Insights into the adoption of innovative clinical trials across therapeutic areas using clinical trials registry data and large Language models

Abstract

Innovative clinical trial designs, such as adaptive and Bayesian methodologies, have gained traction as solutions to the challenges of traditional trials, including their high costs and complex regulations. When they adhere to relevant ethical and regulatory requirements, these designs can improve efficiency, flexibility, and ethical standards. However, their application outside of oncology, particularly in fields such as neuroscience and rare diseases, remains underexplored. We analyzed data from ClinicalTrials.gov for interventional trials registered between 2005 and 2024. The trials were classified as innovative or traditional using a keyword-based algorithm. Therapeutic areas were identified using a large language model (LLM), with classification accuracy evaluated using a random sample of 2,000 trials. Of the 348,818 trials, 5827 were classified as innovative, with prevalence in neuroscience and rare diseases. These designs were predominantly observed in early-phase trials and pediatric research, with limited representation in elderly-focused or sex-specific studies. Innovative trial adoption has grown since 2011, spurred by regulatory advancements and increased funding from scientific networks and the National Institutes of Health. Survival analysis revealed that innovative trials tend to remain active for longer than traditional trials; however, this trend varies across different medical disciplines. LLM demonstrated a classification accuracy of 94.6% (95%CI = 93.6%-95.5%), supporting its utility for trial categorization. The rise in innovative clinical trial designs reflects a shift toward addressing complex challenges in neuroscience, rare diseases, and other therapeutic areas. Although these designs show promise in improving trial efficiency and patient outcomes, their success depends on rigorous planning and adherence to regulatory standards. Advancing LLM-based tools can further optimize clinical trial monitoring by tailoring research in trial settings and therapeutic fields.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-18488-8.

Introduction

Clinical trials play an important role in evaluating the safety and efficacy of medical treatment. Historically, clinical experimental research has followed a traditional design; however, recent challenges, such as escalating costs in drug development and increasingly complex regulatory requirements, have prompted a shift toward more innovative approaches¹. In this context, innovative clinical trials, which have gained significant traction over the past two decades, can offer a promising solution². These trials are characterized by using novel methodologies to enhance efficiency, cost-effectiveness, and ethical standards in clinical research³.

Novel clinical trials fall into two types: Adaptive Designs and Bayesian Designs. Adaptive Designs are dynamic, permitting modifications to trial parameters based on interim analysis. This promotes real-time learning and decision-making, making trials more efficient and ethical by minimizing unnecessary patient exposure to inferior treatments⁴. Adaptive designs can improve ethical outcomes by enabling real-time adjustments, such as stopping a trial early for efficacy or futility, thereby minimizing patient exposure to ineffective or harmful treatments^3,5. Similarly, these methodologies can enable more efficient resource allocation and reduced recruitment of patients into less promising arms, ensuring that trial participants are more likely to benefit from the experimental interventions⁶.

Moreover, adaptive trial designs can optimize resource allocation by allowing mid-trial modifications based on the interim results. For instance, adaptive randomization can direct more patients toward promising treatment arms while reducing enrollment in less effective ones, maximizing available resources⁵. Additionally, early stopping rules for futility or efficacy help conserve resources by avoiding the unnecessary continuation of unproductive trials. These features ensure that trial funds, time, and participant efforts are concentrated on interventions with the greatest potential for success³.

Conversely, Bayesian Designs incorporate Bayesian statistical methods to integrate prior knowledge of treatment effects by accumulating trial data. This method provides a more holistic view of the data and can be particularly useful in cases where historical data or expert opinions can guide the direction of the trial⁷. For example, Bayesian approaches allow the incorporation of prior data from similar studies, reducing the need for large, homogenous sample sizes, and enabling participation from smaller, underrepresented populations⁸.

Despite innovative trial designs’ advantages, comprehensive evaluations of their implementation and impact across therapeutic domains in medicine are lacking. This shift has far-reaching implications for clinical outcomes, healthcare research, and policy. However, a significant gap remains in assessing these trials’ implementation and impact, particularly outside the well-explored fields of oncology and hematology⁹.

Evaluating the current state of the art and implementing novel trials across various therapeutic areas can have a significant impact on healthcare outcomes. This evaluation should help identify successful strategies and innovations in trial design, which can expedite the development of new and effective treatments and ultimately improve patient outcomes across multiple diseases¹⁰. Additionally, it should aid in better resource allocation by guiding policymakers and funding bodies to support high-impact research areas and ensure the optimal use of available resources¹¹. Ultimately, this assessment aims to promote the adoption of best practices and drive continuous innovation in clinical trial methodologies, thereby enhancing the global quality and efficiency of clinical research¹².

Within this general framework, the ClinicalTrials.gov database serves as a crucial resource for understanding the current state of clinical research worldwide. It provides comprehensive data from clinical trials across various therapeutic areas. Despite its vast information repository, it is often underutilized. This database offers valuable insights into trial methodologies, outcomes, and trends, serving as a crucial resource for researchers, clinicians, and policymakers seeking to advance medical science and patient care¹³.

Moreover, large language models (LLMs) such as ChatGPT can facilitate free-text disease classification¹⁹ in the ClinicalTrials.gov database to monitor state-of-the-art clinical trials, thereby improving efficiency and accuracy. These models can quickly process vast amounts of text data and interpret complex medical terminology to classify diseases without manual effort¹⁹. ChatGPT’s natural language understanding effectively interprets free-text fields in clinical trial databases, ensuring accurate trial categorization¹⁴. Additionally, LLMs are scalable and capable of handling large datasets, making them ideal for extensive databases like ClinicalTrials.gov. They can also be continuously trained and updated, improving their performance and keeping classifications current with evolving medical knowledge¹⁵.

This study has two primary objectives. First, it seeks to quantify the adoption of innovative clinical trial designs, specifically adaptive and Bayesian approaches, and to evaluate how their use has evolved over time using data from ClinicalTrials.gov registry⁹. In addition, we aim to characterize how the implementation of these designs varies according to key trial features, including study population, recruitment status, duration, funding source, and geographic distribution.

Second, we aim to assess the performance of a large language model (LLM)-assisted classification system for assigning clinical trials to therapeutic areas, based on free-text disease condition fields available in ClinicalTrials.gov. This objective reflects the growing need for scalable and automated approaches to manage trial registry data’s increasing volume and complexity. By evaluating the accuracy, sensitivity, precision, and overall reliability of the LLM in reproducing expert-based classification, we seek to determine whether such tools can feasibly support large-scale, real-time monitoring of innovation in clinical trial design across therapeutic disciplines¹⁶.

The broader goal of this study is to better understand how the uptake of innovative trials varies across medical disciplines and over time and to explore the utility of artificial intelligence tools in improving the efficiency and granularity of trial design monitoring at scale.

Methods

Study design and data source

This was a retrospective study based on data from ClinicalTrials.gov, the largest publicly available clinical trial registry. On June 27, 2024, a comma-separated value (CSV) export was executed for all interventional trials registered on ClinicalTrials.gov and uploaded into R.

Study sample

Studies published after 2005 were included. This date was considered because, in 2004, the International Committee of Medical Journal Editors (ICMJE), a group of editors from leading medical journals, issued the first of several statements requiring the prospective registration of clinical trials (i.e., before enrollment of the first patient) by September 2005 as a condition for publication of the study results¹⁷.

Study outcomes

This study focused on two primary research outcomes, each aligned with a distinct objective. The first outcome was the adoption of innovative clinical trial design specifically, the extent to which adaptive and Bayesian methodologies have been implemented across therapeutic areas and over time. This outcome reflects the broader shift in methodological practices in clinical research and captures patterns in how innovation is integrated into trial design.

The second outcome pertained to a large language model (LLM) classification performance in assigning trials to their correct therapeutic areas. This outcome was designed to evaluate whether artificial intelligence tools can support the efficient and scalable categorization of large volumes of trial registry data, thereby enabling automated monitoring of trial design trends by medical specialty.

Study endpoints and measurements

To achieve the study objectives, we defined a set of measurable endpoints extracted from structured and unstructured fields within the ClinicalTrials.gov dataset.

Objective 1: quantifying the adoption of innovative designs

The primary endpoint, necessary to quantify the adoption of innovative trial designs (Objective 1), was each trial’s design identification as innovative or not. Trials were categorized as innovative if their “study description” field contained at least one of 42 predefined keywords referring to adaptive or Bayesian design elements.

The analysis was based on a keyword search in the study description fields to identify innovative designs in the ClinicalTrials.gov database.

These designs incorporate adaptive methodologies, which enable modifications to the trial structure based on accumulating data, and Bayesian methodologies, which facilitate the integration of prior knowledge and real-time updates^5,18.

A search string of 42 keywords identified in the study description field of ClinicalTrials.gov was considered. The keywords were derived from another study searching for advanced trials in ClinicalTrials.gov¹⁹. Therefore, we grouped all trials into Innovative, if identified by our string, and Traditional, if not.

The keywords considered for the search were as follows:

(“treatment switching”, “adaptive”, “Bayes”, “phase II/III”, “phase I/II”, “biomarker adaptive”, “biomarker adaptive design”, “biomarker adjusted”, “adaptive hypothesis”, “adaptive dose-finding”, “pick-the-winner”, “drop-the-loser”, “sample size re-estimation”, “re-estimations”, “adaptive randomization”, “group sequential”, “adaptive seamless”, “adaptive design”, “interim monitoring”, “Bayesian adaptive”, “flexible design”, “adaptive trial”, “play-the-winner”, “adaptive method”, “adaptive dose adjusting”, “response adaptive”, “adaptive allocation”, “adaptive signature design”, “treatment adaptive”, “covariate adaptive”, “sample size adjustment”, “switch from superiority to noninferiority”, “adjustable design”, “multiple adaptive design”, “trial modification”, “flexible trial”, “modified trial”, “adaptive sequential sampling”, “response-adaptive”, “adaptive sampling”, “adaptive urn design”, “multi-stage design”).

• The terms included in the search string, such as “treatment switching,” “adaptive,” “Bayes,” “biomarker adaptive design,” and “adaptive randomization,” reflect some features of innovative designs. Adaptive trials allow changes to the study protocol in response to interim results. Bayesian adaptive trials instead use prior knowledge to update probabilities as data accumulate^5,7,20. The broader terms “Bayes” and “Bayesian” are referred to Bayesian design as reported in other review performed on Clinicaltrials.gov²¹.

• Other terms, such as “phase ii/iii” and “phase i/ii,” highlight seamless trial designs that transition between phases without interrupting patient recruitment, thereby optimizing trial resources. “Biomarker adaptive,” “biomarker adjusted,” and “adaptive hypothesis” indicate trials that tailor interventions based on biological markers, aligning with the principles of precision medicine²².

• Terms such as “sample size re-estimation,” “adaptive randomization,” and “group sequential” designs ensure that trials remain adequately powered while dynamically modifying sample size and treatment allocation as new data become available¹⁸. Strategies such as “play-the-winner” and “drop-the-loser” dynamically adjust treatment assignments, prioritizing promising interventions while minimizing patient exposure to ineffective treatments²³.

• More flexible trial approaches, as captured by terms like “trial modification,” “multiple adaptive design,” and “switch from superiority to non-inferiority,” enable greater flexibility in regulatory pathways, ensuring trials remain aligned with evolving clinical knowledge⁸. Finally, “adaptive urn design,” “multi-stage design,” and “response-adaptive sampling” refine patient recruitment and treatment adaptation, ensuring that trial designs optimize resource use and patient benefit²⁴.

Secondary Endpoints.

To explore how adoption varied across contexts and populations, we defined several secondary endpoints, describing the characteristics of both traditional and innovative trials:

Other secondary endpoints included:

• Trial status identification, as reported in ClinicalTrials.gov (completed, active, terminated, suspended, withdrawn), which allowed for survival analysis of early discontinuation;

• Trial duration, calculated as the number of days between the start date and the completion date or censoring date (for ongoing or suspended studies);

• Population focus, determined from structured fields indicating whether the study targeted pediatric or elderly populations, or was sex-specific (female-only or male-only);

• Trial phase, as reported (Early Phase I to Phase IV);

• Funding source, categorized from the sponsor and collaborator fields into NIH, industry, federal government, individual, network, or other;

• Geographic location, based on trial site locations grouped by continent; and.

• Availability of results, a binary variable indicating whether summary results were posted in the registry.

Objective 2: assessing LLM classification accuracy

To address Objective 2, which aimed to assess the performance of a large language model (LLM) in classifying clinical trials into therapeutic areas, we defined an additional endpoint: the LLM’s classification performance.

The therapeutic area was identified by applying a large language model (ChatGPT 4.0) to each trial’s free text “condition” field. Using the ChatGPT algorithm, we categorized trial diseases in therapeutic areas according to the main free-text disease field, analyzing trials initiated after 2025.

In this study, we adopted a predetermined list of therapeutic areas, as reported in Table 1, to categorize clinical trials based on their primary disease condition. The list was created based on other published research identifying therapeutic areas in ClinicalTrials.gov²⁵. This list structures medicine into broad, organ- or system-based specialties (e.g., Cardiology, Neurology, Oncology)¹⁰. The authors have further integrated the list with cross-cutting fields (e.g., Public Health, Geriatrics, Nutrition) and specialized therapeutic areas for special populations (e.g., Pediatrics, Women’s Health). Research is organized to accommodate emerging or interdisciplinary fields (e.g., Rare Diseases and Genetics, Health Informatics) that are increasingly prominent in global research priorities.

Table 1.

Descriptive table of innovative versus traditional trials within therapeutic areas. Percentages in parentheses represent the proportion of trials within each therapeutic area, relative to the total number of trials classified as innovative or traditional.

	Traditional N = 342,991	Innovative N = 5827
Hematology and Oncology	58,739 (17%)	1922 (33%)
Psychiatry, Substance Abuse, and Mental Health	20,986 (6.1%)	686 (12%)
Neurology and Neurosurgery	26,418 (7.7%)	575 (9.9%)
Infectious Diseases	20,552 (6.0%)	407 (7.0%)
Rare Diseases and Genetics	8547 (2.5%)	290 (5.0%)
Gastroenterology and Hepatology	19,469 (5.7%)	244 (4.2%)
General Preventive Medicine, Public Health, and Geriatrics	25,475 (7.4%)	226 (3.9%)
Endocrinology and Metabolism	21,857 (6.4%)	183 (3.1%)
Dental and Otolaryngology	14,691 (4.3%)	175 (3.0%)
Cardiology and Cardiothoracic Surgery	22,748 (6.6%)	169 (2.9%)
Dermatology	9611 (2.8%)	152 (2.6%)
Musculoskeletal and Orthopedics	14,762 (4.3%)	131 (2.2%)
Autoimmune Immunology	6429 (1.9%)	96 (1.6%)
Pulmonology and Respirology	10,927 (3.2%)	93 (1.6%)
Pain Management and Rehabilitation	10,771 (3.1%)	80 (1.4%)
Sexual Health Sleep Disorders and Sports Medicine	6005 (1.8%)	75 (1.3%)
Urology	7411 (2.2%)	71 (1.2%)
Ophthalmology	6927 (2.0%)	55 (0.9%)
Pediatrics and Neonatology	3269 (1.0%)	50 (0.9%)
Radiology and Diagnostic Imaging	6402 (1.9%)	55 (0.9%)
Anesthesiology Critical Care and Transplantation	4765 (1.4%)	31 (0.5%)
Women’s Health and Reproductive Medicine	7734 (2.3%)	25 (0.4%)
Other	638 (0.2%)	13 (0.2%)
Vascular Medicine and Surgery	1734 (0.5%)	6 (0.1%)
General and Plastic Surgery	2947 (0.9%)	6 (0.1%)
Allergy and Immunology	626 (0.2%)	4 (< 0.1%)
Nutrition	747 (0.2%)	2 (< 0.1%)
Health Informatics Pharmacy and Others	802 (0.2%)	3 (< 0.1%)
Nephrology	1002 (0.3%)	2 (< 0.1%)

ChatGPT was then used to map each trial’s free-text “condition” field to one of these author-defined therapeutic areas. Due to the lack of precise information about the medical branch in our database, we used the algorithm ChatGPT 4o to classify innovative trial designs in different therapeutic areas. We interfaced with ChatGPT, asking it to find an exact clinical specialization based on the free text reported for each trial in the “condition” field of the trial disease. The results have been added to the database in a new column named “Therapeutic Area.”

Statistical analysis

Quantifying the adoption of innovative designs

Innovative Trial Adoption Trend. The absolute and relative frequencies, according to the overall trial characteristics (Innovative vs. Traditional), are reported for the categorical variable, and the median with interquartile ranges are provided for the quantitative variables. The odds Ratio (OR) and 95% confidence intervals (CI) for a Univariable Logistic Regression model comparing the innovative trial with the others are also provided.

The trial publication year on ClinicalTrials.gov was categorized according to the regulatory agencies’ publication timelines for guidelines concerning novel, Bayesian, and adaptive designs, as reported in the supplementary material (Table S1).

Innovative trial adoption and outcome in Therapeutic fields. Plots depicting the proportion of innovative trials in the therapeutic field are also included. 95% confidence intervals were calculated using the mid-P approach to mitigate the conservatism of the exact Clopper-Pearson interval²⁶.

Data was prepared by deriving intervals between trial dates (first posted date, start date, completion date, and last update date) to define the time-to-event outcomes. Trials that were terminated before completion were coded as events, and all other statuses were coded as censored observations. Survival times were calculated from the start date to the event or censoring date (e.g., completion date and last update for suspended studies). Studies withdrawn were excluded from the analysis. Kaplan–Meier curves and log-rank tests were used to compare termination-free survival across groups, and a Cox proportional hazards model, including interaction terms between trial design (Traditional or Innovative) and therapeutic area, was fitted. The final estimates quantify the hazard of termination as a function of trial design and therapeutic area. Predicted survival curves were derived from the Cox model to visualize the differences over time. Effects are adjusted for funder type, trial initiation year, sex, pediatric and elderly populations, trial phase, enrollment size, and geographic continent.

Adoption of Innovative Trials in Predominant Therapeutic Areas. We conducted a descriptive analysis and developed a multivariable logistic regression model, focusing on the therapeutic area with the highest adoption of innovative trial designs outside of oncology, to assess how trial characteristics influence the likelihood of adopting such designs.

Assessing LLM classification accuracy

A manual review of a random sample of 2000 trials was performed. Using standard sample size estimation for proportions, a sample of 1865 trials would enable the analysis of a 95% accuracy rate, similar to other LLM classification tasks reported in the literature¹⁹, with a 95% confidence interval (CI) of ± 2%. To ensure a sufficient margin and accommodate stratification across therapeutic areas, we rounded up to 2000 trials. The validation sample was selected before applying the large language model (LLM) classification. We drew a random sample from the entire dataset of registered trials. Two independent reviewers (VS and RC) reviewed the “condition” field and assigned each trial to one of the predefined therapeutic areas. In case of disagreement, a third reviewer adjudicated the final assignment. The manually assigned therapeutic area was then compared to ChatGPT’s classification. A classification was considered accurate if ChatGPT’s category matched the manually assigned label.

The accuracy rate with F1, precision, and recall was computed for the GPT classification of disease conditions with 95% confidence intervals. Analyses and data extraction were conducted using R version 3.4.2²⁷.

Results

Data

In the overall dataset of 499,740 studies, 383,235 were interventional trials, and 351,713 were registered from January 1, 2005, to June 2024. Among the 348,818 records reporting the study description field, 5,827 were identified as innovative trials (Fig. 1).

Fig. 1.

ClinicalTrials.gov studies identification flowchart.

Quantifying the adoption of innovative designs

Innovative Trial Adoption Trend.

The distribution of trials over time shows an increasing trend in innovative trials, particularly from 2011 onwards, with these trials being more frequently active, suspended, terminated, or withdrawn than traditional trials (Table S2).

Innovative trials were less likely to include sex-specific populations (i.e., only male or only female participants) and more likely to include participants of all sexes. Although pediatric trials were more common among the innovative trials, there was no significant difference in the proportion of elderly trials. Early-phase (I–II) trials dominated both categories, but were more prevalent in innovative trials, whereas late-phase (III–IV) trials were less frequent. Enrollment sizes did not differ significantly (Table S2).

In terms of funding, innovative trials received more support from networks and the NIH, whereas there were no significant differences in industry or individual funding. Geographically, innovative trials were more prevalent in the Americas, Europe, and Oceania and less so in Africa and Asia. Innovative trials are also likely to have their results publicly posted. Furthermore, they tended to have significantly longer durations, with a higher proportion lasting more than five or even ten years compared to traditional trials (Table S2).

Innovative trial adoption and outcome in Therapeutic fields.

Table 1 shows a significantly greater number of innovative trials in some therapeutic regions than traditional trials. Hematology and Oncology had the highest proportion of creative studies, reaching 33% versus 17% in conventional settings. Similarly, the proportion of innovative trials in Psychiatry, Substance Abuse, and Mental Health rose from 12 to 6.1%, and in Neurology and Neurosurgery, It increased from 9.9 to 7.7%. Other fields, including Infectious Diseases, Rare Diseases and Genetics, and Gastroenterology and Hepatology, also exhibited higher percentages of innovative trials.

Figure 2 presents Kaplan–Meier survival curves comparing the time until trial termination between innovative and traditional trials. The log-rank test revealed a statistically significant difference in the termination-free survival between the two groups.

Fig. 2.

Trial Termination Free survival Kaplan Meier estimation with Log Rank test between Innovative and Traditional trials.

Panel A of Table 2 illustrates the model-based predicted termination-free survival curves across different therapeutic areas for both the innovative and traditional trials. Although there are visible differences in the shape and duration of these survival curves, the impact of Innovative design on termination-free survival is not uniform across all therapeutic fields. The predicted curves suggest divergence between trial types for some areas, such as neurology and neuroscience.

Table 2.

Predicted trial termination-free survival and Cox model Estimation of early termination risk. Panel A: predicted survival curves showing the probability of remaining trial termination-free across therapeutic areas, comparing traditional (red) and innovative (blue) study designs. Shaded regions represent 95% confidence intervals. Panel B: results from a Cox proportional hazards model estimating the hazard of trial termination. Hazard ratios (HR), 95% confidence intervals (CI), and p-values are reported for innovative design and therapeutic area effects, including interactions. Effects are adjusted for funder type, trial initiation year, sex, pediatric and elderly populations, trial phase, enrollment size, and geographic continent.

Panel B provides the results of the Cox proportional hazard model. The overall effect of adopting an innovative design, compared to a traditional design, adjusted for the therapeutic area, did not reach statistical significance. In contrast, several therapeutic areas demonstrated significantly lower global hazards of early termination than the reference group, particularly Hematology and Oncology, Rare Disease, Neurology, and Neurosurgery. The interaction terms indicate that the benefits of an innovative approach can vary depending on the therapeutic area. For example, in Neurology and Neurosurgery, innovative designs have been associated with a reduced risk of early termination.

Innovative Trials in Neuroscience.

Neuroscience trials exhibit broader trends observed in the overall dataset, including the increased adoption of innovative trials over time, the emergence of more innovative designs in early-phase trials, and a preference for NIH funding; however, the intensity and pattern of these effects differ. Neuroscience trials demonstrate a stronger time-based acceleration toward novel designs, which are highly adopted for pediatric populations, and a pronounced avoidance of elderly-focused trials. Regarding trial phases, both the overall and neuroscience subsets demonstrated a preference for novel designs in early-phase trials. However, the contrast between the early and late phases is somewhat less extreme in neuroscience, while innovative trials tend to have a longer overall duration. Moreover, within the trial phases, a higher representation of Phase II/III trials and Phase II trials compared to Phase I is shown (Table 3).

Table 3.

Descriptive table of innovative versus traditional trials in neurosciences (Neurology, Neurosurgery, Psychiatry, substance Abuse, and mental health trials). The univariable logistic regression model odds ratio (OR) with 95% confidence interval and P-values was also reported. Percentages in parentheses represent the proportion within each therapeutic area relative to the total number of trials classified as innovative or traditional.

	Overall	Traditional	Innovative	OR	P-value
	N = 348,818	N = 342,991	N = 5827
Characteristic	N = 48,665	N = 47,404	N = 1261
Year:
(2005,2007]	1098 (5.07%)	1082 (5.16%)	16 (2.33%)	Ref.	Ref.
(2007,2010]	2098 (9.68%)	2064 (9.84%)	34 (4.96%)	1.11 [0.62;2.08]	0.737
(2010,2019]	9831 (45.4%)	9531 (45.4%)	300 (43.7%)	2.11 [1.31;3.65]	0.001
(2019,2023]	7182 (33.1%)	6911 (32.9%)	271 (39.5%)	2.63 [1.63;4.55]	< 0.001
(2023, 2024]	1463 (6.75%)	1398 (6.66%)	65 (9.48%)	3.12 [1.84;5.62]	< 0.001
Status1:
Completed	25,629 (59.4%)	25,073 (59.7%)	556 (48.7%)	Ref.	Ref.
Active	13,044 (30.3%)	12,549 (29.9%)	495 (43.4%)	1.78 [1.57;2.01]	< 0.001
Suspended	192 (0.45%)	187 (0.45%)	5 (0.44%)	1.24 [0.44;2.73]	0.645
Terminated	2825 (6.55%)	2772 (6.60%)	53 (4.65%)	0.86 [0.64;1.14]	0.307
Withdrawn	1422 (3.30%)	1390 (3.31%)	32 (2.80%)	1.04 [0.71;1.47]	0.82
Sex of Trial Participant:
All	45,368 (93.3%)	44,170 (93.2%)	1198 (95.1%)	Ref.	Ref.
Female	2137 (4.39%)	2094 (4.42%)	43 (3.41%)	0.76 [0.55;1.02]	0.068
Male	1142 (2.35%)	1123 (2.37%)	19 (1.51%)	0.63 [0.38;0.96]	0.032
Pediatric Trial:
No	39,089 (80.3%)	38,367 (80.9%)	722 (57.3%)	Ref.	Ref.
Yes	9576 (19.7%)	9037 (19.1%)	539 (42.7%)	3.17 [2.83;3.55]	< 0.001
Elderly Trial:
No	14,284 (29.4%)	13,710 (28.9%)	574 (45.5%)	Ref.	Ref.
Yes	34,381 (70.6%)	33,694 (71.1%)	687 (54.5%)	0.49 [0.44;0.55]	< 0.001
Trial Phases aggregated:
Early I-II	10,734 (59.9%)	10,421 (59.5%)	313 (75.2%)	Ref.	Ref.
Late III-IV	7196 (40.1%)	7093 (40.5%)	103 (24.8%)	0.48 [0.38;0.60]	< 0.001
Trial Phases:
Early Phase I	2018 (29.8%)	1979 (29.9%)	39 (25.7%)	Ref.	Ref.
Phase II	1810 (26.7%)	1745 (26.3%)	65 (42.8%)	1.89 [1.27;2.85]	0.002
Phase II|Phase III	295 (4.35%)	283 (4.27%)	12 (7.89%)	2.17 [1.07;4.08]	0.033
Phase III	1173 (17.3%)	1154 (17.4%)	19 (12.5%)	0.84 [0.47;1.44]	0.532
Phase IV	1481 (21.9%)	1464 (22.1%)	17 (11.2%)	0.59 [0.32;1.04]	0.067
Enrollment sample size	60.0 [25.0;132]	60.0 [25.0;132]	60.0 [30.0;150]	1.00 [1.00;1.00]	0.625
Funder Type:
Federal or Governative	2002 (4.11%)	1956 (4.13%)	46 (3.65%)	Ref.	Ref.
Individual	76 (0.16%)	71 (0.15%)	5 (0.40%)	3.07 [1.02;7.32]	0.046
Industry	8098 (16.6%)	7942 (16.8%)	156 (12.4%)	0.83 [0.60;1.17]	0.291
Network2	173 (0.36%)	165 (0.35%)	8 (0.63%)	2.09 [0.90;4.29]	0.084
NIH3	361 (0.74%)	338 (0.71%)	23 (1.82%)	2.90 [1.70;4.81]	< 0.001
Other4	37,951 (78.0%)	36,928 (77.9%)	1023 (81.1%)	1.17 [0.88;1.61]	0.282
Continent:
Africa	767 (1.72%)	759 (1.75%)	8 (0.68%)	Ref.	Ref.
Americas	22,742 (51.0%)	22,012 (50.6%)	730 (62.5%)	3.08 [1.64;6.82]	< 0.001
Asia	8402 (18.8%)	8263 (19.0%)	139 (11.9%)	1.57 [0.82;3.51]	0.191
Europe	12,351 (27.7%)	12,071 (27.8%)	280 (24.0%)	2.16 [1.14;4.80]	0.016
Oceania	369 (0.83%)	358 (0.82%)	11 (0.94%)	2.90 [1.15;7.65]	0.024
Trial Duration (Years)	1.05 [0.19;2.32]	1.03 [0.18;2.30]	1.50 [0.43;3.00]	1.13 [1.08;1.18]	< 0.001
Study Results:
Not Posted	41,004 (84.3%)	39,953 (84.3%)	1051 (83.3%)	Ref.	Ref.
Posted	7661 (15.7%)	7451 (15.7%)	210 (16.7%)	1.07 [0.92;1.24]	0.368
Trial Duration (Categories)
Less than 5 years	14,144 (95.5%)	13,654 (95.5%)	490 (93.0%)	Ref.	Ref.
5–10 years	640 (4.32%)	604 (4.23%)	36 (6.83%)	1.67 [1.16;2.33]	0.007
>= 10 years	34 (0.23%)	33 (0.23%)	1 (0.19%)	0.96 [0.04;4.43]	0.971
Therapeutic Area:
Neurology and Neurosurgery	26,993 (55.5%)	26,418 (55.7%)	575 (45.6%)	Ref.	Ref.
Psychiatry Substance Abuse and Mental Health	21,672 (44.5%)	20,986 (44.3%)	686 (54.4%)	1.50 [1.34;1.68]	< 0.001

¹The Active studies include 1) not yet Recruiting: The study has not started recruiting participants; 2)Recruiting: The study is currently recruiting participants; 3)Enrolling by invitation: The study selects its participants from a population or group of people, decided on by the researchers in advance. These studies are not open to everyone who meets the eligibility criteria, but only to people in that particular population, who are specifically invited to participate. 4)Active, not recruiting: The study is ongoing, and participants are receiving an intervention or being examined; however, potential participants are not currently being recruited or enrolled.

Suspended: The study has stopped early, but may start again.

Terminated: The study was stopped early and will not start again. The participants were no longer being examined or treated.

Completed: The study has ended normally, and participants are no longer being examined or treated (i.e., the last participant’s last visit has occurred).

Withdrawn: The study was stopped early before enrolling the first participant.

² Community-based organizations.

³U.S. National Institutes of Health.

⁴All others including Universities.

The multivariable logistic regression model in Table 4 indicates that neuroscience trials initiated between 2011 and 2024 are significantly more likely to adopt innovative designs than those initiated earlier. Furthermore, within neuroscience, trials in Psychiatry, Substance Abuse, and Mental Health stand out as having greater odds of being innovative than other neuroscience fields (Panel A). The predicted probabilities depicted in Panel B confirm a steady upward trend in innovative trial designs from 2010 to 2024.

Table 4.

Multivariable innovative trial logistic regression model OR (Odds Ratio) with 95% confidence intervals and p-values (Panel A). The model estimated the probability of innovative trials to be published on clinicaltrial.gov according to publication year (Panel B).

LLM validation

Additionally, on a manually classified random sample of 2,000 trials, the LLM classification’s accuracy rate was 94.6% (95% CI = 93.6%−95.5%).

The comparison between the distribution of manual classifications and population data revealed similarities in coverage across major therapeutic areas. The manual sample proportionally represented key areas, such as Hematology and Oncology, Neurology and Neurosurgery, and Psychiatry, reflecting their prominence in the broader dataset (Tables S3 and 1).

Although the Radiology and Diagnostic Imaging fields were underrepresented in the overall dataset (Table 1), they exhibited a higher misclassification rate in the manual sample (Figure S1). The pattern is similar in Dental and Otolaryngology, rare Diseases, and Genetics.

Instead, Neurology and Neurosurgery, which were well-represented in both the manual sample and the broader dataset, displayed no misclassifications (Figure S1).

Discussion

This research highlights that neuroscience, like oncology, is adopting innovative trial designs, including adaptive and Bayesian methodologies. These approaches, aligned with ethical and methodological standards, improve efficiency, cost-effectiveness, and ethical integrity, offering a promising avenue for treating complex psychiatric and neurological disorders²². These advantages align with the broader movement toward precision medicine, where treatment strategies can be personalized based on evolving data^3,22.

The adoption of innovative trials, such as Adaptive and Bayesian Designs, has surged since 2011, reflecting advancements in computational and statistical methods. Technological advancements, particularly in neuroimaging and genetics, have revolutionized our understanding of neurological disorders. These technologies enable researchers to identify biomarkers and genetic markers associated with various conditions, leading to more personalized and targeted treatments²⁸. Innovative and Bayesian designs are promising in neuroscience as they enable real-time adjustments based on accumulating data. For instance, adaptive trials can modify parameters such as sample size or treatment arms in response to interim results, which is crucial in fast-evolving fields like neuroimaging and genetics⁵. Bayesian methods are especially useful in fields with extensive prior research, as they combine existing knowledge with new data, thereby improving the precision and reliability of outcomes⁵. In this framework, innovative Bayesian designs are promising in neuroscience as they enable real-time adjustments based on accumulating data. Adaptive trials can modify parameters, such as sample size or treatment arms, in response to interim results, which is crucial in fast-evolving fields like neuroimaging and genetics²⁹.

Despite their potential, innovative trial designs, such as adaptive and Bayesian approaches, face challenges. In 2007, the European Medicines Agency (EMA) published the “Reflection Paper on Methodological Issues in Confirmatory Clinical Trials Planned with an Adaptive Design,” introducing considerations for confirmatory adaptive trials, including maintaining trial integrity, controlling Type I error, and pre-specified adaptation rules³⁰.

Regulatory agencies emphasize that adaptive designs should not compromise trial integrity. Maintaining blinding and avoiding bias in interim analyses is critical. These designs may be labor-intensive and costly, requiring advanced expertise and simulations to ensure valid conclusions³¹. Bayesian trials offer flexibility and efficiency but require careful planning to meet regulatory standards. Key challenges include selecting and justifying prior distributions, addressing computational complexity, and conducting sensitivity analyses for robustness. Regulatory agencies emphasize potential bias in prior selection and stress transparent reporting for stakeholder interpretation. These demands require advanced expertise, clear communication, and collaboration with regulators³¹.

Regulatory agencies, including the FDA and EMA, have primarily focused on confirmatory trials (Phase III) when guiding adaptive and Bayesian designs³⁰. The FDA’s 2010 guidance reinforced these principles, outlining requirements for interim analyses, blinding, and statistical rigor³¹. Subsequent updates, such as the EMA’s 2019 revised paper³⁷ and the FDA’s 2019 final guidance²⁴, clarified best practices and provided concrete examples to support the application of adaptive designs in Phase III trials. In addition to EMA and FDA guidance, the International Council for Harmonisation (ICH) has played a relevant role. The updated ICH E6(R3) guideline on Good Clinical Practice promotes flexibility and risk-based approaches compatible with adaptive and decentralized designs³². The forthcoming ICH E20 guideline focuses specifically on adaptive clinical trials, providing a global framework for their design and interpretation³³. These initiatives underscore that regulatory authorities not only tolerate but actively encourage innovation in trial design.

Our findings highlight that neuroscience trials often employ these designs in early phases, particularly Phases II and II/III, to facilitate flexibility in optimizing dose-response relationships and assessing early efficacy signals. Early-phase trials focus on understanding disease mechanisms, identifying biomarkers, and determining optimal dosages³⁴. For example, in a study of a new Alzheimer’s treatment, an adaptive trial could adjust dosages in real-time to maximize efficacy while minimizing side effects³⁵. Instead, Bayesian methods can integrate prior knowledge, such as previous research findings on brain activity patterns, to improve the precision of early-phase trials, leading to more informed decision-making and faster progression to later phases^36,37.

This shift reflects broader acceptance of innovative designs beyond confirmatory trials, acknowledging their potential to accelerate drug development while maintaining rigorous standards, even in the exploratory phase. However, robust methodological standards are essential, including comprehensive pre-planning of adaptive strategies, clear pre-specification of adaptation rules, and careful control of biases during interim analyses to ensure trial validity and reproducibility^36,37. The increased use of Phase II/III trials in neuroscience shows these designs are not limited to confirmatory purposes, but connect exploration and validation in clinical research. In line with this trend, EMA’s 2023 Bayesian methods guidelines support the application of innovative designs by providing a framework for their use across all trial phases, emphasizing their potential beyond traditional confirmatory settings³⁸. However, thorough monitoring of study design compliance with recent regulatory recommendations should also be encouraged, particularly for early phase trials in neuroscience and other fields¹⁸.

This research shows that within neuroscience, innovative trials are more prevalent in psychiatry than neurology, likely due to the complex nature of psychiatric disorders. Psychiatric conditions like depression and schizophrenia exhibit high variability in treatment responses, and outcome evaluation necessitates flexible trial designs. Adaptive trials can modify parameters based on interim results, offering personalized treatments. For instance, an adaptive trial of antidepressants can adjust dosages based on patient responses to ensure better outcomes³⁹.

Innovative clinical trials in pediatric care are a critical area of focus. Children have historically been underrepresented in clinical research, leading to a gap in evidence-based treatment for this population⁴⁰. The increased emphasis on pediatric participation in innovative trials is a significant step towards closing this gap and ensuring that children benefit equally from medical advancements⁴¹. Innovative trials in neuroscience are mainly pediatric because of the critical need for effective treatments for neurological conditions in children. Pediatric neurological disorders such as epilepsy, autism, and developmental delays often require early and tailored interventions^42–44. For example, adaptive trials can modify dosage levels and treatment protocols based on a child’s response, which is important given the variability in how children metabolize medications compared to adults⁴⁵.

However, innovative trials targeting the elderly and sex-specific populations remain underrepresented. Elderly patients present several challenges owing to complex and comorbid psychiatric conditions⁴⁶. Innovative trial designs, with their adaptive nature, are well-suited to address these complexities, allowing for real-time modifications based on interim results. Gender-specific research in psychiatry is another area where innovative trial designs offer significant benefits. Women often experience psychiatric disorders differently from men in terms of both prevalence and symptomatology⁴⁷. Innovative trials could help to uncover these differences more effectively, leading to personalized treatment strategies. Despite their potential benefits, these studies are limited, partly because of insufficient industry funding. Extending innovative trials to underrepresented populations would provide numerous benefits tailored to specific study settings⁴⁸. Properly conducting novel trials requires substantial organizational and structural resources, which can pose significant challenges for newly developing countries. This limitation is reflected in the lower representation of trials from these regions, as noted in prior studies that emphasized the resource-intensive nature of innovative trial designs⁴⁹.

Regarding trial outcomes, our survival analysis demonstrates that innovative trials are more likely to remain active over time than traditional trials. This finding aligns with previous research highlighting the flexibility of innovative designs, particularly adaptive methodologies^3,5. Interaction analysis in the therapeutic field-adjusted survival model showed that the benefits of innovative trial designs vary by therapeutic area. In Neurology and Neurosurgery, these designs are linked to a reduced risk of early termination, aligning with the complexity and variability of neurological disorders, where disease progression and patient responses are unpredictable. Adaptive methodologies enable real-time adjustments, such as modifying endpoints or stratifying patient populations⁵⁰, mitigating risks that might otherwise lead to premature trial discontinuation. This variability in outcomes highlights the need to align innovative approaches with each therapeutic area’s specific demands. While Neurology and Neurosurgery may benefit from reduced termination risks, other fields might not. These findings underscore the importance of tailoring trial designs to each field’s specific needs and complexities while maintaining methodological rigor and adaptability.

In this framework, LLM tools can facilitate state-of-the-art monitoring and regulatory compliance of clinical trials in neuroscience and other fields, handling vast amounts of free-text data and documents, and even for therapeutic area-specific research⁵¹. Consequently, we invite the scientific community to investigate the factors behind the prevalence of innovative designs in neuropsychiatric sciences and address the equity challenges in accessing innovative clinical research and compliance with the recommendations of international guidelines.

Limitations and future research developments.

Innovative trial designs can provide flexibility and efficiency; their implementation is often complex and demands advanced statistical expertise, which may hinder their adoption in resource-limited settings⁸.

The term innovative designs in this research encompasses a broad range, including a comprehensive set of terms to capture these methodologies. Some terms, such as “interim analysis” and “group sequential trials,” are widely considered standard in confirmatory trials, particularly in fields like cardiovascular and oncology research. Though classified as part of adaptive strategies, these methodologies have been integral to traditional trial frameworks for decades⁸. Future research could refine this classification by differentiating innovative designs based on their application in early-phase exploration trials versus later confirmatory trials.

Ensuring compliance with methodological standards in innovative trials is a significant challenge. Although this study offers a snapshot of adoption trends across therapeutic areas, particularly in neuroscience, it does not evaluate adherence to ethical and methodological standards. Prior research, such as that of Collignon et al. (2018)⁴⁹, highlights how well-planned adaptive trials can achieve regulatory approval. Future studies should deepen into these aspects and analyze protocols, statistical plans, and regulatory outcomes to understand better how innovative designs align with ethical and regulatory expectations.

The application of innovative designs in rare disease research is noteworthy. These methodologies allow for real-time adaptations and efficient use of small datasets, making them helpful in addressing the peculiar challenges of these studies⁸. Nonetheless, these trials are not sufficiently represented in the overall sample to comprehensively understand their long-term outcomes. The high misclassification rate in rare disease trials emphasizes the complexity and overlapping aspects in this area, underscoring the need for customized classification methods that consider rare disease research characteristics and its reliance on innovative techniques. Further research is needed to establish success indicators for innovative trials in these domains, aligning with regulatory standards.

Although ChatGPT’s classification accuracy is high, it is not perfect, and errors in categorizing trials could affect the study’s findings; a few therapeutic areas showed lower recall, likely due to overlapping terminology or broad scope. These limitations are acknowledged in interpreting results at the category level, but they do not significantly impact overall trends or comparisons.

Moreover, the focus on innovative trials in neuroscience and psychiatry may not fully represent trends in other therapeutic areas, necessitating further research to generalize these findings across different medical fields.

Conclusion

This study highlights the increasing use of adaptive and Bayesian trial designs across therapeutic areas, with an expanding adoption in neuroscience and psychiatry. These fields often involve complex, heterogeneous conditions, such as neurodevelopmental and mental health disorders, where traditional trial designs may be insufficient. Innovative designs can offer greater flexibility in evaluating emerging therapies and tailoring interventions to individual needs.

These approaches are more common in early-phase and pediatric trials but still underused in elderly and sex-specific populations. Innovative neuroscience trials were associated with longer durations and lower early termination risk.

Moreover, using a large language model to classify trials proved to be accurate support for large-scale monitoring of trial design trends.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Wrote Manuscript DA, Designed Research DA, Performed Research VS, Analyzed Data VS, DI, DA, MRK, Coordination DG, DA, writing review and editing LV, RC, MBM, MRK.

Funding

No funding was received for this study.

Data availability

The datasets used and/or during the current study are available from the corresponding author reasonable request.

Declarations

Competing interests

The authors declare no competing interests.