Basket and umbrella trials in pediatric precision medicine: a systematic review of designs, opportunities, and challenges

Abstract

Background

Personalized medicine, driven by genomic insights, has catalyzed the emergence of innovative clinical trial designs such as basket and umbrella trials. These designs are particularly suited for evaluating targeted therapies in biomarker-defined subgroups and rare pediatric conditions where traditional trials face challenges of small sample sizes and disease heterogeneity.

Objectives

This systematic review aimed to characterize and synthesize the current literature on basket and umbrella trial designs in pediatric drug development, with a focus on their methodological, regulatory, and statistical aspects.

Methods

A systematic search was conducted in the electronic databases PubMed, Scopus, and Web of Science to identify all literature related to basket and umbrella trials. A text mining analysis using unsupervised machine learning technique was performed with relevant articles to automatically identify the primary topics within publications on basket and umbrella trials. A systematic search of PubMed, Scopus, and Web of Science identified 1867 records. After screening and eligibility assessment, 28 studies were included in the final review. Topic modelling using Latent Dirichlet Allocation (LDA) was performed on 76 pertinent articles to identify dominant themes. Statistical convergence, topic coherence, and classification accuracy (> 85%) were validated.

Results

A systematic search of PubMed, Scopus, and Web of Science identified 1867 records. After screening and eligibility assessment, 28 studies were included in the final review. Basket trial designs were more prevalent than umbrella trials, particularly in early-phase oncology and rare disease research. Topic modelling using Latent Dirichlet Allocation (LDA) was performed on 76 relevant articles to identify dominant themes. Statistical convergence, topic coherence, and classification accuracy (> 85%) were validated.

Conclusions

Basket and umbrella trials offer substantial advantages for pediatric drug development by increasing trial efficiency, enabling precision targeting, and supporting adaptive decision-making. Their success depends on robust statistical planning, careful use of Bayesian methods, and attention to regulatory guidance.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12874-026-02914-0.

Introduction

In the last several decades, with the advancement of genomic technology, personalised medicine has become a new paradigm in drug development. Based on molecular data researchers can optimize diagnosis, prognosis, and therapeutic decision making. Precision medicine uses information about the genes, proteins, and other features of a person’s genomic profile to diagnose or treat their disease [1].

The development of personalized medicine has necessitated more sophisticated and flexible clinical trial designs. To address this need, basket and umbrella trials have emerged as innovative approaches that allow for the evaluation of therapies in a more stratified and precise manner. Basket trials evaluate the safety and efficacy of a single targeted therapy across multiple diseases or disease subgroups, which are typically grouped based on shared molecular alterations or biomarkers [2, 3]. It enables early signal detection, support tumor-agnostic drug approvals, and help identify promising indications for further study [4]. By the assessment of a novel therapeutic across a broad range of conditions at an early stage, basket trials provide critical proof-of-concept data and inform subsequent go or no-go development decisions [2].

Conversely, umbrella trials are designed to assess the efficacy of multiple targeted therapies within a single disease population [2, 3]. Patients are often screened for specific molecular markers and subsequently enrolled or randomized into distinct sub-studies or treatment arms, each tailored to a particular molecular subtype. This design allows the simultaneous evaluation of several therapeutic hypotheses within a unified trial structure [2, 5]. When umbrella trials incorporate adaptive randomization and allow for the dynamic addition or discontinuation of treatment arms, they may be classified as platform trials, reflecting their flexibility and efficiency in therapeutic evaluation [6–9]. The terms are often used together because a platform trial becomes an umbrella design if it is investigating multiple matched therapies for one disease.

Despite their promising potential, the development of new targeted drugs has been slow and scattered in the pediatric, especially compared with adult indications. Only a few drugs have been approved by the European Medicines Agency (EMA) for children with cancer; meanwhile, the list of possible drivers in this population reaches about 200 oncogenic alterations [10]. The pediatric patient population has been historically excluded from the clinical drug development enterprise for many reasons, including enrolment challenges due to unfortunately small patient populations [11, 12] and lack of market forces that provide the incentive for drug discovery and development [13].

The Food and Drug Administration (FDA) released its draft guidance document outlining recommendations for basket trials and umbrella trials in March 2022, highlighting their support for a wider dissemination of these designs [14]. This guidance specifically recommends developing trial designs that test multiple drugs and/or multiple cancer subpopulations in parallel under a single protocol, using terms like umbrella, basket, or platform to describe these designs. The EMA has also published a similar guidelines relevant to complex trial designs and subgroup analysis [15].

Innovative clinical trial designs, such as basket, and umbrella address the unique challenges of conducting a drug trial, particularly in the context of pediatric, rare diseases and biomarker-defined subgroups. Traditional clinical trials often struggle with small sample sizes, disease heterogeneity, and ethical constraints, making it imperative to adopt more efficient and adaptive approaches.

Given the rarity of pediatric diseases, recruiting sufficient patients for conventional trials is difficult, especially for biomarker-defined subgroups. Additionally, the molecular heterogeneity [16, 17] of pediatric malignancies necessitates personalized treatment strategies. The urgency to develop new therapies is another critical factor, as traditional trial designs are often too slow to meet the pressing need for effective pediatric treatments. Ethical concerns also play a significant role as randomizing children to a placebo or an ineffective standard-of-care arm can be problematic, particularly when better alternatives might exist.

Basket and umbrella offer several advantages over conventional designs. These approaches improve efficiency by allowing the simultaneous evaluation of multiple treatments or patient subgroups within a single trial, reducing both time and cost. They also provide greater flexibility, as treatment arms can be added, removed, or adjusted based on emerging data [3]. Furthermore, these designs expand patient access to promising new therapies and effectively address disease heterogeneity [18–20].

Several real-world examples highlight the effectiveness of these trial designs. The iMATRIX platform facilitates the study of multiple agents across different pediatric tumor types based on their mechanisms of action [21]. The ODYSSEY trial used Bayesian methods to evaluate dolutegravir-based antiretroviral therapy in HIV-positive children, incorporating data from older children to inform results in younger patients [22]. The SNAP trial investigates treatments for Staphylococcus aureus bloodstream infections in both children and adults, utilizing cross-population data to enhance trial efficiency [23]. Other examples include the BELIEVE trial, which offers treatment opportunities for patients with relapsed or refractory disease [24]; the A12PT platform for pediatric immunotherapy screening; and the INFORM2 NivEnt trial, which applies a biomarker-driven, Bayesian adaptive approach to maximize early activity detection while minimizing risks [25].

The successful implementation of these trials depends on robust statistical and methodological approaches. Given the small sample sizes typical in pediatric trials, advanced statistical techniques are required to address challenges such as borrowing information to enhance statistical power, implementing adaptive designs that allow dynamic trial modifications, controlling the risk of type I error inflation, and managing multiple comparisons. In addition, sophisticated methods are necessary to account for variability in treatment responses across patient subgroups and to determine optimal dosing, including approaches such as the modified continual reassessment method (mCRM) [26, 27], and leverage Bayesian frameworks for improved decision-making [20].

Over the last decade, there has been a significant increase in both the interest in and the volume of studies related to basket and umbrella trial designs [28]. Despite the growing adoption of basket and umbrella trials, the statistical methodologies underpinning these designs are diverse and evolving, and a comprehensive synthesis of current approaches remains limited [9, 29]. This highlights the need for a focused review to systematically examine and compare the statistical frameworks used in these complex trial designs.

Primary objective of this systematic review is to synthesize the state-of-the-art methodologies employed in basket and umbrella trials in the context pediatric trials while also identifying key methodological and regulatory gaps. There are very few efforts to characterise the literature on basket and umbrella trials according to its perspective or topic. To address this gap, we systematically classify existing basket and umbrella trials.

Materials and methods

This systematic review was designed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA 2020) checklist [30] and was registered with Open Science Framework (OSF) (DOI: https://doi.org/10.17605/OSF.IO/GEPSQ).

Search strategy

A systematic search was conducted on 9 Dec 2024 in the electronic databases PubMed, Scopus, and Web of Science (WOS) to identify all papers related to basket and umbrella trials (detailed search string for each database is available in Additional file). To ensure a comprehensive search, we used the terms “basket trial”, “umbrella trial”, “pediatric”, and “paediatric”. All retrieved references were imported into Zotero for deduplication and further screening.

Study inclusion and exclusion

Studies were eligible for inclusion if they were original, peer-reviewed research articles with target population < 18 years that explicitly described the design, methodology, or the data analysed from basket, umbrella, or platform trials in any therapeutic area (Table 1). The studies that did not use a basket, umbrella, or platform design, as well as those without a pediatric focus were excluded. Opinion pieces, editorials, commentaries, reviews, conference abstracts, case reports and publications not available in English were excluded.

Table 1.

Population, Interventions, Comparator, Outcome and Study design (PICOS) criteria

Characteristics	Inclusion criteria
Population	Studies with study population under 18 years of age
Interventions	No restrictions
Comparator	No restrictions
Outcome	No restrictions
Study design	Studies using following study designs or data analysed from: 1. Basket trials 2. Umbrella trials 3. Platform trials
Other	▪ Original research and peer-reviewed publications published in English language. ▪ Conference abstracts, opinion pieces, editorials, and case reports, letter will be excluded.

Two reviewers (MRK and DA) independently reviewed all titles and abstracts of the identified records. Full texts of potentially relevant (pertinent) articles were then retrieved and assessed for eligibility. Any discrepancies in study selection were resolved through discussion, and when consensus could not be reached, a third reviewer (MS) was consulted.

Data extraction

Data related to study design, target populations and sample size were extracted independently by two reviewers (MRK and DA) using a standardized, piloted data extraction template. We extracted information on trial design, trial phase, target population, statistical approach, sample size, primary endpoint, and simulation study. Data extracted by the two reviewers were matched and there was no ambiguity present.

Text mining

A text mining analysis using topic modelling was performed with pertinent articles to automatically identify the primary topics within publications on basket and umbrella trials. Topic modelling, an unsupervised machine learning technique, can automatically group words into topics and identify expressions that best describe a collection of documents [31].

Data preprocessing

The titles and abstracts of basket and umbrella trials pertinent papers have been pre-processed. Punctuation, stop words, white spaces, and numbers were removed. Redundant words such as background, objectives, results, conclusion, year, abstract which were repeatedly found in every abstract and do not serve the purpose were removed, and all words were converted to lowercase. Once the text corpus has been cleaned, a Document-Term Matrix (DTM) has been created. A DTM is a matrix, reporting documents (articles) by rows and words by columns; a generic element of DTM is the word count. To detect topics, a Latent Dirichlet Allocation (LDA) [31] analysis has been performed on the DTM matrix of pertinent articles. LDA is a technique leading to the automatic discovery of themes in a collection of documents. The method assumes that each document (article) is a mixture of topics. Documents and words are observed elements, instead topics are latent structures discovered by the LDA algorithm. The method aims to infer the latent topic structure given the words and document. LDA recreates the documents in the corpus by adjusting the relative importance of topics in documents iteratively using a Gibbs sampler algorithm [32]. Gibbs sampling works by performing a random walk. The starting point of the walk is chosen at random; for this reason, it may be useful to discard the first steps (burn-in period). Overall, 10,000 iterations have been considered in the computation, and 100 draws have been discarded as burn-ins. A total of five Markov chains with different starting points were generated.

Number of the topic

The number of topics has been chosen following the maximization criterion of the Deveaud and Griffith metrics [33, 34]. The method is based on the idea of computing distances between pairs of topics over several instances of the model while varying the number of topics. The model iterations are done by varying the number of topics of the LDA model, and then estimating again the Dirichlet distributions. The optimal number of topics is reached when the crossing point of griffith2004 and deveaud curve [33, 34].

The Deveaud metric determines the optimal number of topics by maximizing the distinctness between them. Each topic is represented by top-n most probable words and then it calculates the information deviation (i.e. distance) between all possible topic pairs [33]. Griffiths2004 metric uses Bayesian model to extract the number of topics, the method employs a Markov chain Monte Carlo algorithm, specifically Gibbs sampling. It considers the posterior distribution over the assignments of words to topics [34].

Validation and convergence assessment

The algorithm has been validated on the clinical trial pertinent articles. Furthermore, the overall accuracy has been calculated, comparing the manual and automatic classification. The convergence of the LDA algorithm has been evaluated showing the Log-Likelihood in correspondence of the first 500 iterations [Additional file]. If the Log-Likelihood estimate stabilises after the first few iterations, then the convergence is deemed acceptable [32].

Computations have been performed using R 4.4.0 [35] System with “topicmodels” package [36].

Results

Literature search

The systematic search identifies a total of 1867 number of records (Scopus 1476, Pubmed 343, and Web of Science 48). After removing duplicates, ineligible, books/book sections, conference papers, and datasets, 1693 articles were retained for title and abstract screening. Upon screening of the title and abstract, 76 pertinent articles were identified for full text review. After reading the full text of the articles, 28 were selected for the final review (Fig. 1). The literature on basket and umbrella trials in the pediatric field, including articles specific to the methodological challenges present in basket and umbrella trials, was screened by reviewing the titles and abstracts.

Fig. 1.

Study flowchart

Trend of basket and umbrella trials

Basket and umbrella trial designs are emerging trial designs in clinical research. The innovative trial design offers more suitability for personalised therapies in pediatric setting. There has been almost an exponential increase in the number articles by the years (Fig. 2). In 2015, around 10 years back very few articles were published on basket and umbrella trials related to pediatric patients.

Fig. 2.

Number of publications by year on basket and umbrella trials in pediatric setting

Results synthesis

A systematic search of PubMed, Scopus, and Web of Science identified 1867 records. After screening and eligibility assessment, 28 studies were included in the final review. The majority of which were basket trials (75%) and early-phase studies (82%). The majority targeted pediatric populations under 18 years (64%), while the remainder included both pediatric and adult participants. Bayesian methods were applied in 35% of trials, frequentist approaches in 50%, and hybrid (Bayesian + frequentist) methods in 14%, reflecting the increasing uptake of Bayesian designs in small-sample pediatric settings. Sample sizes varied considerably, ranging from 16 to 1000 participants, although most trials enrolled fewer than 200. Objective response rate (ORR) was the most frequently reported primary endpoint. Nearly half of the studies (46%) incorporated simulations, predominantly within Bayesian frameworks, to assess operating characteristics and borrowing strategies (Table 2). Overall, there has been a significant increase in the number of articles on basket and umbrella trial since 2015 and also the findings highlights a growing shift toward Bayesian and adaptive approaches in pediatric oncology and rare disease trials.

Table 2.

Basic characteristics of the selected studies

Study ID	Title	Trial Design	Phase	Target Population	Statistical Methods	Sample Size (overall)	Primary Endpoint	Simulation Performed
Mahar et al. 2023 [23]	A blueprint for a multi-disease, multi-domain Bayesian adaptive platform trial incorporating adult and paediatric subgroups: the Staphylococcus aureus Network Adaptive Platform trial	Platform	Phase 2	Pediatric and adults	Bayesian	844	All-cause mortality	Yes
Uguen et al. 2023 [21]	Advancing drug development in pediatric oncology, a focus on cancer biology and targeted therapies: iMATRIX platform	Platform	Phase 1/2	Pediatric and adults	Bayesian and Frequentist	214	Safety	Yes
Desai et al. 2022 [37]	Entrectinib in children and young adults with solid or primary CNS tumors harboring NTRK, ROS1, or ALK aberrations (STARTRK-NG).	Basket	Phase 1/2	Pediatric and adults	Frequentist	43	RP2D (PK) and ORR	Yes
Perreault et al. 2019 [38]	A phase 2 study of trametinib for patients with pediatric glioma or plexiform neurofibroma with refractory tumor and activation of the MAPK/ERK pathway: TRAM-01.	Basket	Phase 2	Pediatric (< 18 years)	Frequentist	150	ORR	No
Gounder et al. 2020 [39]	Tazemetostat in advanced epithelioid sarcoma with loss of INI1/SMARCB1: an international, open-label, phase 2 basket study.	Basket	Phase 2	Pediatric and adults	Frequentist	62	ORR	No
Briggs et al. 2022 [40]	Estimating Long-Term Survival Outcomes for Tumor-Agnostic Therapies: Larotrectinib Case Study.	Basket	Phase 1/2	Pediatric and adults	Frequentist	102	ORR	No
van Tilburg et al. 2020 [25]	INFORM2 NivEnt: The first trial of the INFORM2 biomarker driven phase I/II trial series: the combination of nivolumab and entinostat in children and adolescents with refractory high-risk malignancies.	Basket	Phase 1/2	Pediatric (< 18 years)	Bayesian	128	RP2D	No
Le Teuff et al. 2020 [41]	Phase II study of temozolomide and topotecan (TOTEM) in children with relapsed or refractory extracranial and central nervous system tumors including medulloblastoma with post hoc Bayesian analysis: A European ITCC study.	Basket	Phase 2	Pediatric and adults	Bayesian and Frequentist	91	ORR	No
Mangat et al. 2018 [42]	Rationale and Design of the Targeted Agent and Profiling Utilization Registry (TAPUR) Study.	Basket	Phase 2	Pediatric and adults	Frequentist	1000	ORR	No
Moore et al. 2021 [22]	ODYSSEY clinical trial design: a randomised global study to evaluate the efficacy and safety of dolutegravir-based antiretroviral therapy in HIV-positive children, with nested pharmacokinetic sub-studies to evaluate pragmatic WHO-weight-band based dolutegravir dosing.	Basket	Phase 2/3	Pediatric (< 18 years)	Bayesian and Frequentist	792	Clinical outcome	No
Simon et al. 2016 [43]	The Bayesian basket design for genomic variant-driven phase II trials.	Basket	Phase 2	Pediatric and adults	Bayesian	70	Probability of response	Yes
Ishimaru et al. 2024 [24]	Platform trial for off-label oncology drugs using comprehensive genomic profiling under the universal public healthcare system: the BELIEVE trial.	Platform	Phase 2/3	Pediatric and adults	Bayesian	80	ORR	No
Italiano et al. 2020 [44]	Larotrectinib versus Prior Therapies in Tropomyosin Receptor Kinase Fusion Cancer: An Intra-Patient Comparative Analysis.	Basket	Phase 1/2	Pediatric and adults	Frequentist	72	Progression-free survival (PFS)	No
Wei et al. 2024 [45]	A Bayesian platform trial design with hybrid control based on multisource exchangeability modelling	Platform	Phase 2/3	Pediatric (< 18 years)	Bayesian	150	Probability of response	Yes
Wang et al. 2024 [46]	rBMA: A robust Bayesian Model Averaging Method for phase II basket trials based on informative mixture priors	Basket	Phase 2	Pediatric (< 18 years)	Bayesian	30	ORR, DCR, PFS	Yes
Bouffet et al. 2023 [47]	Efficacy and Safety of Trametinib Monotherapy or in Combination with Dabrafenib in Pediatric BRAF V600-Mutant Low-Grade Glioma	Basket	Phase 1/2	Pediatric (< 18 years)	Frequentist	139	RP2D	No
Ouma et al. 2022 [48]	Bayesian modelling strategies for borrowing of information in randomised basket trials	Basket	Phase 2	Pediatric (< 18 years)	Bayesian	336	Change in a composite score	Yes
Turner et al. 2022 [49]	Borrowing information across patient subgroups in clinical trials, with application to a paediatric trial	Basket	Phase 3	Pediatric (< 18 years)	Bayesian	792	Absolute change of 10%	Yes
He et al. 2022 [50]	Efficiency of a randomized confirmatory basket trial design constrained to control the family wise error rate by indication	Basket	Phase 3	Pediatric (< 18 years)	Frequentist	483	PFS	Yes
Subbiah et al. 2022 [51]	Tumour-agnostic efficacy and safety of selpercatinib in patients with RET fusion-positive solid tumours other than lung or thyroid tumours (LIBRETTO-001): a phase 1/2, open-label, basket trial	Basket	Phase 1/2	Pediatric (< 18 years)	Frequentist	45	ORR	No
Schöffski et al. 2021 [52]	Long-term efficacy update of crizotinib in patients with advanced, inoperable inflammatory myofibroblastic tumour from EORTC trial 90,101 CREATE	Basket	Phase 2	Pediatric (< 18 years)	Frequentist	20	ORR	No
Mu et al. 2021 [53]	A Bayesian adaptive phase I/II platform trial design for pediatric immunotherapy trials	Platform	Phase 1/2	Pediatric (< 18 years)	Bayesian	78	Optimal biological dose (OBD)	Yes
Jiang et al. 2021 [54]	Shotgun: A Bayesian seamless phase I-II design to accelerate the development of targeted therapies and immunotherapy	Basket	Phase 1/2	Pediatric (< 18 years)	Bayesian	24	Safe and effective	Yes
Jiao et al. 2019 [55]	Utilizing shared internal control arms and historical information in small-sized platform clinical trials	Platform	Phase 2	Pediatric (< 18 years)	Bayesian and Frequentist	230	Response rate	Yes
Dhaenens et al. 2024 [56]	Platform trial design for neurofibromatosis type 1, NF2-related schwannomatosis and non-NF2-related schwannomatosis: A potential model for rare diseases	Platform-Basket trial	Phase 1/2	Pediatric (< 18 years)	Frequentist	80	Binary response	No
Kerioui et al. 2024 [57]	New clinical trial design borrowing information across patient subgroups based on fusion-penalized regression models	Basket	Phase 2	Pediatric (< 18 years)	Frequentist	16	ORR	Yes
Plotkin et al. 2024 [58]	Brigatinib in NF2-Related Schwannomatosis with Progressive Tumors.	Basket	Phase 2	Pediatric (< 18 years)	Frequentist	40	Radiographic response in target tumors	No
Vo et al. 2024 [59]	Phase II Study of Ulixertinib in Children and Young Adults With Tumors Harboring Activating Mitogen-Activated Protein Kinase Pathway Alterations: APEC1621J of the National Cancer Institute-Children’s Oncology Group Pediatric MATCH Trial	Basket	Phase 2	Pediatric (< 18 years)	Frequentist	20	ORR	No

Phase 1/2 and Phase 2/3 suggest seamless trial design

RP2D Recommended Phase 2 Dose, PK Pharmacokinetics, ORR Objective response rate, PFS Progression free survival, DCR Disease Control Rate

For topic modelling we utilized LDA to extract latent thematic structures across the 76 relevant articles. Upon reliability assessment of the model’s thematic extraction, the classification output was cross validated against a manual thematic analysis of a representative subset of the corpus, achieving a classification accuracy exceeding 85%.

To explore these findings in more detail we structured the synthesis around the main themes identified in the review. In the sections that follow, we describe each of theme, its representation among the included studies, and the key insights drawn from the synthesis.

Sample size

Across the reviewed studies, sample sizes ranged from16 to 1000 participants depending on the phase of the trial, number of groups and type of study. Several studies justified their sample sizes or adapted their designs based on specific considerations related to the pediatric population, primarily due to the rarity of certain diseases in children and the associated recruitment challenges [22, 37, 38, 41, 45, 49, 53, 56]. Some of them adopted traditional frequentist approaches based on predefined power and significance levels [22, 50], whereas exploratory studies employed flexible, feasibility-driven sample sizes approach [24, 51]. A subset of them applied Bayesian methods that incorporated prior information and allowed adaptive adjustments [43, 48] and some emerging hybrid designs combined frequentist and Bayesian principles [45]. Overall, the literature demonstrates diverse strategies shaped by both methodological aims and practical constraints.

Safety and efficacy endpoints

Endpoint selection is a pivotal component of clinical trial design, carefully tailored to the therapeutic area and specific objectives of the study. Efficacy endpoints, which directly estimate the treatment’s impact on the disease or condition were most prevalent across the reviewed literature for example objective response rates (ORR) [24, 37–42, 46, 52, 57, 59] progression-free survival (PFS) [44, 46, 50] complete response (CR) or partial response (PR) [25], radiographic response [58], binary endpoints [23, 56] such as all-cause mortality, responder, and composite indicator as an endpoint [48]. In HIV trials [22, 49], authors adapted to the unique challenges of the disease, focusing on virological and clinical failure, where the treatment fails to suppress the virus, and clinical failure, marked by new or recurrent AIDS-defining events signalling disease progression. In addition to efficacy, studies reporting safety endpoints [21, 25, 37, 47, 53, 54] are aiming to find the right dose, assessing adverse events, their severity, and their relationship to the treatment, ensuring participant safety.

Multiple comparisons

As the basket trials inherently involve multiple testing because each indication analysis represents a separate hypothesis test. As the number of tests increases, so does the probability of observing at least one false-positive result simply by chance. Some of the selected studies notice the multiple comparison or multiple hypothesis testing and account for its consequences [22, 50, 57]. In the event of multiple hypothesis testing to control the type 1 error rate, the “Shotgun” design proposed CBHM specifically to address the tendency of standard BHM to produce inflated type I errors when drug efficacy is heterogeneous [54]. To address the biased inference and inflated error rates in heterogeneous scenarios the rBMA method calibrates its posterior probability thresholds to specifically set the family-wise type I error rate (FWER) to a target level under the global null hypothesis [46, 50].

Bayesian adaptive trials

Bayesian adaptive designs provide the statistical framework that enables the flexibility and efficiency required for basket and umbrella trials. Bayesian methods are quite frequently used in basket and umbrella trials. Studies have shown that Bayesian methods offer flexibility to handle complex trial designs like platform trials [45, 53], basket trials [43, 48, 54], and seamless phase I-II designs [53, 54]. They can accommodate multiple endpoints, different dose levels, and evolving research questions within a trial [53]. Bayesian adaptive designs often incorporate interim analyses enabling real-time trial modifications. Based on posterior probabilities, trials can stop early for futility or efficacy [43, 45, 48], drop ineffective treatment arms [53], adaptively randomize patients to promising treatments [23, 53], or adjust sample sizes as needed [50].

Bayesian models

Specific Bayesian models used in the context of basket and umbrella trials include Bayesian hierarchical model [21, 23, 45, 50, 55, 57], Bayesian hierarchical model with multisource exchangeability modelling [23, 45] Bayesian logistic model [53, 54], Bayesian model averaging (BMA) [57] and robust Bayesian model averaging (rBMA) [46].

A brief summary table of the different Bayesian models and their intended use cases.

Bayesian Model	Intended Use Case / Application
Bayesian Hierarchical Model (BHM) [46, 57]	Information borrowing across subgroups to increase statistical power, particularly in small-sample oncology trials.
Clustered BHM (CBHM) [54]	Minimizing bias and inflated Type I errors in basket trials where drug efficacy varies significantly across tumor types.
Bayesian Model Averaging (BMA) [46, 54]	Handling uncertainty about whether subgroups are related or independent when the relationship is not known a priori.
Robust BMA (rBMA) [46]	Robust information borrowing in trials with mixed endpoints or significant potential for prior-data conflict.
Multisource Exchangeability (MEM) [45]	Facilitating hybrid control arms by borrowing data from multiple historical sources based on verified similarity.
Exchangeability-Nonexchangeability (EXNEX) [46]	Allowing specific groups to share information while others remain distinct to protect against heterogeneity.
Meta-Analytic-Predictive (MAP) [45, 55]	Leveraging documented historical control information to reduce the number of patients required for a concurrent control arm.
Commensurate Predictive Prior (CPP) [48]	Borrowing over treatment effects (TEB) or groupwise responses (TRB) in randomized basket trials.
Cure Rate Models [53]	Modelling time-to-event endpoints (e.g., PFS) in immunotherapy trials where long-term durable responses are common.
Continual Reassessment Method (CRM / mCRM) [21]	Sophisticated Optimal Biological Dose (OBD) finding when monotonic assumptions are challenged.
Bayesian Optimal Interval (BOIN) [53]	Phase I dose-finding to quickly and safely identify the Maximum Tolerated Dose (MTD) or Recommended Phase II Dose.

Bayesian adaptive design in early phase

For early-phase trials, Bayesian adaptive phase I/II designs seamlessly integrate dose-finding (phase I) and efficacy evaluation (phase II) to optimize treatment selection. The commonly used approach is the Bayesian optimal interval (BOIN) design and modified continual reassessment model (mCRM) applied in Phase I trials to explore the dose space and collect preliminary toxicity and efficacy data to establish the recommended pediatric dose (RP2D) [21, 53, 54].

Information borrowing

Clinical trial designs, especially basket and platform trials, utilize information borrowing strategies to enhance statistical power and efficiency. These approaches are particularly valuable when addressing small sample sizes or rare conditions, which are often encountered in pediatric trials. The various borrowing methods such as across strata or subgroup borrowing [43, 45, 49, 54, 57], historical information or historical control [45, 55], and adaptive borrowing [49] leverage Bayesian and frequentist techniques to implement information borrowing.

Type of priors

Bayesian models in basket and umbrella use various types of prior distributions to incorporate existing knowledge, manage uncertainty, and borrow information across subgroups. The selected studies have used priors such as non-informative prior [45, 46, 48, 54], minimally informative (vague) gamma, normal and inverse gamma [53], beta-binomial [54], several types of informative prior such as slab and spike while borrowing information from subtrials [48, 53], historical prior power priors to discount the influence of historical data [41, 49, 53], skeptical and enthusiastic prior [41], normal and inverse-gamma prior [2, 20, 23, 48], flat normal [49, 53] and expert elicitation [49].

Dose finding and dose escalation

The selected studies have explored various dose-finding and dose-escalation methodologies within a seamless Phase I/II design, including employment of the conventional 3 + 3 design [21, 25, 37, 53], the rolling six design [37, 59], and the Bayesian optimal interval (BOIN) design [53, 54], modified continual reassessment method (mCRM) [21]. Additionally, certain studies determined the recommended Phase 2 dose (RP2D) based on pharmacokinetic/pharmacodynamic (PK/PD) data and dose-limiting toxicities (DLTs) [25, 37, 47].

The 3 + 3 design is a widely used, simple rule-based dose-escalation method, typically enrolling patients in cohorts of three and adjusting doses based on observed dose-limiting toxicities (DLTs). In contrast, CRM and BOIN methods are model based by continuously update toxicity probabilities using Bayesian modelling. These methods offer superior accuracy in estimating the optimal dose and allow integration of prior knowledge, such as adult data, which is especially beneficial in pediatric trials.

Regulatory compliance

Several studies cite adherence to international standards such as ICH-GCP, the Declaration of Helsinki, and guidelines from the FDA and EMA [21, 25, 37–39, 47, 52], yet also acknowledge the lack of unified, consensus-based protocols. The SNAP trial explicitly calls out the absence of consensus checklists for essential content in platform trials [23].

Topic modelling analysis

The algorithm was applied to a corpus of 76 articles. Once the Latent Dirichlet Allocation (LDA) algorithm was validated, it produced two key outputs: a distribution of words associated with each topic and the probability of topic assignment for each article. The final four-topic model was selected based on both algorithmic output and interpretability. Topic names were assigned after reviewing the top 10 most representative terms under each topic.

Initially, topic modelling using the Griffiths and Deveaud metrics suggested an optimal solution between four and five topics (Additional file: Figure S2). However, upon closer examination of the word distributions and adherence to the principle of mutual exclusivity, a four-topic model was deemed more appropriate. To improve topic coherence, commonly used but semantically redundant words—such as “will,” “object,” “month,” “within,” “year,” “conclus,” “result,” “number,” “base,” “author,” “overall,” “also,” “background,” “first,” “part,” “without,” and “popul”—were removed from the LDA input.

Clinical trial design and implementation

Focuses on the structure and execution of clinical trials, especially in pediatric and cancer-related studies. It includes aspects like trial protocols, randomization, participant outcomes, and overall efficacy assessment.

Targeted therapy and application

Covers the development and application of targeted cancer therapies based on molecular and genomic insights. Emphasizes personalized treatment, biomarkers, pediatric relevance, and clinical benefit.

Statistical methodologies in innovative trials

Highlights modern clinical trial frameworks like basket trials and Bayesian methods. Focus is on efficient design, information sharing across studies, data analysis techniques, and statistical modelling.

Patient outcomes and treatment evaluation

Explores evaluation metrics for cancer therapies, patient responses, dose strategies, survivorship, and disease progression. It includes mutation profiling and primary outcome measures in clinical settings.

Based on the word distributions associated with each topic as generated by the model (Table 3), appropriate titles were assigned to each topic. As previously noted, there has been a marked increase in the overall volume of literature concerning basket and umbrella trials; this trend is further evident in Fig. 3, which illustrates the growth across individual sub-themes.

Table 3.

Top-10 words distribution for each LDA topic. In bold are represented the most important words

	Clinical Trial Design and Implementation	Targeted Therapy and Application	Statistical Methodologies in Innovative trials	Patient Outcomes and Treatment Evaluation
1	trial	cancer	trial	patient
2	studi	therapi	design	tumor
3	clinic	drug	basket	respons
4	children	target	data	phase
5	particip	patient	inform	surviv
6	protocol	treatment	treatment	dose
7	outcom	pediatr	control	safeti
8	random	molecular	bayesian	rate
9	treatment	develop	approach	efficaci
10	research	clinic	analysi	progress

Fig. 3.

Cumulative trend of publications for each topic by year

Following topic assignment by the LDA algorithm, each document (i.e., article) was manually reviewed by the author and co-authors to verify the classification. This manual validation revealed that 11 articles out of 76 had been incorrectly classified, resulting in an overall classification accuracy exceeding 85% (Additional file: Table S2). It is important to note that this level of performance was achieved using only the titles and abstracts of the articles as input for the analysis.

Discussion

The systematic review table highlights that the majority of selected studies employ basket trials followed by platform trials and a smaller subset of umbrella or hybrid designs. These trials are predominantly conducted in early-phase settings (Phase 1/2 or Phase 2), reflecting their exploratory or biomarker-driven objectives.

Statistically, Bayesian methods are frequently adopted either alone or integrated with frequentist approaches to enhance flexibility in dose-finding, subgroup analysis, or borrowing information across cohorts. The most common primary endpoints include ORR, safety/tolerability metrics (e.g., MTD/RP2D), and survival outcomes like PFS. Notably, nearly half of the trials incorporate simulation studies to validate design feasibility or optimize operating characteristics, emphasizing the methodological rigor in these complex adaptive frameworks.

Although these trial designs are currently concentrated within oncology and rare disease research underscoring their utility for heterogeneous and underserved patient populations their foundational principles are broadly applicable. By stratifying patients according to shared biological or molecular mechanisms rather than traditional disease labels, these approaches hold significant promise for diverse therapeutic areas, including HIV [22], bloodstream infections [23], autoimmune disorders, and rare hereditary syndromes such as neurofibromatosis [48, 56].

The applications of basket and umbrella trials in the pediatric setting have increased over the years. This can be attributed to the fact that these innovative trial designs offer more flexibility, less patient accrual, and faster decision making. As we can see, basket trial design is more popular than umbrella or platform trial, the reason could be it is more adoptable in oncology, and rare diseases.

The recent increase in the adoption of innovative trial designs has been attributed not only to their inherent flexibility but also to the growing integration of Bayesian methodologies in clinical research. Danila and colleagues similarly observed this dual driver, noting that Bayesian frameworks have become central to the operationalisation of innovative trial designs [60]. In our review, 14 out of 26 studies (Table 2) reported the use of Bayesian frameworks in the context of basket and umbrella trials. However, it is important to note that this figure includes not only implemented clinical trials but also methodological and simulation-based studies. As such, these findings should not be interpreted as a direct measure of real-world adoption. Rather, they reflect a growing interest in and methodological development of Bayesian approaches within this domain.

Selected studies collectively emphasize that innovative designs methodologies, particularly dose finding, Bayesian adaptive and information-borrowing, are essential for advancing drug development in pediatric and rare disease settings. They offer compelling advantages in efficiency, precision, and patient benefit, but their successful implementation requires meticulous statistical planning, rigorous validation through simulation, and transparent communication to navigate the inherent complexities and potential pitfalls, especially concerning statistical validity and ethical considerations.

Various Bayesian approaches offer tailored advantages for innovative trial designs. Bayesian methods allow information borrowing across subgroups can reduce the sample size requirement and enhance statistical power, but inappropriate or excessive borrowing can inflate the Type I error. If information is borrowed from a subgroup where the treatment is effective to a subgroup where it is not, it can artificially inflate the perceived efficacy in the ineffective subgroup, leading to a false positive conclusion for that specific subgroup. To mitigate this issue, the degree of borrowing, often an ethical necessity in rare disease trials, must be carefully calibrated based on the degree of exchangeability between subgroups.

Borrowing information across age groups in clinical trials aims to ethically strengthen pediatric evidence while minimizing risk. It reduces the number of children exposed to ineffective treatments and avoids reliance on informal adult extrapolation by using formal Bayesian methods, enabling faster, evidence-based access to therapies. However, ethical concerns arise if biological (e.g., drug metabolism, development) or social differences (e.g., adherence patterns) make age groups non-comparable.

Different Bayesian method have been employed to address the risk of Type I error inflation and information borrowing constraints in the reviewed studies. In the basket trials, Bayesian methods promote efficient data sharing across subgroups, which can lead to surprisingly large posterior probabilities of drug activity even with limited data in a specific stratum [43, 54]. However, the potential of getting inflated Type I error rates or biased estimates persist if information is borrowed inappropriately, especially when there is unacknowledged or significant heterogeneity between subgroups or between historical and current data [46]. In umbrella and platform trials, they enable simultaneous testing of multiple treatments across populations, improves inference by leveraging information across different disease and or population subgroups, including allowing adult data to support inference for pediatric populations and vice versa [21, 23]. However, the potential for confounding is present if eligibility of the participant changes over time or lack of randomization [23]. Multisource Exchangeability Modelling (MEM) and hybrid control models explicitly account for variability across historical datasets and allow pre-specified borrowing limits, which simplifies communication with collaborators and regulatory agencies, yet it may inflate error rates when internal and historical data differ [45]. Robust Bayesian Model Averaging (rBMA) accommodates uncertainty between indications, enhances power, and improves Type I error control through inclusion of non-informative priors but it is constrained to binary endpoints and is most suitable for early-phase trials [46]. Treatment Effect Borrowing (TEB) and Treatment Response Borrowing (TRB) both boost efficiency; TEB is more robust in heterogeneous settings, while TRB excels when effects are consistent across subtrials. Choosing between TEB and TRB can affect the results as TRB perform poorly in highly heterogeneous settings [48]. Clustered Bayesian Hierarchical Models (CBHM) balance power and error control by clustering similar indications and borrowing within subgroups but may still inflate errors in small samples or limit borrowing if error control is too strict [54]. Fusion-penalized regression models offer easy implementation, fast computation and do not require any prior information. However, they do not adapt borrowing dynamically to the heterogeneity and may overlook estimation uncertainty when making decisions about drug activity [57].

Many basket and umbrella trials are non-randomized in nature, particularly in early phases and for rare diseases, is often a pragmatic choice driven by ethical considerations, challenges in patient recruitment, and the primary objectives of early-phase research. However, this design choice carries implications such as an increased risk of bias, difficulties in establishing definitive causality, and the need for careful statistical methodologies and potential confirmatory studies.

Studies focusing exclusively on pediatric populations typically address rare, biologically distinct childhood conditions, often necessitating small sample sizes and statistical borrowing across pediatric subgroups. These trials prioritize child-specific dosing, formulations, safety, and developmental or quality-of-life endpoints, and are often inherently exploratory. In contrast, trials incorporating mixed (pediatric and adult) populations emphasize tumor-agnostic therapies, leveraging larger adult cohorts and cross-age statistical borrowing to satisfy standard oncology endpoints. Our review identified several protocols that successfully integrate both populations, suggesting that recruitment across all age groups is feasible when disease characteristics are shared rather than age-specific. Integrating these populations within a single master protocol offers distinct strategic advantages, most notably the ability to accelerate evidence generation through information sharing and improved accrual rates [49].

Bayesian adaptive designs offer a powerful and flexible framework for clinical trials often require small sample sizes and permitting early termination of ineffective treatment, particularly beneficial in pediatric trials and rare diseases where patient accrual is very challenging. Their ability to incorporate informative priors from different sources, such as historical trial data or expert opinions can effectively reduce the size of concurrent control arms. However, eliciting informative priors can be challenging and requires close collaboration with clinical experts. There is also a need to ensure that priors are scientifically valid and not overly informative. Moreover, Bayesian adaptive designs can simultaneously consider multiple endpoints, such as immune response, toxicity, and efficacy, allowing for a more comprehensive assessment of treatments, especially in immunotherapy trials.

Basket and umbrella trials inherently involve multiple comparisons, raising concerns about inflated Type I error rates. Multiple testing without proper adjustments inflates the Family-Wise Error Rate (FWER). He et al. [50] specifically address the need to control the FWER by indication. Their modified basket trial design incorporates several steps to manage this increased risk of false positives across multiple indications. However, relatively few studies have explicitly addressed the increased risk of false positives, despite the fact that such risks are intrinsic to any scenario involving multiple hypothesis testing. For example, Moore et al. [22] identify this issue but do not provide a detailed resolution.

Studies suggest a shift from traditional rule-based dose-finding methods to more advanced Bayesian adaptive designs that handle complex outcomes and efficiently leverage information. The 3 + 3 design remains popular due to its simplicity but is inefficient, lacking statistical modelling of dose-toxicity and potentially exposing patients to suboptimal doses [21, 25, 37, 53, 54, 59]. The Continual Reassessment Method (CRM) improves dose accuracy through Bayesian modelling and prior knowledge but is complex and requires statistical expertise [21]. The Bayesian Optimal Interval (BOIN) design offers a middle groundsimpler than CRM with good performance and built-in overdose control, though still more complex than rule-based methods [53, 54]. PK-guided approaches are valuable in pediatrics for optimizing doses with limited data and extrapolating from adults, but their reliability hinges on accurate models, with errors likely if assumptions or variability are mis-estimated [21, 37].

Several articles collectively highlight persistent regulatory gaps in drug development, particularly in areas such as precision medicine, rare diseases, pediatric drug development, and innovative clinical trial designs. A key issue is the fragmented and evolving nature of regulatory frameworks. For instance, the BELIEVE trial in Japan demonstrates the lack of integration between genomic testing and the actual administration of matched targeted therapies, despite national coverage of profiling tests [24]. Similarly, the ODYSSEY trial underscores a significant delay in pediatric drug formulations, especially in low-resource settings, and illustrates how regulatory guidelines must adapt swiftly to emerging safety concerns [22].

The TAPUR study identifies a critical lack of structured mechanisms and incentives to collect outcome data for off-label applications of FDA-approved therapies informed by genomic profiling [42]. Basket and platform trials, as discussed in studies like those by He et al., the SNAP trial, and iMATRIX [21, 23, 50], face regulatory ambiguity, particularly around control of family-wise error rates, subgroup analysis, and methodological transparency. Regulatory bodies such as the EMA and FDA have not yet reached a global consensus on best practices for these complex trial formats.

Platform trials for conditions like Neurofibromatosis, along with studies such as Larotrectinib and Tazemetostat, expose significant data scarcity, such as incomplete natural history data [56], small sample sizes [39, 40], and unvalidated patient-reported outcome measures (PROMs) across languages and contexts [56]. Pediatric trials continue to suffer from slow regulatory progress and over-reliance on data extrapolated from adult studies. In response, some trials have adopted Bayesian methods to strengthen pediatric-specific decision-making.

When regulatory pathways are fragmented or lagging behind the pace of medical innovation, the resulting design choices are typically reactive designed to navigate ambiguity while maintaining scientific and ethical standards. In essence, the regulatory gap is forcing a move toward “agile evidence generation.” Researchers are no longer just designing trials to test a drug; they are designing trials to build the regulatory evidence framework simultaneously, often by necessity rather than preference.

The application of topic modelling using Latent Dirichlet Allocation (LDA) provided a structured approach to classifying the selected studies into distinct thematic subcategories related to basket and umbrella trial literature. By identifying the optimal number of topics, the LDA model facilitated an organized synthesis of the evidence base, enabling clearer thematic differentiation across the corpus. Notably, the model achieved a classification accuracy exceeding 85% when compared to manual categorization, suggesting a high degree of reliability. This demonstrates that topic modelling can serve as a valuable methodological tool in systematic reviews, particularly when the objective is to highlight and analyze subthemes within a broader research area.

Conclusion

This review highlights the increasing use of basket and umbrella trial designs, particularly in oncology, pediatrics, and rare diseases. When combined with Bayesian methodologies, these designs offer important advantages over traditional trials, including greater efficiency, reduced sample size requirements, and increased flexibility. Basket trials were more prevalent, likely due to their suitability for biomarker-driven contexts. The growing adoption of Bayesian adaptive approaches reflects a broader shift toward more flexible and data-efficient trial methodologies, particularly valuable in small pediatric populations.

Methodologically, several challenges remain. Non-randomized early-phase designs may limit causal inference, and inappropriate information borrowing across subgroups can inflate Type I error if heterogeneity is not properly addressed. The complexity of Bayesian modelling, along with multiplicity control across multiple arms or endpoints, requires careful statistical planning and transparent reporting. From a practical perspective, regulatory fragmentation, operational complexity of master protocols, and considerations around combining adult and pediatric populations may influence trial feasibility and design decisions. Assumptions of exchangeability across age groups must be critically evaluated.

Future review would benefit from more formally stratifying analyses by study type, distinguishing primary trial manuscripts from methodological and simulation-based articles to better differentiate trends in actual trial conduct from developments in the methodological literature. Additionally, the topic modelling approach may expand by incorporating full-text analyses or clustering studies according to specific design characteristics to further refine methodological insights.

Overall, basket and umbrella trials represent valuable tools for advancing pediatric precision medicine, provided that both methodological rigor and practical challenges are carefully addressed.

Limitations and challenges

This systematic review has several limitations that should be considered when interpreting the findings.

• Our search strategy focused specifically on the terms “basket” and “umbrella” and did not include “platform” trials. Given the substantial conceptual overlap and frequent mislabelling of master protocols in the literature, some relevant studies may have been missed, particularly those self-identifying primarily as platform trials or using hybrid designs that combine basket, umbrella, and platform elements.

• Master protocols, especially umbrella trials, are susceptible to reporting fragmentation, where sub-studies are published independently without clear reference to the parent protocol. This may have led to incomplete identification or characterization of certain trials within our review.

• The literature search was limited to three electronic databasesPubMed, Scopus, and Web of Science and did not include other sources such as Embase, ClinicalTrials.gov, or grey literature databases. This may have led to the omission of relevant studies, particularly those involving unpublished data or ongoing trials.

• After full-text screening 28 were included in the final qualitative synthesis. This is relatively small sample size may limit the generalizability of the findings, especially across different therapeutic areas and regulatory contexts.

• Pooling different types of studies (primary trials, methodological work, and secondary analyses) may affect the interpretation of overall trends and could overrepresent emerging statistical approaches relative to their current use in practice.

• Although topic classification was validated manually with over 85% accuracy, some degree of misclassification or thematic overlap is inevitable, given the complexity and multidimensionality of the articles.

Despite these limitations, the review provides a structured and comprehensive overview of the methodological landscape for basket and umbrella trials in pediatric precision medicine contexts, while highlighting key areas for future development and standardization.

Supplementary Information

Additional file 1. ^{(57.6KB, docx)}

Authors’ contributions

DA conceived the idea and study design. MRK conducted the literature search and data extraction. MRK, DA, and MS screened the articles and assessed study quality. MRK drafted the manuscript, and MRK and DA performed the data analysis and interpretation. MS, IB, LV, PB, SB, and DG critically reviewed and provided input on the manuscript. DA supervised the study throughout its conduction. All authors read and approved the final manuscript.

Funding

This research did not receive any specific grant or funding.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.