The evolving landscape of inflammation-related clinical trials: a 25-year registry-based network analysis from a clinical pharmacology perspective

Abstract

The longitudinal evolution of inflammation-related clinical research has not been systematically characterized at the registry level, and the differential impact of the coronavirus disease 2019 (COVID-19) pandemic across inflammatory disease domains remains incompletely resolved. We analyzed 15,051 inflammation-related interventional trials registered on ClinicalTrials.gov (2000–2025), stratified into pre-pandemic (≤ 2019; n = 10,366), pandemic (2020–2022; n = 2,356), and post-pandemic (2023–2025; n = 2,329) epochs. We quantified volumetric, phase, and disease-category metrics; assessed thematic evolution using temporally stratified keyword co-occurrence networks; and characterized the static disease-mechanism architecture through bipartite network analysis using the H2′ specialization and Nestedness based on Overlap and Decreasing Fill (NODF) nestedness indices. Annual registration sustained near pandemic-era rates through the post-pandemic window, peaking in 2025 (871 trials), alongside a sustained shift toward earlier-stage development. Disease-category composition reorganized into two distinct trajectories: a transient pulmonary surge (13.5%→21.4%→15.3%) and a sustained oncology expansion (12.3%→16.9%→21.3%; 73.2% relative increase), with oncology overtaking rheumatology as the largest category by 2023–2025. Co-occurrence networks underwent structural simplification (density 0.426→0.247; modularity 0.527→0.393), while the betweenness centrality of immune increased monotonically (0.171→0.333), identifying it as a progressively load-bearing bridging concept. Bipartite analysis showed moderate specialization (H2′ = 0.415) and substantial nestedness (NODF = 51.07), with cross-domain coupling of immunologic and oncologic vocabularies. Formal artificial intelligence and machine learning (AI/ML) incorporation remained negligible (0.14%). COVID-19 functioned as a discrete inflection point amplifying a longer-term reorientation of inflammation research toward oncology-anchored, immune-centered investigation; AI/ML methodologies have not yet translated into prospective trial design.

INTRODUCTION

Chronic inflammatory conditions, driven by dysregulated immune activation, represent a growing contributor to global disease burden, with epidemiological analyses documenting rising prevalence worldwide [1,2]. According to the Global Burden of Disease Study 2021, the age-standardized incidence rate across six major immune-mediated inflammatory conditions reached 823.28 per 100,000 population [3]. These conditions also drive substantial therapeutic development, spanning small-molecule agents and biologics targeting cytokine pathways, immune checkpoints, and inflammatory mediators, making the inflammation trial landscape a uniquely informative window into modern clinical pharmacology priorities.

These conditions share common pathophysiological mechanisms such as aberrant cytokine signaling and chronic tissue inflammation, yet they differ markedly in clinical trial design [3]. Tracking how research evolves across heterogeneous categories requires methods capturing multidimensional patterns over time, beyond simple trial counts. Keyword co-occurrence network analysis is well suited to this task: trial title terms serve as nodes and their pairwise co-occurrence as edges, providing a structured way to visualize how dominant topics emerge, persist, or recede across successive periods [4,5]. Applied to a large registry, this approach reveals thematic transitions that count-based analyses tend to obscure.

Prior inflammation-related registry analyses, however, have largely been confined to specific disease areas or limited time windows [6,7]. A comprehensive longitudinal view of the entire inflammation trial ecosystem is therefore still lacking, and the pandemic’s differential impact across disease categories has not been examined using network-based approaches. The present study addresses these gaps by applying network analysis across heterogeneous categories. Such an analysis is meaningful because the composition of registered trials at any given moment reflects the scientific priorities, regulatory frameworks, and unmet medical needs of that era.

The coronavirus disease 2019 (COVID-19) pandemic offers a particularly informative perturbation, as its onset was a discrete, globally synchronized shock within a compressed timeframe. Following the World Health Organization’s pandemic declaration on March 11, 2020, clinical research infrastructure was disrupted on an unprecedented scale; registry-based analyses of ClinicalTrials.gov, after excluding COVID-19 studies, show a deficit of up to 5,000 ongoing non-COVID trials in 2020 relative to pre-pandemic forecasts [8,9,10]. COVID-19 itself simultaneously became a major inflammatory disease paradigm, as cytokine release syndrome, post-acute pulmonary fibrosis, and immune dysregulation drew fresh research interest at the intersection of infectious disease and inflammation biology. Registration behavior in the ClinicalTrials.gov registry also shows a policy-driven discontinuity around 2005 following implementation of the International Committee of Medical Journal Editors (ICMJE) prospective registration requirement; this is addressed through a sensitivity analysis restricting the pre-pandemic epoch to 2005–2019.

Another contemporary feature is the incorporation of artificial intelligence and machine learning (AI/ML) [11,12,13,14]. The extent to which these methods have been formally adopted in registered interventional trials for inflammatory diseases—as opposed to being discussed in literature—remains unquantified, providing a third axis for characterizing landscape evolution. The present study aimed to characterize the evolving landscape of inflammation-related interventional trials registered on ClinicalTrials.gov from 2000 to 2025. We quantified longitudinal registration and category trends across pandemic-defined epochs, mapped thematic evolution via keyword co-occurrence and bipartite network analyses, and tracked AI/ML incorporation.

METHODS

Data source

Trial data were obtained from the Aggregate Analysis of ClinicalTrials.gov (AACT) database, a publicly available relational database maintained by the Clinical Trials Transformation Initiative that is synchronized daily with the ClinicalTrials.gov application programming interface [15]. AACT provides structured access to all protocol and result elements, including trial characteristics (phase, sponsor, enrollment, status), eligibility criteria, intervention details, outcome measures, and condition terms. Medical Subject Headings (MeSH) descriptors are additionally embedded through the browse_conditions table, supporting ontology-based categorization independent of free-text variability. A static snapshot dated March 8, 2026 was used as a fixed reference dataset for all analyses (Fig. 1). This study analyzed only publicly available, fully de-identified aggregate registration records and did not involve human subjects, individual patient-level data, or any interaction with human participants or animals. It therefore does not constitute human subjects research under the Declaration of Helsinki, and Institutional Review Board review and informed consent were not applicable.

Figure 1. Overall study scheme. Trial data were obtained from Aggregate Analysis of ClinicalTrials.gov database. After applying inflammation-related keyword filters to tile fields, the analytic cohort of 15,051 trials was stratified into three pandemic-defined epochs.

Created in BioRender. Lee S. (2026) https://BioRender.com/l33nmqd.

Eligibility criteria

The analytic window spanned first-submission dates (study_first_submitted_date field) from January 2000 to December 2025. Only interventional studies were retained; observational studies, expanded access records, and patient registries were excluded. Inflammation-related trials were identified by applying a curated set of 18 inflammation-related terms (e.g., inflammation, inflammatory, autoimmune, colitis; Supplementary Data 1) to the brief_title and official_title fields using case-insensitive regular expression matching without explicit word-boundary constraints; a trial was classified as inflammation-related if at least one term appeared in either title field. Because the term set independently enumerates all major lexical variants (e.g., immune, autoimmune, immunotherapy, immunology, neuroinflammation, systemic inflammation), the absence of word boundaries produces overlap between related terms but does not introduce spurious inclusions. Title fields were chosen over broader conditions or detailed_description fields to maximize specificity, since inflammation-related terminology in titles typically signals the primary therapeutic focus. Of 23,975 inflammation-titled interventional candidates, 8,525 were excluded due to missing, Not Applicable, Early Phase 1, or unclassifiable phase designations, and a further 399 were excluded due to first-submission dates outside the 2000–2025 window. This yielded the final analytic cohort of 15,051 inflammation-titled interventional trials with complete and classifiable phase designations. The cohort was stratified into three epochs by first-submission year: pre-pandemic (≤ 2019), pandemic (2020–2022), and post-pandemic (2023–2025).

Keyword identification

Each trial was assigned to one of 18 predefined disease categories through a rule-based curation process applied to MeSH terms from the downcase_mesh_term field of the browse_conditions table; when MeSH annotations were unavailable, free-text entries from the conditions field were used as supplementary input [15]. When a trial’s MeSH profile spanned multiple categories, assignment followed a clinical specificity principle: disease-specific categories (e.g., rheumatology, inflammatory bowel disease [IBD]) took precedence over broader or residual categories (e.g., other), ensuring mutually exclusive single-category assignment. The 18 categories were: rheumatology, IBD, pulmonary, dermatology, infection/sepsis, neuroinflammation, metabolic, cardiovascular, oncology, hepato-renal, periodontal/oral, allergy, hematology, gynecology, urology, ophthalmology, musculoskeletal, and other. Complete mapping rules are provided in Supplementary Table 1.

Keyword vocabularies for network construction were derived in two ways. For the co-occurrence network, per-epoch keyword extraction was performed on a concatenated text field (brief_title, official_title, conditions) after lowercase normalization. Within each epoch, the top 22 keywords by raw frequency were retained after applying a stopword list that excluded generic clinical research terminology (e.g., therapy, disease, treatment, intervention, parallel, multi) carrying limited thematic discriminability within an inflammation-specific corpus. The union of epoch-specific keyword sets formed the combined vocabulary.

AI/ML incorporation was assessed by text mining across five trial text fields (brief_title, official_title, brief_summary, detailed_description, conditions) using a predefined set of 11 AI/ML-related keywords: machine learning, deep learning, artificial intelligence, prediction model, neural network, natural language processing, NLP, random forest, support vector, classification model, and clustering. Case-insensitive regular expression matching was applied with word-boundary constraints. The term algorithm was initially included but excluded after systematic manual review of all 58 matching trials, which uniformly reflected non-AI/ML clinical usages (treatment algorithms, dose-escalation schemas, randomization procedures). Trials were classified as AI/ML-incorporated if at least one keyword matched in any of the five text fields. The complete keyword list and search strategy are provided in Supplementary Data 2.

Network representation

Two complementary networks were constructed to characterize the structure and temporal evolution of inflammation research: a disease–mechanism/treatment bipartite network capturing the static organization of the field over the full cohort, and a temporally stratified keyword co-occurrence network capturing thematic change across epochs. In both, nodes represented keywords extracted from each trial's concatenated text field, and edges represented co-occurrence of two keywords within the same field, with edge weight scaled by co-occurrence frequency and node size by keyword frequency. Keyword category was encoded by node fill color: Disease/Condition (red), Biological/Mechanism (blue), Treatment/Drug (teal), and Other (gray). All network construction, layout, and metric computation were performed using the NetworkX library in Python, and all layouts used the Fruchterman–Reingold force-directed algorithm with a fixed random seed to ensure reproducibility [16,17].

The bipartite network used three predefined keyword groups: a disease/condition vocabulary (e.g., arthritis, colitis, cancer, COVID; n = 45 terms), a biological mechanism vocabulary (e.g., cytokine, interleukin, antibody, receptor; n = 14 terms), and a treatment/drug vocabulary (e.g., immunotherapy, methotrexate, chemotherapy, vaccine; n = 10 terms); the mechanism and treatment vocabularies were treated as a combined mechanism/treatment set during network construction (Supplementary Table 2).

Bipartite network and community detection

A weighted undirected bipartite graph was constructed across the full cohort (2000–2025). One partition represented disease/condition terms; the other represented mechanism/treatment terms (biological mechanisms and treatments/drugs). An edge connected a disease and a mechanism/treatment term whenever both co-occurred within the same trial's text field, with edge weight equal to the number of co-occurring trials. Nodes were retained only if the corresponding keyword appeared in ≥ 0.2% of all trials, and edges only if the disease–mechanism pair co-occurred in ≥ 0.1% of all trials; isolated nodes were removed.

Two one-mode projections were derived to resolve internal structure on each side. Rather than weighting projected edges by raw number of shared partners, each shared partner was weighted inversely to its degree following Newman’s collaboration weighting [18]. Community structure was identified on the weighted projections by modularity optimization using the Leiden algorithm, with the Louvain algorithm applied in parallel for consistency [19,20]. The random seed was fixed (seed = 7) across all runs. For visualization, the top 60% (disease-side) and top 50% (mechanism/treatment-side) of edges by weight were retained to preserve dominant relational structure while maintaining readability (Fig. 2).

Figure 2. Community structure of disease-mechanism/treatment bipartite projections in inflammation-related interventional clinical trials (2000–2025). (A) Disease-side projection. Two disease terms are connected when they share mechanism/treatment partners, with edge weights computed via Newman’s collaboration weighting [18] to downweight shared partners proportional to their degree. Three communities were identified by the Leiden algorithm (modularity Q = 0.0426; resolution = 1.0; seed = 7; stability confirmed across 10 random seeds): an immune-related disorder cluster (red shading; n = 11: cancer, carcinoma, lymphoma, leukemia, melanoma, coronavirus disease, hepatitis, fibrosis, sclerosis, asthma, rhinitis), an autoimmune/chronic inflammatory cluster (purple shading; n = 9: arthritis, rheumatoid, lupus, autoimmune, colitis, dermatitis, atopic, diabetes, sepsis), and a singleton (osteoarthritis). (B) Mechanism/treatment-side projection. Two mechanism/treatment terms are connected when they share disease partners, weighted by the same Newman scheme. Three communities were identified (modularity Q = 0.0627): a targeted biologic and immune-modulation cluster (red shading; n = 9: antibody, biologic, monoclonal, receptor, immune, methotrexate, steroid, vaccine, antiviral), an oncology-cytokine pathway cluster (purple shading; n = 7: cytokine, interleukin, interferon, chemotherapy, immunotherapy, radiotherapy, transplant), and a singleton (corticosteroid). For visualization clarity, only the top 60% (A) and top 50% (B) of edges by weight are displayed; this preserves dominant relational structure while reducing visual clutter from weak ties. Node size is proportional to the number of trials containing the keyword. Node fill color denotes category (red, disease/condition; blue, biological/mechanism; teal, treatment/drug). Node positions were determined by the Fruchterman–Reingold force-directed layout (seed = 7).

Node-level structural roles were characterized by degree centrality, identifying terms with the broadest co-occurrence reach, and betweenness centrality, identifying terms lying on shortest paths between otherwise distinct communities (structural bridges). For the bipartite network, bipartite-aware degree centrality was computed for each partition, together with two bipartite-specific indices on the disease × mechanism/treatment biadjacency matrix: the H2′ specialization index, derived from the Shannon entropy of observed interaction frequencies (0 = full generalist, 1 = full specialization), and the Nestedness based on Overlap and Decreasing Fill (NODF), computed on the binary incidence matrix (0–100; higher values indicate stronger nestedness) [21,22].

Temporal co-occurrence network

To track thematic change over time, separate keyword co-occurrence networks were constructed for three pandemic-defined epochs (~2019 [2000–2019], 2020–2022, and 2023–2025). Pairwise co-occurrence counts were computed for keyword pairs co-appearing within the same trial text field, and a weighted undirected graph was built for each epoch in which nodes represented keywords (sized by frequency) and edges represented co-occurrence (thickness scaled by frequency). Vocabulary was determined by the union of the top-22 keywords per epoch by raw frequency (minimum per-epoch top-22 count > 140 in all epochs), implicitly excluding low-frequency terms. Edges were retained if co-occurrence count was ≥ 3; this absolute threshold preserves comparability of raw counts across epochs but corresponds to different relative proportions (0.03% pre-pandemic vs. 0.13% pandemic and post-pandemic). Node sizes and edge weights were expressed in raw counts without normalization.

The integrated temporal evolution was rendered as a Sankey diagram, in which three columns represented the three epochs and vertical block heights reflected raw keyword frequency within each epoch; flows between adjacent epochs were proportional to the shared keyword frequency and colored by category (Fig. 3). The top 25 keywords by total frequency were displayed, with keywords appearing in fewer than 5 trials in a given epoch omitted from that column. Period-specific networks were additionally rendered as small multiples on a shared force-directed layout computed once on the union network, enabling direct visual comparison across epochs (Supplementary Fig. 1). Modularity was computed for each epoch network via the Louvain algorithm to index thematic clustering over time.

Figure 3. Temporal flow of inflammation-related keywords across pandemic-defined epochs. Sankey diagram visualizing the longitudinal trajectory of the top 25 inflammation-related keywords across three pandemic-defined epochs (pre-pandemic, ≤ 2019; pandemic, 2020–2022; post-pandemic, 2023–2025). Keywords were extracted from the trial title and conditions fields of all 15,051 interventional trials after stopword removal, and selected as the union of the most frequent terms in each epoch followed by re-ranking by total frequency across the entire study period. Each vertical block represents a keyword in a given epoch, with block height proportional to the raw number of trials containing that keyword in that epoch (no frequency normalization). Block fill color denotes the keyword category (red, disease/condition; blue, biological/mechanism; teal, treatment/drug; gray, other). Flow ribbons connect the same keyword across consecutive epochs; ribbon thickness is proportional to the smaller of the two epoch-specific frequencies, representing the carried-over usage. Within each epoch, keywords are vertically ordered by category (disease, biological, treatment, other) and by raw frequency within each category. Keywords with fewer than 5 trials in a given epoch are not shown for that epoch; absence of an incoming or outgoing ribbon for a keyword indicates that the keyword fell below the threshold or was absent in the adjacent epoch.

Statistical analysis

Categorical variables are reported as counts and proportions. Annual trends in trial registration volume were visualized using phase-stratified stacked bar charts, with the COVID-19 pandemic period (2020–2022) highlighted by a shaded band (Fig. 4). For phase-stratified analyses, AACT phase designations were standardized as follows: phase 1/2 was reclassified as phase 1, and phase 2/3 as phase 2. Changes in disease category proportions across epochs were expressed as relative percentage changes, calculated as [(Period B% − Period A%)/Period A%] × 100, and visualized using a heatmap with a diverging color scale (Fig. 5).

Figure 4. Annual trends in inflammation-related interventional clinical trials registered on ClinicalTrials.gov by phase (2000–2025). Stacked bars represent the annual number of inflammation-titled interventional trials registered on ClinicalTrials.gov, stratified by clinical trial phase. Trials designated as Phase 1/2 were reclassified as Phase 1, and Phase 2/3 as Phase 2; trials with missing, Not Applicable, Early Phase 1, or unclassifiable phase designations were excluded prior to cohort assembly. Numbers above each bar indicate the total annual trial count. Shaded background regions denote the COVID-19 pandemic period (2020–2022, red) and the post-pandemic period (2023–2025, blue). Annual counts rose monotonically across 2023–2025 (691→767→871), with 2025 representing the highest single-year count of the 25-year series.

COVID-19, coronavirus disease 2019.

Figure 5. Temporal distribution of inflammation-related interventional clinical trials across disease categories by pandemic-defined period (2000–2025). Each cell represents the proportion (%) of trials within a given disease category relative to the total number of trials registered in the corresponding period, with absolute trial counts shown in parentheses. Color intensity reflects the magnitude of proportional representation within each period. Disease categories were assigned based on Medical Subject Headings terminology following a predefined priority hierarchy.

IBD, inflammatory bowel disease.

Inter-epoch similarity of the co-occurrence networks was quantified by the Jaccard coefficient, defined as the ratio of shared edges to the union of edge sets between two networks (1.00 = identical; 0 = disjoint), computed for all three epoch pairs. Two quantities were reported: the observed similarity of the full-cohort networks (point estimate) and the expected similarity at a common sample size of n = 2,329 trials per epoch (bootstrap distribution). Nonparametric bootstrap resampling at the trial level (resampling unit = individual trial; with replacement; 1,000 iterations; resampling size n = 2,329 trials per epoch, matched to the smallest epoch) yielded 95% percentile confidence intervals (CIs) and bootstrap bias (bootstrap mean–point estimate). For each iteration, trial-level bootstrap samples were independently drawn from each epoch and a keyword co-occurrence network was reconstructed from the resampled trials under the same construction parameters as the primary analysis; pairwise Jaccard similarities were then computed between the resulting bootstrap networks. The 2.5^th and 97.5^th percentiles of the empirical distribution defined the 95% CI bounds. Edge-level changes between consecutive epochs were classified into four mutually exclusive categories: emerging (absent earlier, present later), disappeared (present earlier, absent later), strengthened (present in both, increased weight), and weakened (present in both, decreased weight). In parallel, emerging keywords were identified from per-keyword frequency fold-changes between the post-pandemic and pre-pandemic epochs, with the largest increases highlighted as drivers of post-pandemic thematic shift.

All analyses were performed using Python 3.14.2 (pandas, NumPy, matplotlib, NetworkX, and leidenalg/igraph, and plotly). No formal hypothesis tests were performed, as the study was descriptive.

Sensitivity analysis

To assess whether the observed bridging trajectory was influenced by cohort-defining vocabulary, a sensitivity analysis was conducted in which the 18 inflammation-related terms used for cohort definition (20 word-level tokens) were excluded from the network vocabulary. Top-22 per epoch keywords were re-selected from the remaining word pool, and network construction, centrality, and community detection were repeated under identical parameters (Supplementary Table 3). In addition, because annual trial submissions jumped substantially in 2005 after the ICMJE adopted its prospective registration requirement, the pre-pandemic epoch was restricted to 2005–2019 in a supplementary sensitivity analysis to test whether pooling pre- and post-ICMJE registration behavior had distorted the pre-pandemic reference (Supplementary Table 4).

RESULTS

General characteristics of the trials

Between January 2000 and December 2025, 15,051 inflammation-related interventional clinical trials met the eligibility criteria (Fig. 1). Annual registration volume expanded through the 2000s and 2010s, apart from a 2005 policy-driven step (ICMJE registration requirement), reaching a pre-pandemic peak that was abruptly truncated by the COVID-19 pandemic onset in 2020 (Fig. 4). Pandemic-defined stratification yielded 10,366 trials (68.9%) pre-pandemic (≤ 2019), 2,356 (15.7%) during the pandemic (2020–2022; from the WHO declaration on March 11, 2020 to the global vaccination rollout marking the end of the acute phase), and 2,329 (15.5%) post-pandemic (2023–2025). Over an equivalent three-year window, the post-pandemic period sustained registration at near pandemic-era levels (n = 2,329 vs. 2,356; annual mean 776 vs. 785), with monotonic year-over-year growth (691→767→871) and 2025 representing the highest single-year count of the 25-year series; recent-year totals likely underestimate true rates owing to registration lag.

Baseline trial characteristics rebalanced across epochs (Table 1). Phase 2 trials rose from 35.3% pre-pandemic to 43.5% during and 44.0% after the pandemic, while phase 4 trials contracted from 17.8% to 12.0% and 11.9%; phase 1 and phase 3 proportions remained largely invariant. The distribution of trial status varied with registration recency, with completed trials falling from 69.3% of the pre-pandemic cohort to 13.1% post-pandemic and recruiting trials rising reciprocally from 1.2% to 49.2%, a pattern attributable to administrative censoring rather than a change in trial conduct. Median per-trial enrollment was stable across epochs (58, 54, and 60 participants, respectively).

Table 1. Baseline characteristics of inflammation-related interventional clinical trials registered on ClinicalTrials.gov by pandemic-defined period (2000–2025).

Category		Total	Pre-pandemic (≤ 2019)	Pandemic (2020–2022)	Post-pandemic (2023–2025)
Overall
	Total trials	15,051	10,366 (68.9)	2,356 (15.7)	2,329 (15.5)
Phase
	Phase 1	3,851 (25.6)	2,607 (25.1)	610 (25.9)	634 (27.2)
	Phase 2	5,705 (37.9)	3,656 (35.3)	1,025 (43.5)	1,024 (44.0)
	Phase 3	3,094 (20.6)	2,262 (21.8)	438 (18.6)	394 (16.9)
	Phase 4	2,401 (16.0)	1,841 (17.8)	283 (12.0)	277 (11.9)
Overall status
	Completed	8,356 (55.5)	7,186 (69.3)	864 (36.7)	306 (13.1)
	Recruiting	1,622 (10.8)	128 (1.2)	348 (14.8)	1,146 (49.2)
	Active, not recruiting	694 (4.6)	238 (2.3)	260 (11.0)	196 (8.4)
	Terminated	1,503 (10.0)	1,229 (11.9)	227 (9.6)	47 (2.0)
	Other/unknown	2,876 (19.1)	1,585 (15.3)	657 (27.9)	634 (27.2)
Enrollment (per trial)
	Median (IQR)	59 (24–150)	58 (23–156)	54 (23–130)	60 (30–156)

Data are presented as number (%) unless otherwise stated.

Proportions for each subcategory are calculated relative to the total number of trials within the corresponding period. Phase 1/2 trials were reclassified as Phase 1; Phase 2/3 trials were reclassified as Phase 2. Trials with missing or unclassifiable phase information are reported separately. Enrollment is reported as median with IQR.

IQR, interquartile range.

Disease category distribution and pandemic-driven shifts

Across the observation window, registered trial activity was unevenly distributed across the eighteen pre-defined disease categories (Supplementary Table 1). Rheumatology was the largest domain overall (2,904 trials; 19.3%), followed by pulmonary (2,258; 15.0%), oncology (2,166; 14.4%), dermatology (1,696; 11.3%), and infection/sepsis (1,562; 10.4%). These five categories accounted for 70.4% of all trials, whereas seven of the eighteen each contributed fewer than 1.5%.

Stratification by epoch revealed category-specific reorganization (Supplementary Table 1, Fig. 5). Pulmonary trials rose from 13.5% of pre-pandemic registrations to 21.4% during the pandemic (58.5% relative increase) before retracting to 15.3% afterward, a transient surge consistent with the temporary inclusion of pulmonary inflammatory conditions (acute lung injury, post-acute pulmonary fibrosis, acute respiratory distress syndrome [ARDS]) into the trial pipeline during 2020–2022. Oncology followed a distinct course, expanding monotonically across all three epochs from 12.3% to 16.9% to 21.3% (a cumulative 73.2% relative increase) and overtaking rheumatology as the single largest category by 2023–2025.

Several traditionally prominent categories showed a declining trajectory. Rheumatology contracted from 21.7% pre-pandemic to 12.3% during the pandemic and recovered only partially to 15.8% thereafter (Supplementary Table 1, Fig. 5). Infection/sepsis and neuroinflammation both declined monotonically across the three epochs. Mid-tier categories, including dermatology, IBD, and hematology, fluctuated within narrow bounds, indicating that the reallocation was concentrated at the high-volume end of the disease spectrum.

Structural organization of the inflammation research landscape

To characterize the static architecture underlying these category-level shifts, a weighted disease–mechanism bipartite network was constructed across the full cohort. After frequency thresholding (node ≥ 0.2%, edge ≥ 0.1% of all trials), the network comprised 21 disease nodes and 17 mechanism/treatment nodes connected by 7,959 disease–mechanism co-occurrences. Connectance was 0.255, the H2′ specialization index 0.415, and the NODF nestedness index 51.07, indicating intermediate specialization combined with moderate nestedness. This reflects a hub-and-spoke architecture in which a small set of widely applicable mechanistic strategies—particularly immune modulation, cytokine signaling, and antibody-based therapy—is deployed across many disease contexts, while peripheral disease-mechanism couplings (e.g., osteoarthritis-corticosteroid) preserve residual specialization. Bipartite-aware degree centrality identified cancer, carcinoma, arthritis, hepatitis, and rheumatoid as the principal disease anchors, and immunotherapy, immune, methotrexate, interferon, and chemotherapy as dominant mechanism/treatment integrators.

Two one-mode projections, each resolved by modularity optimization, characterized the internal structure of each partition (Fig. 2). The disease-side projection partitioned into three communities (Q = 0.0426): an immune-related disorder cluster (n = 11), an autoimmune/chronic inflammatory cluster (n = 9), and a singleton (osteoarthritis). The mechanism/treatment-side projection partitioned into three communities (Q = 0.0627): a targeted biologic and immune-modulation cluster (n = 9), an oncology-cytokine pathway cluster (n = 7), and corticosteroid as a singleton (Fig. 2). Community assignments were fully concordant between the Leiden and Louvain on the disease side (100%) and partially concordant on the mechanism/treatment side (79.4%), with the community number stable across 10 random seeds in both partitions. Absolute modularity was low (Q < 0.1), as expected for densely interconnected bipartite projections after Newman weighting, so these partitions are best read as a coarse organizational summary rather than sharply separated communities; the assignments were nonetheless stable across multiple seeds and resolution parameters. Across both partitions, immune, immunotherapy, and cancer emerged as the central anchors.

Thematic evolution across pandemic-defined epochs

Having established this static architecture, we next resolved how the thematic structure of the field evolved over time by constructing period-specific keyword co-occurrence networks for each epoch (Table 2, Fig. 3, Supplementary Fig. 1). The number of nodes remained essentially invariant across epochs (31, 32, and 31, respectively), indicating that the breadth of dominant research topics was preserved despite the reduction in trial volume. Edge counts, by contrast, contracted from 198 in the pre-pandemic network to 126 during and 115 after the pandemic, a 41.9% reduction in co-occurrence relationships. This contraction was paralleled by a decline in every global connectivity metric (Table 2). Network density fell from 0.4258 pre-pandemic to 0.2540 during and 0.2473 after; average degree from 12.7742 to 7.8750 and 7.4194; and modularity from 0.5268 to 0.4392 and 0.3932. The average clustering coefficient followed a non-monotonic trajectory, partially recovering post-pandemic.

Table 2. Temporal evolution of keyword co-occurrence network characteristics across pandemic-defined periods in inflammation-related interventional clinical trials (2000–2025).

Metrics		Pre-pandemic (≤ 2019)	Pandemic (2020–2022)	Post-pandemic (2023–2025)
Network structure
	Number of nodes	31	32	31
	Number of edges	198	126	115
	Density	0.4258	0.2540	0.2473
	Average degree	12.7742	7.8750	7.4194
	Average clustering coefficient	0.7291	0.5812	0.6362
	Modularity (Louvain)	0.5268	0.4392	0.3932
	Number of communities	5	4	5
Inter-epoch similarity
	Jaccard vs. Previous epoch (point estimate)	-	0.5140	0.7090
	Bootstrap mean (95% CI)	-	0.5540 (0.4970–0.6090)	0.6160 (0.5540–0.6790)
	Bootstrap bias (mean-point)	-	0.0400	−0.093
	Jaccard vs. pre-pandemic (point estimate)	1.0000	0.5140	0.5270
	Bootstrap mean (95% CI)	-	-	0.5570 (0.4960–0.6150)
	Bootstrap bias (mean-point)	-	-	+0.030
Edge dynamics vs. previous epoch
	Emerging edges	-	16	15
	Disappeared edges	-	88	26
	Strengthened edges	-	10	40
	Weakened edges	-	89	31
Top 5 nodes by degree centrality		Immune (0.9670)	Immune (0.7100)	Immune (0.7670)
		Cancer (0.8000)	Cancer (0.5480)	Cancer (0.5000)
		Immunotherapy (0.7000)	Immunotherapy (0.5160)	Immunotherapy (0.4670)
		Vaccine (0.7000)	Inflammatory (0.4520)	Inflammatory (0.4670)
		Inflammatory (0.6670)	COVID (0.3870)	Carcinoma (0.4330)
Top 5 nodes by betweenness centrality		Immune (0.1710)	Immune (0.2690)	Immune (0.3330)
		Cancer (0.0780)	Cancer (0.2070)	Cancer (0.1750)
		Inflammatory (0.0670)	Inflammatory (0.1230)	Infection (0.1200)
		Vaccine (0.0560)	Immunotherapy (0.0860)	Inflammatory (0.1120)
		Immunotherapy (0.0500)	Dermatitis (0.0650)	Dermatitis (0.0670)
Top 5 emerging keywords (fold-change vs. pre-pandemic)		-	-	Radiotherapy (10.2-fold)
		-	-	Neoadjuvant (8.1-fold)
		-	-	Squamous (5.4-fold)
		-	-	Autoimmune (3.6-fold)
		-	-	Thrombocytopenia (3.2-fold)

The keyword co-occurrence network was constructed from brief_title, official_title, and conditions fields of inflammation-related interventional trials within each pandemic-defined epoch, using the union of the top 22 keywords by frequency per epoch after stopword removal (implicitly excluding low-frequency terms). Nodes represent keywords; edges represent co-occurrence within the same trial. Edges were retained if co-occurrence count was ≥ 3 (absolute threshold, corresponding to 0.03–0.13% of trials across epochs). Network density is defined as the ratio of observed edges to the maximum possible edges. The average clustering coefficient reflects the mean probability that two neighbors of a node are also mutually connected. Modularity was computed via the Louvain algorithm (seed = 7); higher values indicate more clearly delineated community structure. Inter-epoch network similarity was quantified by the Jaccard coefficient (ratio of shared edges to the union of edge sets; 1.000 = identical, 0 = disjoint). Two Jaccard quantities are reported: the observed similarity of the full-cohort networks (point estimate) and the expected similarity at a common sample size of n = 2,329 trials per epoch (bootstrap distribution). Bootstrap 95% percentile CIs and bias (bootstrap mean–point estimate) were derived from 1,000 nonparametric resampling iterations at the trial level with replacement. Degree and betweenness centrality values are normalized to a 0–1 range. Fold-change values represent post-pandemic frequency divided by pre-pandemic frequency.

CI, confidence interval; Jaccard, Jaccard similarity coefficient.

Inter-epoch similarity quantified these transitions (Table 2), reported as two quantities: observed full-cohort similarity (point estimate) and expected similarity at a common sample size of n = 2,329 trials per epoch (bootstrap distribution). The pre- and during-pandemic networks shared 51.4% of their edges (point estimate = 0.5140; bootstrap mean = 0.5540; 95% CI, 0.4970–0.6090; bias +0.040), whereas the during- and post-pandemic networks shared 70.9% (point estimate = 0.7090; bootstrap mean = 0.6160; 95% CI, 0.5540–0.6790; bias −0.093). The pre- vs. post-pandemic comparison yielded a point estimate of 0.5270 (bootstrap mean = 0.5570; 95% CI, 0.4960–0.6150; bias +0.030), indicating that the post-pandemic network did not converge back toward its pre-pandemic configuration. The negative bias for pandemic-vs-post reflects differential subsampling of the larger pre-pandemic network relative to the pandemic and post-pandemic networks whose sizes are close to the common resampling size; the two quantities describe distinct estimands. The higher point estimate for the during-post comparison suggests greater structural continuity between the latter two epochs than across the pre-pandemic transition; however, the substantially overlapping bootstrap 95% CIs indicate that at matched sample size this difference is more tentative than the point estimates alone suggest.

Edge-level dynamics dissected the mechanism of these transitions (Table 2). The pre- to during-pandemic transition was dominated by edge loss: 88 edges disappeared and 16 emerged, with 89 weakened against 10 strengthened. The during- to post-pandemic transition instead showed selective consolidation, with 26 disappeared, 15 emerged, and 40 strengthened against 31 weakened, the only transition in which strengthened edges outnumbered weakened ones.

Node-level centrality identified a small set of integrative keywords anchoring the network across all epochs. Immune and cancer held the top two degree-centrality ranks in every period, with immune the most connected node throughout, recapitulating the structural anchors identified in the bipartite analysis. The betweenness of immune increased monotonically across the three epochs, nearly doubling between the pre- and post-pandemic periods, consistent with the concurrent decline in modularity. Vaccine, among the top five degree-central terms pre-pandemic, was displaced by carcinoma in the later networks.

Fold-change analysis identified a coherent set of emerging post-pandemic terms that were absent or rare pre-pandemic: radiotherapy (10.2-fold), neoadjuvant (8.1-fold), squamous (5.4-fold), autoimmune (3.6-fold), and thrombocytopenia (3.2-fold). Four of the five are oncology-adjacent. The emergence of autoimmune as a top-five term, despite the contracted overall share of rheumatology, suggests a within-domain shift from disease-specific descriptors toward the more generic autoimmune designation. While the disease category analysis identified oncology as the most expanded domain, the network-level fold-change analysis additionally revealed the specific lexical drivers (radiotherapy, neoadjuvant, squamous) underlying this expansion—a within-domain refinement not visible at the category level.

Incorporation of AI/ML methodologies

Keyword-based text mining across five trial text fields identified 21 AI/ML-incorporated trials, corresponding to 0.14% of the 15,051-trial cohort (Supplementary Fig. 2). Two methodological considerations bear on this estimate. First, the term algorithm, although initially included, was excluded after manual review of all 58 matching trials showed uniformly non-AI/ML usages (treatment algorithms, dose-escalation schemas, randomization procedures); its inclusion would have inflated the estimate three-to four-fold without adding valid AI/ML trials, a deliberate trade-off favoring specificity. Second, the analysis captured only AI/ML methodologies explicitly disclosed in the registry record and therefore likely underestimates the true incorporation rate. Within these constraints, fewer than 1 in 700 registered inflammation trials formally incorporated AI/ML methodologies into protocol design as of December 2025.

DISCUSSION

This 25-year analysis of 15,051 inflammation-related interventional trials registered on ClinicalTrials.gov used the COVID-19 pandemic to characterize shifts in trial volume, disease category composition, and thematic structure. Three main findings emerged. First, registration sustained near pandemic-era annual rates through the post-pandemic window, with continued growth and a sustained shift toward earlier-stage development. Second, the pandemic coincided with a category-level reorganization—a transient pulmonary surge layered on a sustained, longer-term rise in oncology-focused research—a pattern independently confirmed by temporal keyword and bipartite network analyses. Third, formal AI/ML incorporation in registered protocols remained negligible (0.14%), indicating that this interface has not yet translated into prospective trial design.

The sharp 2020 drop in our cohort mirrors registry-wide disruption seen in independent analyses, with time-series modeling estimating a peak shortfall of roughly 5,000 non-COVID-19 trials and cross-country data linking more severe outbreaks to lower trial screening rates [8,10,15,23]. While overall registry activity reportedly recovered to near pre-pandemic growth by 2022, the inflammation cohort sustained near pandemic-era annual rates into the post-pandemic period (mean 776/year in 2023–2025 vs. 785/year in 2020–2022), with monotonic growth culminating in the 25-year peak in 2025—a pattern consistent with reorientation rather than reversion to baseline, identifying inflammation as a therapeutic area in which the pandemic-era restructuring has stabilized at elevated levels rather than receded [24]. The phase mix shifted concurrently: phase 2 share rose by 8.7 percentage points and phase 4 fell by 5.9, while the median enrollment stayed flat (58, 54, and 60 participants). The volume drop therefore reflects fewer trials rather than smaller ones, with post-approval programs shrinking the most.

Disease category reorganization was the central finding. The pulmonary trajectory, a brief pandemic-era surge to 21.4% that partly fell back to 15.3%, is best read as pulmonary inflammatory conditions being pulled into the trial pipeline during the pandemic, including acute lung injury, ARDS-associated cytokine dysregulation, and post-acute pulmonary fibrosis; this matches the disease-specific surges reported in other registry analyses [10]. Oncology followed a different shape, rising steadily across all three epochs to a 73.2% relative increase over the full period and ultimately overtaking rheumatology as the largest category. Because it carries no pandemic-specific timing, it is better read as the continuation and acceleration of immuno-oncology growth over the past decade [25,26]. The contrast is informative: had the pandemic driven both, both would exhibit the same time-limited bump. Instead, the pandemic acted as an inflection point that coincided with, and amplified, a longer-running portfolio shift, with oncology growing faster than an expanding whole so that several established domains lost share, infection/sepsis and neuroinflammation also declining in absolute terms.

Two complementary network analyses reinforce this reading independently. The temporal analysis works from trial titles: four of the five top emerging post-pandemic keywords are oncology-adjacent, and vaccine dropped out of the top five most-connected terms, replaced by carcinoma. The network as a whole grew simpler—density and modularity both fell steadily and edge counts dropped by 41.9%—a pattern described in other biomedical co-occurrence networks as a field consolidating around fewer, more influential concepts. The betweenness of immune nearly doubled across the three epochs (0.171→0.333), reflecting the same structural change as the concurrent decline in density and modularity: as edges thin, shortest paths concentrate through remaining hub nodes [4]. This bridging pattern was preserved in a sensitivity analysis excluding cohort-defining vocabulary, with cancer showing monotonic betweenness increase (0.082→0.311→0.399) and infection and vaccine displaying epoch-consistent bridging trajectories (Supplementary Table 3), indicating a broader consolidation around integrative concepts rather than a vocabulary artifact. The bipartite analysis offers a structural counterpart: a handful of widely used mechanistic strategies are applied across a growing range of inflammatory conditions, while disease-specific links preserve specialization at the edges [27,28]. On the mechanism side, immunomodulatory and oncology-related terms form a tightly linked core, evidence that the line between the two has become porous—consistent with the recognized convergence of inflammation and tumor immune [29]. The two analyses converge: the recent oncology-focused shift is an amplification of latent features present before the pandemic.

Three quantitative shifts observed in our cohort have direct implications for trial design within the clinical pharmacology framework. First, the Phase 2 expansion (35.3%→44.0%) concentrates inflammation drug development at the stage where dose-finding, exposure-response characterization, and proof-of-mechanism biomarker integration are central, indicating that the post-pandemic portfolio has shifted toward earlier, mechanism-anchored evaluation. Second, as the network simplified, immune remained the dominant bridging node across disease clusters (betweenness 0.171→0.333); in clinical-pharmacology terms, this convergence around shared pathways—JAK/STAT, IL-6, TNF, and immune checkpoint biology—prioritizes target-engagement biomarkers and exposure-response endpoints applicable across what were previously distinct therapeutic areas. Third, Phase 4 activity did not scale with the expanding early-phase portfolio, its share contracting from 17.8% pre-pandemic to 11.9% post-pandemic, raising specific concerns about long-term real-world evidence for chronic immunomodulatory regimens with sustained immunosuppressive action. These registry-observable shifts identify trial-design priorities—earlier-phase concentration, mechanism-anchored biomarker selection, and post-marketing surveillance not keeping pace with development—that situate the manuscript’s clinical-pharmacology framing in measurable changes rather than narrative inference; direct pharmacokinetics/pharmacodynamics characterization in each therapeutic area lies beyond the scope of a title-based registry cohort and warrants follow-up in dedicated pharmacological cohorts.

That only 0.14% (n = 21) of registered inflammation trials formally incorporated AI/ML methods stands in sharp contrast to the large and growing body of AI/ML-oriented translational work in this field [11,14]. The gap is most parsimoniously interpreted as a translational lag rather than lack of interest, given that a similar gap between methodological discussion and actual use in trial protocols has been described in an adjacent therapeutic area [30]. Within the clinical-pharmacology frame established above, AI/ML-enabled approaches to population pharmacokinetic modeling, exposure-response analysis, and biomarker-guided patient stratification have proliferated in the methodological literature, yet our cohort indicates that translation into registered inflammation protocols remains minimal. This quantifies the trial-level adoption barrier that clinical pharmacologists working in inflammation must address; closing the gap will require regulatory and protocol-design frameworks that make room for AI/ML components in the registered protocol itself.

Several limitations should be acknowledged. First, the cohort was drawn solely from ClinicalTrials.gov and identified by a curated 18-term title search prioritizing specificity, likely under-representing the inflammation-relevant trials without inflammation-titled labels and early-phase research from outside the United States. Second, network analyses rely on raw word frequency and a curated vocabulary; this may favor domains with mechanistic titles (oncology) over disease-specific ones (rheumatology), potentially contributing to the apparent contraction of rheumatology-specific vocabulary alongside the rise of the generic ‘autoimmune’ designation. Additionally, absolute edge-count thresholds (≥ 3) correspond to different relative proportions across epochs of unequal size (0.03% pre-pandemic vs. 0.13% pandemic and post-pandemic), potentially contributing to the observed density decline. Third, the pre-pandemic epoch spans two decades including the 2005 ICMJE-driven inflection; sensitivity analysis restricted to 2005–2019 preserved all principal findings (Supplementary Table 4), but epoch-by-epoch metrics remain cumulative rather than rate-matched. Finally, this descriptive analysis cannot establish causality; the AI/ML estimate should be interpreted as a lower bound given possible undeclared methodological components.

SUPPLEMENTARY MATERIALS

Supplementary Data 1

Inflammation-related keyword list

Supplementary Data 2

AI/ML-related keyword list

Supplementary Table 1

Distribution of inflammation-related interventional clinical trials by disease category based on MeSH terminology across pandemic-defined periods (2000–2025)

Supplementary Table 2

Predefined keyword vocabularies used in the bipartite disease-mechanism/treatment network analysis

Supplementary Table 3

Sensitivity analysis of keyword co-occurrence network characteristics with cohort-defining vocabulary excluded (2000–2025)

Supplementary Table 4

Sensitivity analysis restricting the pre-pandemic epoch to 2005–2019 to address the 2005 ICMJE registration discontinuity

Supplementary Figure 1

Temporally stratified keyword co-occurrence networks across pandemic-defined epochs. Small-multiples representation of inflammation-related keyword co-occurrence networks across three epochs: (A) pre-pandemic (≤ 2019), (B) pandemic (2020–2022), and (C) post-pandemic (2023–2025). The common vocabulary was defined as the union of the top 22 most frequent keywords in each epoch; only keywords in ≥ 5 trials and keyword pairs co-occurring in ≥ 3 trials (raw counts, no normalization) within a given epoch were retained for that panel. To enable cross-panel comparison, node coordinates were computed once from the union network using the Fruchterman–Reingold layout (seed = 7) and applied identically to all panels. Node size is proportional to raw keyword frequency in that epoch; node fill color denotes category (red, disease/condition; blue, biological/mechanism; teal, treatment/drug; gray, other); edge color denotes epoch (gray, pre-pandemic; red, pandemic; green, post-pandemic), with edge thickness proportional to within-epoch co-occurrence count. Nodes and edges below threshold in a given epoch are shown as light-gray ghost elements at the same shared coordinates.

Supplementary Figure 2

Frequency of AI/ML-related keywords identified in inflammation-related interventional clinical trials registered on ClinicalTrials.gov (2000–2025). Keywords were searched across trial titles, brief summaries, and detailed descriptions. The term ‘algorithm’ was excluded from the final keyword set following manual review of all matching trials (n = 58), which confirmed that occurrences predominantly reflected non-AI/ML clinical contexts (e.g., treatment algorithms, dose-escalation schemas, randomization procedures) with no co-occurrence of AI/ML-specific terminology.