since 2014, highlighting the strong academic connections between Switzerland and Germany. Additionally, institutions in the Berlin area, including Freie Universität Berlin and Humboldt Universität, have emerged as new cooperative nodes since 2019, likely influenced by Heinzel’s involvement. Furthermore, the participation of the Texas A&M University System in the USA and Chongqing Medical University in China illustrates a growing trend of cross-regional collaboration, emphasizing the increasing significance of cooperation among Europe, America, and Asia, and reflecting the internationalization of research efforts in this field.

This scientometric analysis systematically mapped the evolving neurobiological mechanisms linking exercise and depression across 2006-2024, revealing a clear three-phase progression: neurotransmitter-focused studies (2006–2014), pharmacological comparison trials (2015–2019), and precision intervention approaches (post-2020). The co-citation network analysis identified seven research clusters (Q=0.9299, S=0.9794) illustrating the field’s evolution from broad neurochemical investigations to molecular-level mechanisms, particularly emphasizing brain-derived neurotrophic factor (BDNF) signaling as a central pathway mediating exercise’s antidepressant effects (42). These findings provide clinicians with a roadmap for evidence-based exercise prescription, transitioning from generic recommendations to biomarker-guided, personalized interventions.

One notable limitation of our study is the exclusive reliance on the Web of Science Core Collection, which may introduce citation bias by potentially excluding relevant studies from emerging journals or underrepresented regions. Additionally, the inherent data update lag may lead to omission of the most recent publications, affecting temporal analysis accuracy and identification of emerging trends.

To enhance representativeness and robustness of future bibliometric analyses, we propose integrating additional databases such as Scopus and PubMed. This multi-database approach would provide a more comprehensive literature pool, mitigate citation bias, and ensure analysis reflects current field developments (10).

Future research should prioritize multi-modal approaches integrating neuroimaging, metabolomics, and genomics to elucidate dose-response relationships and optimize exercise parameters. The development of smart digital intervention systems capable of providing real-time feedback and dynamically adjusting exercise protocols represents a tangible strategy for translating neurobiological insights into clinical practice. While our study provides foundational mapping of the field’s evolution, targeted clinical trials testing these precision intervention approaches are needed to validate their therapeutic efficacy.