Clinical registry metadata as a hidden bottleneck in AI-driven drug discovery: a computational audit of translational phase data in glioma research Authors: Iryna Oliynyk Published: 2026 ABSTRACT AbstractClinical translation in glioma and glioblastoma remains inefficient despite advances in computational drug discovery. An underrecognized contributor to this gap is structural degradation of clinical registry metadata. We analyzed the complete set of 2.357 glioma-related clinical trial records available in the WHO ICTRP registry at the time of extraction (3 January 2026). A deterministic Python-based validation pipeline was developed to normalize phase and study-type annotations while distinguishing technical voids, methodological non-applicability, and structurally inconsistent entries. Two quantitative indices were introduced: the Reporting Gap (), reflecting phase metadata completeness, and the Maturity Ratio (), describing the balance between early and late translational stages. Phase annotation showed substantial structural degradation, with large fractions of records lacking machine-interpretable phase labels, whereas study-type fields demonstrated high completeness but severe terminological fragmentation. This asymmetry indicates a systemic mismatch between real-world research designs and the classical phase ontology. Consequently, clinically relevant evidence remains invisible to algorithms, which may lead to biased AI-based translational evaluation. The results suggest that semantic validation of registry metadata is a critical prerequisite for robust integration of clinical data into computational drug development cycles.Keywords: Clinical trial registries, Metadata quality, Reporting completeness, Translational research, Machine learning bias, GlioblastomaSubject terms: Cancer, Computational biology and bioinformatics Introduction Modern drug development increasingly relies on computational approaches, from virtual screening to complex machine learning models. AI holds the potential to revolutionize clinical trials by accelerating medical breakthroughs and optimizing drug development pathways1. In this paradigm, clinical registries are transformed from reporting tools to a critical component of the data-driven drug discovery infrastructure. Trial data are used as downstream layers for in silico prediction validation and translational risk assessment. Therefore, the quality and semantic consistency of this metadata directly determine the reliability of computational solutions. In modern AI-driven drug discovery, clinical registries function as a downstream validation layer linking molecular predictions with real-world translational outcomes. Errors at this level propagate upstream, distorting model training and evaluation. Ideally, clinical registry data could support a wide range of machine learning approaches, including survival modeling, trajectory prediction, translational outcome classification, and unsupervised pathway clustering 2. Despite this, most efforts in the field are focused on optimizing algorithms, while the structure of the input data is often ignored 3. Structural degradation of clinical metadata generates algorithmic biases. Non-standardized clinical phase records generate false signals for machine learning, leading to a distorted view of translational progress 4. The necessity for such structural discipline is further reinforced by emerging methodological frameworks for trustworthy medical AI, particularly the METRIC. Rather than focusing solely on software implementation, the METRIC framework conceptualizes data integrity, consistency, completeness, traceability, and semantic coherence as foundational methodological requirements for reliable AI deployment in medicine. This problem is particularly acute in neuro-oncology. Glioblastoma shows a persistent discrepancy between the volume of research and actual clinical progress 5,6. Traditionally, this has been explained by the biological heterogeneity of tumors or the blood–brain barrier, but the stagnation of survival rates also points to nonbiological limitations, particularly in the processes of evidence generation and interpretation. Registries such as the WHO ICTRP use phase classification (I–IV) as a surrogate for trial maturity. However, this linear model does not always reflect the modern ecosystem: accelerated drug repurposing cycles, observational studies, and hybrid protocols often go beyond the boundaries of classical categories. In such circumstances, the question arises: is apparent translational stagnation a purely biological factor, or is it partly due to “translational blindness” (“out-of-distribution learning”) – the limited computational accessibility of clinical records. Biomedical ontologies were originally developed to support semantic interoperability, controlled vocabulary standardization, and formal representation of biomedical knowledge structures7. However, ontology-driven normalization alone does not guarantee preservation of methodological context, temporal continuity, or higher-order translational relationships within real-world clinical registry ecosystems. This paper systematically analyses over 2300 glioma trial registrations from the WHO ICTRP database. Using semantic normalization techniques, a computational pipeline (Python) is developed that transforms unstructured descriptions into machine-readable categories. This provides a preprocessing framework for integrating registry data into artificial intelligence systems, allowing for a quantitative assessment of the completeness and translational maturity of the current neuro-oncology landscape. Results Analysis of Table 1 demonstrates a pronounced inconsistency in phase reporting across global clinical trial registries, including datasets aggregated by the World Health Organization. Importantly, this inconsistency does not merely reflect missing values but reveals a systemic degradation of metadata structure, which substantially limits the interpretability of translational progress in glioma and glioblastoma research. Table 1Types of Phase Reporting Inconsistency.PhenomenonExample from dataProblem/ Loss of computational interoperabilityOver-specificationHuman pharmacology (Phase I): Yes; Therapeutic exploratory (Phase II): No…Redundancy that blocks automatic processingStructural InconsistencyII (Phase II study) vs 2 vs Phase-2Lack of a single linguistic standardMethodological AmbiguityNot selected, Not applicable, Retrospective studyInability to classify according to the standard FDA scaleTechnical VoidNaN (550 records)Complete absence of a translational marker Types of Phase Reporting Inconsistency. The most critical finding is the identification of 550 records lacking any phase information (NaN). Within the context of neuro-oncology, such records represent studies whose existence is documented, yet whose translational status cannot be inferred. As a result, these trials remain effectively invisible in phase-based analyses, creating a distorted representation of the research landscape. This absence of annotation introduces a systematic blind spot, where active research efforts cannot be positioned along the translational continuum. An opposing but equally problematic pattern is excessive narrative reporting of study phases. Entries such as “Human pharmacology (Phase I): Yes; Therapeutic exploratory (Phase II): No…” replace standardized phase labels with free-text descriptions. While informative at the individual record level, such formatting is incompatible with automated data aggregation and large-scale analytics. Consequently, over-specification transforms potentially structured metadata into computational noise, increasing the dependency on manual curation and reducing the scalability of translational monitoring. Substantial variability was observed in the linguistic representation of identical trial phases, including the concurrent use of Roman numerals (e.g., II), Arabic numerals (2), and hybrid textual forms (Phase-2). This heterogeneity indicates the absence of a harmonized reporting standard and prevents reliable phase-based querying. Within a single registry, identical developmental stages may be encoded in multiple non-equivalent formats, thereby fragmenting analytically identical data across artificial categories. Entries labeled as “Not selected” or “Not applicable” highlight a deeper methodological limitation. The classical FDA phase model (I–IV), originally designed for commercial drug development, does not adequately capture the full spectrum of biomedical research8, including observational, retrospective, or registry-based studies common in neuro-oncology. In such cases, investigators are compelled to select neutral or undefined categories, which effectively exclude these studies from phase-oriented translational analyses. Analysis of Table 2 points to another aspect of the translational barrier, and differs from the information in Table 1 where the phase annotation is characterized by significant missing data and structural degradation. The Study_type field exhibits the opposite phenomenon: high completeness, but at the same time accompanied by excessive terminological variability. This contrast provides critical insight into the nature of the observed data inconsistency. Table 2Types of Study_type reporting inconsistency.Record type (Raw Data)QuantityUnification ProblemInterventional (total variants)2065Semantic fragmentation: 6 + variants (Interventional, Interventional study, interventional, etc.)Observational (total variants)242Over-specification: Division into invasive/non-invasive without a unified standardMedicinal product specific227Classification overlap: Merging study type with the object of research (medicinal product)Specialized/Niche terms23Terminology drift: Use of BA/BE, Expanded Access, or Diagnostic test instead of main typesTechnical Voids (NaN/Empty)0Complete coverage: Unlike the “Phase” field, “Study_type” is a mandatory field with 100% filling rate Types of Study_type reporting inconsistency. Unlike the Phase field, which shows deep structural degradation (550 missing values), the Study_type field demonstrates complete administrative integrity. This stark difference indicates that investigators consistently identify and report the general methodological nature of their work (e.g., interventional or observational). Therefore, the loss of translational interpretability does not originate from incomplete reporting, but rather from the inability to map methodological intent onto a rigid phase-based classification system. In other words, researchers know what type of study they conduct, but lack a coherent framework to express where that study resides within the drug development continuum. Interventional studies represent the dominant category (2065 records); however, this group is fragmented into multiple orthographic and semantic variants (e.g., Interventional, Interventional study, interventional). Such variability suggests the absence of controlled vocabularies or automated input normalization within clinical trial registries. As a result, a conceptually uniform methodological category is artificially subdivided into multiple labels, complicating aggregation and large-scale quantitative analysis. Within the observational group (242 records), investigators frequently introduce additional qualifiers such as invasive, non-invasive, or patient registry. This pattern reflects an attempt to convey biologically and clinically relevant distinctions that are not accommodated by existing registry standards. While scientifically meaningful, these unregulated extensions generate heterogeneous metadata that resist harmonization and cross-study comparison at the global level. A distinct subset of records (n = 227) classifies studies according to the regulatory or material focus (e.g., medicinal product–specific) rather than study design. This conflation of methodological category with the object of investigation introduces an additional semantic layer that obscures the underlying experimental structure. Consequently, methodological classification becomes entangled with administrative or regulatory descriptors, further degrading data interpretability. Other specialized categories, including diagnostic tests and bioequivalence studies, account for the remaining 23 records. Note that there were entries that fell into several categories at once. The combined analysis of Tables 1 and 2 identifies a central anomaly in the translational data structure of glioblastoma research. While study authors reliably report the methodological nature of their investigations (Study_type where 100% of records are filled), they frequently fail to provide—or cannot meaningfully assign—a corresponding developmental phase (Phase where a significant portion remains empty or non-applicable). This asymmetry supports the hypothesis that inconsistencies in registry data arise not from technical oversight, but from a structural mismatch between academic research practices and the phase-based framework derived from pharmaceutical development. Note that before developing the algorithm, the entire dataset (n = 2357) was manually annotated by the Study_type and Phase columns to create a “gold standard” (Ground Truth). The algorithm was refined to achieve full concordance with expert annotations within the audited dataset with expert labels. This ensures that the developed Python pipeline accurately reproduces the logic of semantic validation, eliminating subjectivity when further scaling the analysis to other ICTRP datasets. This allowed the developed Python pipeline to be used as a reliable tool for automated auditing of large metadata datasets without the risk of losing semantic context. Aligning the registered study phases with the study design types in the four translational quadrants provides a structured visualization of the translational landscape in glioma research (Fig. 1). The definition of quadrants has a deterministic conceptual nature determined by the fundamental clinical methodology: the type of study (Interventional vs Observational). A total of 2357 clinical records were mapped into the two-dimensional phase design space, revealing a high degree of asymmetry in the distribution across translational domains. Each data point represents an individual study and is color-coded according to its registered phase of development. Translational phase-space mapping of glioma studies. The resulting phase space mapping delineates two fundamentally different zones. The first corresponds to a dense interventional domain dominated by phase-structured clinical trials, while the second includes sparsely populated observational, registry, and expanded-access studies that are poorly integrated into the traditional phase hierarchy. This spatial segregation highlights the pronounced translational discontinuity between experimental development and the generation of real-world clinical evidence. The intervention quadrant (Q1) corresponds to the vast majority of entries (n = 2077), highlighting the dominance of experimentally driven therapeutic trials in current glioma research. The distribution in this quadrant has a clear structure, indicating a high level of regulatory and methodological standardization. The largest group is made up of studies in phase 2 (n = 674). This is followed by projects in phase 1 (n = 559) and combined phase 1/2 trials (n = 303). It is precisely the concentration of activity at the phase 2 stage that is a critical “bottleneck” of translation. Despite intensive work in the early and middle stages, only a few (n = 136) proceed to the confirmatory phase 3. This indicates serious barriers between obtaining the first positive results and their large-scale clinical validation. The popularity of mixed protocols (phase 1/2) is explained by the specifics of neuro-oncology: in a high-risk environment, developers are forced to simultaneously verify both safety and first signs of efficacy in order to make decisions more quickly on the feasibility of further investments. At the same time, studies that reached phase 4 are minimally represented (n = 21), indicating a lack of systematic post-marketing surveillance. Only a small proportion of Q1 records contain missing information about the phase in which the studies are located (< 4%), a pattern that is best interpreted as a consequence of strict regulatory and administrative reporting requirements rather than as a marker of translational immaturity. While this structural rigor enhances internal validity, it also limits the representation of clinical heterogeneity and long-term outcomes, thereby reinforcing a cyclical focus on early-phase studies. In contrast, the surveillance quadrant (Q2; n = 249) exhibits a markedly different phase structure. Here, the dominant categories are records classified as “Blank” (technical void; n = 154) or characterized by methodological ambiguity (n = 69). The lack of clearly defined study phases suggests that the classical phase model (0–4) is poorly suited—or intentionally not applied—to observational and non-interventional study designs. Importantly, this pattern does not reflect a lack of data generation, but rather a structural mismatch between observational methodologies and phase-based study classification systems. The few Phase 3 study records observed in Q2 likely correspond to postmarketing surveillance or long-term cohort follow-up of patients receiving established standard treatments. Patient registry studies (Q3) and expanded access programs (Q4) are minimally represented, with only 16 and 15 records, respectively. Almost all studies in these quadrants lack structured phase metadata, with most records being empty classifications. This near-total absence of phase designation is methodologically justified, as registry-based and expanded access studies are not designed to test therapeutic interventions within the usual development hierarchy. However, their exclusion from phase-oriented metrics renders these data streams largely invisible in assessments of translational progress. Taken together, the observed phase-space distribution demonstrates that the current translational architecture in glioma research is largely optimized for the design of interventional trials but is ill-equipped to integrate observational and real-world data. The resulting imbalance highlights a systemic translational gap, where a large body of clinically relevant data accumulates outside the dominant phase-oriented paradigm, limiting their contribution to longitudinal outcome assessment and evidence-based drug development. Further analysis of the structural distribution of phases (Fig. 2) allowed us to quantitatively confirm that the multiphase architecture of clinical trials is a unique feature of the interventional quadrant Q1. Unlike other sectors, only Q1 demonstrates a continuous phase continuum from early to late stages, which creates the formal illusion of a completed translational chain. However, 12.1% of the trials could not be qualified as a research phase by the researchers and were identified as N/A. At the same time, the normalized distribution of phases reveals an internal disproportion, in which most of the activity is concentrated in phases focused on the initial assessment of safety and efficacy signals, while the final stages of clinical validation remain marginal. Conceptual framework of training-space compression and distribution shift across the translational continuum. Quantitative comparison of phase profiles between quadrants demonstrates that observational, registry, and expanded access programs do not form an alternative translational trajectory, but exist as isolated information layers. In these sectors, there is no even partial imitation of the phase hierarchy, which may indicate not the low quality of the data, but their fundamentally different epistemological nature. Thus, the accumulation of clinical observations, long-term outcomes, and post-intervention patient trajectories occurs outside the dominant analytical paradigm. Comparative analysis revealed that in the Q2 quadrant, 3.6% of studies reach phase 4, which is 3.6 times higher than the similar indicator in Q1 (1.0%). This unevenness indicates a deficit of translational patency in Q1, confirming the hypothesis of a violation of the integrity of the pathway even under operational control. The structural decomposition of the Phase 4 landscape (Table 3) further demonstrates that the formal presence of a late phase is not equivalent to translational maturity. The predominance of studies aimed at multiple optimization of the standard of temozolomide therapy indicates extensive use of already exhausted pharmacological solutions. At the same time, a significant proportion of technology-oriented studies indicates a gradual replacement of biological innovations with engineering, service or surgical interventions that do not create new pharmacodynamic trajectories. Table 3Structural Decomposition of Phase 4 Landscape in Glioma Research (n = 30).CategoryProportionTypical examples from the datasetFunctional implication for translational continuityOptimization of the Standard (TMZ-centric) ~ 40%TMZ injection, TMZ resistance, TMZ + Anlotinib, TMZ + ApatinibExhaustive iteration of an established pharmacological core. No emergence of new biological mechanismsTechnological / Engineering Substitution ~ 30%Handheld intraoperative microscope, GammaTile, NovoTTF, Telehealth vs In-personReplacement of pharmacodynamic innovation by engineering, procedural, or service solutionsImmunotherapy Dead End ~ 20%Pembrolizumab, Anti-PD-1 neoadjuvant/adjuvant, gp96-based immunotherapyLate-phase labeling without consolidation into a stable clinical standard; lack of Q3 continuityConcomitant / Supportive Research ~ 10%Cognitive function in elderly, immune system monitoring, quality-of-life studiesClinically relevant but non-transformative; does not alter survival or disease trajectory Structural Decomposition of Phase 4 Landscape in Glioma Research (n = 30). Analysis of all identified Phase 4 records since 1995 (n = 30) suggests that formal late-phase status is not equivalent to translational maturity. Despite the availability of alkylating agents with proven ability to cross the blood–brain barrier (BBB)—such as nitrosoureas (lomustine, carmustine, fotemustine) 9—none of these compounds are represented in the current Phase 4 registries for glioblastoma. Instead, the Phase 4 space exhibits an “inertial dynamic,” dominated by iterative modifications of the temozolomide (TMZ)-based standard and an increasing shift toward engineering and service interventions (Table 3). This structure further reveals a profound crisis in translational interpretation systems, as the four-phase classification of clinical trials, originally developed for regulatory progress, introduces significant categorical distortion when trying to use it in large-scale machine-readable datasets. For machine learning algorithms, the “Phase 4” label acts as a false predictive signal of success, hiding real bottlenecks in translation due to excessive semantic heterogeneity10. Within a single phase code, the machine simultaneously processes post-marketing surveillance, drug repurposing, engineering interventions, and quality-of-life assessments, which effectively represents five diametrically different types of translation. Since phase labeling is unable to encode the level of pharmacological novelty or the complexity of the biological mechanism, the effect of a pseudo-translational illusion arises. In this state, truly novel biological approaches such as immunotherapy (20% of the landscape) are lumped together with exhausted iterations of older drugs, artificially inflating the apparent activity of late-stage research against a backdrop of de facto biological stagnation. The calculation of the Reporting Gap () and Maturity Ratio () indices revealed critical asymmetries between study sectors: Intervention sector (Q1): Demonstrated the highest phase occupancy ( or ), but a low maturity ratio (). This indicates a high regulatory discipline during registration, but at the same time a significant accumulation (“stagnation”) of studies in early stages without further progress to the final phases. Statistical significance (Wilcoxon, p < 0.01) confirms that the high occupancy in Q1 is not accidental, while the lack of significance in Q3/Q4 highlights the critical ‘Reporting Gap’. Intervention sector (Q1): Demonstrated the highest phase occupancy ( or ), but a low maturity ratio (). This indicates a high regulatory discipline during registration, but at the same time a significant accumulation (“stagnation”) of studies in early stages without further progress to the final phases. Statistical significance (Wilcoxon, p < 0.01) confirms that the high occupancy in Q1 is not accidental, while the lack of significance in Q3/Q4 highlights the critical ‘Reporting Gap’. Observational sector (Q2): Characterized by a deep reporting gap (only 9.64% of records have a defined phase). However, the high among valid records indicates that observational studies are more often focused on post-marketing phases, although this activity remains almost “invisible” to standard monitoring systems. the differences between early and late stages are highly significant (p = 0.0078 for Q2). Confidence intervals for the main indicators of metadata degradation (95% CI) confirm that the structural anomalies detected are not random deviations (p < 0.05). Observational sector (Q2): Characterized by a deep reporting gap (only 9.64% of records have a defined phase). However, the high among valid records indicates that observational studies are more often focused on post-marketing phases, although this activity remains almost “invisible” to standard monitoring systems. the differences between early and late stages are highly significant (p = 0.0078 for Q2). Confidence intervals for the main indicators of metadata degradation (95% CI) confirm that the structural anomalies detected are not random deviations (p < 0.05). Zero values of phase metadata completeness in Q3 and Q4 mathematically confirm the systematic exclusion of long-term clinical outcomes and access programs from current translational analysis models. As a result, a gap is formed between the actual clinical experience of patients and the analytical tools used to evaluate drug development. This gap is not a result of a lack of data, but rather reflects the limitations of existing registration and classification systems. Discussion The object of this study is not biological translation itself, but its digital registry representation, which increasingly serves as the primary data substrate for AI-driven translational assessment. While clinical trial registries were originally designed for regulatory reporting and administrative harmonization, they are now frequently repurposed as machine-readable resources for computational modeling, translational forecasting, and automated evidence synthesis. As emphasized in recent comprehensive reviews, artificial intelligence offers substantial opportunities for optimizing clinical trial design, patient stratification, and translational decision-making; however, the successful implementation of such approaches remains critically dependent on data quality, structural consistency, and metadata integrity 11. Within this context, our findings suggest that current registry architectures were not designed to encode the temporal, biological, and contextual complexity required for modern machine learning systems.Despite the existence of established clinical data standards such as Clinical Data Interchange Standards Consortium SDTM12 , intended to improve interoperability and machine readability, we identified a pronounced gap between formal standardization frameworks and their practical implementation within glioma registries. The observed semantic fragmentation, inconsistent phase annotation, descriptive reporting practices, and large proportion of incomplete metadata collectively reduce the structural coherence of registry datasets. In this context, metadata integrity should not be viewed merely as a reporting-quality issue, but rather as a computational constraint directly affecting AI system performance. Medical concepts in clinical registries are frequently represented through heterogeneous terminology, spelling variation, abbreviations, and narrative descriptions that cannot be reliably resolved using deterministic string-matching approaches alone. Furthermore, standard biomedical ontologies such as Unified Medical Language System often fail to capture operational or non-standard terminology used in real-world registry entries13, generating extensive out-of-vocabulary (OOV) regions within the metadata space. As an international aggregation platform, International Clinical Trials Registry Platform therefore faces substantial challenges related to semantic harmonization and data curation before computational analysis becomes feasible. ICTRP14 as an aggregator faces critical challenges regarding data integrity, requiring the development of complex frameworks for its curation and filtering before analysis. For example, automated extraction of Phase II studies may systematically exclude clinically relevant trials because of inconsistent notation (“Phase II”, “Phase-2”, “II”) or narrative phase descriptions embedded within free text. Consequently, biologically meaningful studies may become effectively invisible to automated computational pipelines, introducing hidden sampling bias into downstream machine learning workflows. These findings indicate that observed translational bottleneck is not solely a biological or therapeutic problem, but also reflects a structural mismatch between fragmented registry metadata, clinical study design, and computational decision-making systems. Importantly, the limitations identified in this study originate at the earliest stage of registry submission15. Although the Study Type field demonstrated relatively high completeness, it exhibited substantial terminological variability that fragmented single methodological categories into multiple semantically inconsistent representations. Simultaneously, the large proportion of technical voids, including missing and empty phase annotations, suggests that the classical FDA phase model (I–IV) may be insufficient to capture the full diversity of contemporary neuro-oncology research. In many cases, the loss of translational transparency appears to arise not from technical error, but from a systemic disconnect between academic research practices and rigid pharmaceutical classification frameworks. Within this context, our audit demonstrates that raw glioma registry metadata frequently fail to satisfy several core dimensions of trustworthy data representation — including consistency, completeness, and explicability — due to semantic fragmentation, inconsistent annotation, and structurally incomplete records. Consequently, without deterministic normalization and rigorous metadata harmonization, registry-derived datasets may remain insufficiently robust for high-stakes clinical prediction, longitudinal modeling, and explainable AI applications in translational neuro-oncology16. This limitation extends beyond ontology mapping. While systems such as the Unified Medical Language System (UMLS)17 can effectively decompose registry text into atomic biomedical entities, their practical application in clinical informatics often remains focused on concept recognition rather than preservation of higher-order structural relationships. As highlighted in recent large-scale reviews of UMLS utilization, the framework performs well in terminology integration but remains less effective at maintaining methodological and contextual coherence between identified entities. Consequently, clinically important information — including the relationships between study phase, intervention design, and temporal progression — may be lost during computational preprocessing even when individual biomedical concepts are successfully recognized. These findings suggest that semantic normalization alone is insufficient for trustworthy translational modeling and that metadata curation frameworks must additionally preserve structural and temporal coherence within registry records. A central observation of our analysis is the absence of a coherent temporal structure within registry metadata. Clinical phases are commonly interpreted as indicators of translational progression over time; however, our findings demonstrate that phase annotations do not form a continuous, standardized, or comparable temporal axis. Trials are not linked into coherent developmental trajectories, and phase labels are frequently fragmented, inconsistently defined, or entirely absent. As a consequence, essential components required for longitudinal modeling — including valid event structure, temporal ordering, and censoring mechanisms — cannot be reliably reconstructed from registry data alone.This structural limitation has important implications for advanced machine learning methodologies. Survival modeling frameworks, including deep-learning approaches such as DeepHit, require temporally resolved trajectories and explicitly defined event structures. However, current registry systems do not encode translational processes as continuous time-resolved sequences, but rather as isolated categorical records. Consequently, the feasibility, interpretability, and robustness of longitudinal AI models applied to such datasets become substantially constrained. From the perspective of computational drug discovery, these findings also highlight a broader generalization problem. Early-phase studies dominate the translational landscape and frequently provide formally high-quality experimental data under tightly controlled conditions. However, these datasets often exclude the temporal, immune, metabolic, and age-related interactions that shape real long-term clinical response in glioblastoma treatment. The observed discontinuity between Phase II and later-phase studies18 therefore resembles a classical generalization failure, in which models perform adequately within a restricted experimental domain but fail to extrapolate into complex clinical environments. The conceptual framework presented in Fig. 3 helps explain the mechanism underlying this gap. Experimental controllability is commonly achieved through compression of biological complexity, resulting in the elimination of interacting dimensions that later become clinically decisive. Consequently, efficacy signals identified in early experimental systems are not necessarily false, but rather confined to a biological space that only partially overlaps with the multidimensional space of long-term therapeutic response. This directly limits the ability of machine learning systems to generalize beyond narrowly constrained training domains. Structural barriers to AI generalization: feature space compression and label space distortion. An additional limitation emerges from the structural organization of clinical registries themselves. The “Technical Void” identified in both registry and observational quadrants indicates that some of the most clinically informative data — particularly long-term patient trajectories and real-world outcomes — remain poorly represented in machine-readable form 19. In the absence of consistent phase annotations or alternative structural descriptors, these datasets cannot be effectively integrated into closed-loop computational pipelines for model training, validation, and iterative correction. As a result, AI systems theoretically intended to learn from real clinical feedback remain partially isolated from it. One of the most notable findings of this audit is the extremely low proportion of Phase IV studies (n = 30), representing only 1.27% of the total 2357 registry records. This quantitative depletion of the translational pipeline provides empirical support for the concept of limited translational progression. Moreover, structural analysis demonstrated that this segment is additionally affected by semantic heterogeneity, where genuine pharmacological innovation is frequently replaced by service-oriented, engineering, or optimization-focused interventions. Although some Phase IV activity may remain obscured within technically incomplete registry entries, the current registry architecture creates a form of “pseudotranslational progression,” in which formal phase advancement does not necessarily correspond to therapeutic novelty or meaningful long-term therapeutic transformation. These observations are consistent with broader findings in translational bioinformatics and clinical AI research, where failures of computational systems increasingly arise not from insufficient algorithmic sophistication, but from structurally constrained data representations that exclude clinically relevant dimensions such as temporal evolution, comorbidity structure, and system-level biological interactions. In this context, the limitations of AI-driven translational prediction may reflect deficiencies in data representation architecture rather than deficiencies in machine learning algorithms themselves. The reliability of large-scale clinical registry ecosystems is receiving increasing attention in the context of trustworthy AI and computational reproducibility20. The metrics proposed in this study — Reporting Gap () and Maturity Ratio () — therefore have not only descriptive value but also methodological significance. Rather than evaluating isolated therapeutic candidates, these metrics allow quantitative assessment of the structural coherence and translational maturity of the research ecosystem itself. Importantly, elevated values within observational and technically incomplete quadrants should not automatically be interpreted as indicators of poor-quality research. Instead, they reveal a structural incompatibility between rigid phase-based ontologies and the epistemological nature of real-world and observational clinical investigation. Beyond the structural fragmentation identified in our study, recent evidence has demonstrated substantial discrepancies between trial registry records and corresponding peer-reviewed publications in AI-driven clinical research 21. Similar limitations have been reported in the broader context of artificial intelligence applied to clinical trials, where algorithmic performance is constrained not by model architecture but by fragmented registries, rigid phase ontologies, and the lack of feedback from real-world clinical trajectories 22. Finally, an important issue requiring critical consideration is the growing use of synthetic or AI-generated data to compensate for missing clinical information. Although such approaches may reproduce statistical properties of existing datasets, they also introduce substantial risks, including amplification of pre-existing biases, suppression of rare but clinically meaningful conditions, and generation of statistically plausible yet biologically implausible associations. In structurally fragmented registry environments, synthetic augmentation may therefore conceal rather than resolve underlying representational deficiencies. Taken together, our findings suggest that current clinical registries remain optimized primarily for regulatory documentation rather than for representation learning or translational systems modeling. Future registry architectures may require trajectory-aware, semantically harmonized, and temporally structured frameworks capable of representing translational processes as continuous dynamic systems rather than isolated categorical events. Without such restructuring, the growing integration of AI into translational medicine will remain fundamentally constrained by the representational limitations of the underlying clinical metadata ecosystem23. Conclusions Overall, the results of this work indicate that overcoming the translational “valley of death” in neuro-oncology requires not only new molecules or algorithms, but also a review of the very structure of the data on which these algorithms are trained. We deliberately restricted the analysis to two fundamental descriptors to demonstrate that structural failure is detectable even at the most elementary metadata level. The analysis by only two basic parameters — phase and type of study — proved that to overcome the translational gap, it is necessary to radically change the approach to filling registries: from simple data fixation to total unification of all parameters at the input stage. This should include the introduction of new classification descriptors — for example, engineering solutions, surgical devices, and complex treatment combinations. Such an approach will allow transforming disparate records into validated datasets suitable for full-fledged use in machine learning and predictive analytics. However, the lack of temporal continuity, standardized phase annotation, and coherent linkage between trials hinders the application of these methods in their current form. Taken together, the results demonstrate that the translational ecosystem of glioma research is characterized by high formal activity but low structural conversion of knowledge into clinical maturity. The identified “blind spots” between interventional, observational, and registry sectors create systemic barriers to integrating real clinical trajectories into the computationally driven drug discovery process. Therefore, only a combination of biologically representative models, standardized metadata, and phase-independent computational metrics will ensure the formation of scalable in silico models. This will allow not only to predict clinical efficacy in controlled conditions, but also to finally integrate data from real therapeutic environments into the information space. Methods To achieve the aim of the study, an analysis of the dataset obtained from the International Clinical Trials Registry Platform (ICTRP) of the World Health Organization (https://trialsearch.who.int/) was conducted using the search term “glioma”. The raw dataset is permanently archived on Zenodo24. Descriptive statistical analysis was used to interpret the distribution of the obtained indices, which allowed to systematize the quantitative indicators of the sample and identify the main trends in the distribution of data (https://github.com/IrynaOliynyk/glioma-metadata-audit). Data preprocessing and validation of translational phase metadata To ensure the integrity and reproducibility of the translational phase analysis, a structured data validation and preprocessing pipeline was implemented for the Phase metadata field. Data source and computational environment The dataset was extracted from a structured clinical trial registry export (Excel format, .xlsx). All analyses were performed using Python (version ≥ 3.9) with the pandas library (version ≥ 1.5), executed in a local development environment (PyCharm IDE, Windows OS). No external APIs or proprietary tools were used. Data ingestion and preservation of semantic markers Data ingestion was performed using the pandas.read_excel() function with default null-value interpretation explicitly disabled (keep_default_na = False). This step was critical to prevent the automatic conversion of semantically meaningful entries—such as “N/A”, “Not Applicable”, or “NA”—into missing values (NaN), a known source of systematic information loss in large-scale registry analyses. This configuration ensured that: Explicit non-applicability markers were preserved as literal strings. Explicit non-applicability markers were preserved as literal strings. Only truly unfilled cells were treated as missing data. Only truly unfilled cells were treated as missing data. Algorithm for phase metadata validation A deterministic, multi-stage validation algorithm was applied to the Phase field: 1. Empty-cell detection (Technical Void identification) Cells were classified as Technical Voids if they met either of the following criteria: Native missing values (NaN); Native missing values (NaN); Cells containing only whitespace characters after string normalization. Cells containing only whitespace characters after string normalization. 2. Literal string integrity analysis. All non-empty entries were processed using exact string matching (no stemming or semantic unification) to enumerate unique phase descriptors. This allowed for the detection of: Over-specified narrative entries (e.g., multi-line phase descriptions); Over-specified narrative entries (e.g., multi-line phase descriptions); Structural variants of equivalent phases (e.g., “Phase II”, “II”, “2”, “Phase-2”). Structural variants of equivalent phases (e.g., “Phase II”, “II”, “2”, “Phase-2”). 3. Separation of absence vs. Ambiguity. A strict conceptual distinction was enforced between: Technical Voids — complete absence of phase metadata; Technical Voids — complete absence of phase metadata; Methodological Ambiguity — explicit declarations of non-applicability (e.g., “N/A”, “Not selected”). Methodological Ambiguity — explicit declarations of non-applicability (e.g., “N/A”, “Not selected”). 4. Cumulative consistency verification. A cumulative frequency check was performed to ensure that the sum of all classified entries exactly matched the total number of records (N = 2357). This step served as a final integrity control, excluding silent data loss during preprocessing. Expert-guided semantic validation (gold standard): To ensure methodological rigor and mitigate the limitations of automated heuristic rules, we performed a 100% manual semantic validation of all unique values identified in the dataset. All distinct terminological variants extracted at step 2 were manually annotated by the author to construct a comprehensive gold standard reference. This exhaustive audit of the entire unique value space served as the benchmark for evaluating the performance of the computational pipeline. We note that this validation strategy is domain-specific and reflects the structure of glioma registry data; adaptation may be required for other registries or disease domains. Metadata classification framework The pipeline was designed as a deterministic audit framework rather than a predictive classification model; therefore, metrics such as ROC-AUC or classification accuracy were not applicable to the primary objectives of this study. For analytical purposes, all phase metadata entries were assigned to one of four mutually exclusive categories (Phase): Standardized — canonical FDA-style phase labels (Phase 0–4, including hybrid phases). Standardized — canonical FDA-style phase labels (Phase 0–4, including hybrid phases). Over-specified — narrative or multi-parameter phase descriptions exceeding standard nomenclature. Over-specified — narrative or multi-parameter phase descriptions exceeding standard nomenclature. Methodological Ambiguity — explicit non-labels indicating uncertainty or non-applicability. Methodological Ambiguity — explicit non-labels indicating uncertainty or non-applicability. Technical Voids — entirely missing or unfilled entries. Technical Voids — entirely missing or unfilled entries. A four-quadrant system based on the ICTRP regulatory descriptors was used to structure the glioma translational landscape (Study_type): Q1 (Interventional Sector): Clinical trials of new therapeutic agents (Interventional). This is the core of classical translation. Q1 (Interventional Sector): Clinical trials of new therapeutic agents (Interventional). This is the core of classical translation. Q2 (Observational Sector): Observational studies, analysis of real-world clinical practice and diagnostics (Observational). Q2 (Observational Sector): Observational studies, analysis of real-world clinical practice and diagnostics (Observational). Q3 (Registries/Epidemiology): Patient registries and large epidemiological datasets (Patient Registry). Q3 (Registries/Epidemiology): Patient registries and large epidemiological datasets (Patient Registry). Q4 (Access/Expanded Use): Expanded access and humanitarian use programs (Expanded Access). Q4 (Access/Expanded Use): Expanded access and humanitarian use programs (Expanded Access). Quantitative outcomes of validation The validation pipeline confirmed full dataset convergence (N = 2357). Comparison with the expert-led Gold Standard demonstrated a high pipeline accuracy of 100%, confirming that the deterministic regex-based approach effectively captures clinical intent while maintaining transparency. Of these: 11.24% (n = 265) constituted complete technical voids; 11.24% (n = 265) constituted complete technical voids; 13.62% (n = 321) used heterogeneous linguistic markers of non-applicability, which are systematically misinterpreted by default registry aggregation tools. 13.62% (n = 321) used heterogeneous linguistic markers of non-applicability, which are systematically misinterpreted by default registry aggregation tools. Phase metadata normalization and categorization To resolve structural inconsistencies in phase reporting, we implemented a hierarchical heuristic pipeline in Python. The algorithm followed a deterministic decision tree to transform heterogeneous registry entries into standardized analytical categories through four primary stages: Preprocessing and String Normalization: All entries underwent uniform lowercasing, whitespace collapse, and preservation of semantic non-applicability markers (e.g., “N/A”), ensuring that methodological intent was not lost to default null-value conversion. Preprocessing and String Normalization: All entries underwent uniform lowercasing, whitespace collapse, and preservation of semantic non-applicability markers (e.g., “N/A”), ensuring that methodological intent was not lost to default null-value conversion. Semantic Decomposition via Pattern Matching: To handle over-specified narrative entries, we employed regular expression (regex)-based filters. Regular expressions were used to identify numerical and Roman indices (1, I, 2, II) in the context of the word ‘phase’ or its abbreviations. These filters were designed to detect affirmative phase markers within multi-parameter descriptions, effectively decoupling the active developmental phase from administrative or descriptive noise. Semantic Decomposition via Pattern Matching: To handle over-specified narrative entries, we employed regular expression (regex)-based filters. Regular expressions were used to identify numerical and Roman indices (1, I, 2, II) in the context of the word ‘phase’ or its abbreviations. These filters were designed to detect affirmative phase markers within multi-parameter descriptions, effectively decoupling the active developmental phase from administrative or descriptive noise. Hybrid and Classical Phase Identification: The pipeline utilized specific logic to identify canonical FDA phases (I–IV) and hybrid oncology-specific designs (e.g., Phase 1/2). This ensured that overlapping safety/efficacy protocols were accurately captured rather than reduced to single-phase categories. Hybrid and Classical Phase Identification: The pipeline utilized specific logic to identify canonical FDA phases (I–IV) and hybrid oncology-specific designs (e.g., Phase 1/2). This ensured that overlapping safety/efficacy protocols were accurately captured rather than reduced to single-phase categories. Classification of Structural Noise: Entries that referenced “phase” but lacked resolvable numerical identifiers were systematically isolated as Over-specified Narrative records. This enabled a clear distinction between structured metadata and unstructured textual residues. Classification of Structural Noise: Entries that referenced “phase” but lacked resolvable numerical identifiers were systematically isolated as Over-specified Narrative records. This enabled a clear distinction between structured metadata and unstructured textual residues. To quantify "Out-of-distribution learning" we introduced two indices 25: Reporting Gap (Completeness Index): Percentage of records in each quadrant that have a defined (not empty and not ambivalent) phase. Reporting Gap (Completeness Index): Percentage of records in each quadrant that have a defined (not empty and not ambivalent) phase. Maturity Ratio: Ratio of the sum of late phases (Phase 3 + 4) to early phases (Phase 0 + 1 + 1/2), which allows us to assess the “exhaustion” of the translational flow. Maturity Ratio: Ratio of the sum of late phases (Phase 3 + 4) to early phases (Phase 0 + 1 + 1/2), which allows us to assess the “exhaustion” of the translational flow. To quantify the information transparency within each translational quadrant we introduced the Reporting Completeness Index (). where: — data completeness level for the quadrant q; — data completeness level for the quadrant q; — number of records containing a clearly defined phase (Phase 0–IV or hybrids 1/2, 2/3); — number of records containing a clearly defined phase (Phase 0–IV or hybrids 1/2, 2/3); — the total number of entries in a particular quadrant. — the total number of entries in a particular quadrant. The Reporting Gap was defined as the inverse of the Completeness Index: The indicator allows for a direct comparison of the level of “digital maturity” between fundamentally different types of studies (e.g., interventional Q1 and observational Q2). High indicator (Gap) indicates that the problem is not the lack of studies themselves, but rather a structural mismatch between the scientific design and the available registry fields. Unlike the standard count of empty cells (NaN), this index takes into account “methodological ambivalence” (N/A, Not applicable), which is de facto lost information for translational monitoring. At the same time, this indicator has limitations. The low in quadrants Q3 (Registries) and Q4 (Expanded Access) may be due to the lack of regulatory requirements to indicate phase, rather than negligence on the part of researchers. Since the indicator depends on the operation of Regex filters, any unaccounted linguistic variation in the “raw” data may slightly underestimate the index. However, our 100% manual validation of unique descriptors ensures that the impact of such variations on the final is minimal. The index only records the presence of a standardized label, but does not validate the correspondence of the specified phase to the real study design (i.e. we take the registrar at his word). To assess the dynamics of the advancement of therapeutic candidates along the translational chain within each quadrant, the () was introduced. It reflects the ability of the research type of study to move from early exploratory stages to large-scale confirmatory trials. The index smooths out the difference in the absolute number of studies, allowing us to assess the "efficiency of exit" from the early phases. A value of < 1 indicates a build-up of projects in the early stages, while > 1 may indicate the development of long-term observational programs. Note that the index is sensitive to small samples (e.g., in Q3 and Q4) and does not take into account the time required to transition between phases. The dataset comprised the complete set of 2357 glioma-related clinical trial records available in the WHO ICTRP registry at the time of extraction (2 January 2026). No sampling or record selection was performed; thus, the analysis represents a full historical registry snapshot used as a stable baseline for manual validation and computational processing. Non-parametric comparisons between early- and late-phase distributions across quadrants were evaluated using the Wilcoxon rank-sum test due to non-normal phase distributions and unequal group sizes. Acknowledgements The author acknowledges the support of the Federation of European Biochemical Societies (FEBS) through a Ukrainian Short-Term Fellowship. Special thanks are extended to Professor Mauro Maccarrone and the members of his laboratory at the University of L’Aquila for their hospitality and assistance. The author further acknowledges the Net4Brain network (COST Action CA22103) for providing a collaborative framework that contributed to the development of this study. Author contributions I. O. conceived the study, designed the methodology, performed the data analysis, interpreted the results, prepared the figures, and wrote and revised the manuscript. Data availability The datasets analyzed in this study are publicly available. Clinical trial metadata were retrieved from the WHO International Clinical Trials Registry Platform (ICTRP). Declaration Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets analyzed in this study are publicly available. Clinical trial metadata were retrieved from the WHO International Clinical Trials Registry Platform (ICTRP).