Informed consent forms (ICFs) are critical to ethical clinical research, safeguarding participant autonomy and trust. Nevertheless, concerns remain that excessive standardisation and poor communication within ICFs may impair participant understanding and contribute to trial failure. This study analysed 85 interventional clinical trials registered on ClinicalTrials.gov, comprising 34 terminated and 51 completed trials, to explore the relationship between ICF characteristics and trial outcomes. Publicly available ICFs were processed using natural language processing techniques, including Sentence-BERT embeddings to quantify linguistic similarity, alongside manual evaluation of ethical clause richness and structural consistency measured via edit distance. Statistical analysis employed Mann-Whitney U tests and logistic regression. The results revealed that trials which were terminated exhibited significantly higher internal linguistic similarity and lower ethical clause richness compared to those that completed successfully, whereas structural formatting differences were not predictive of trial status. Logistic regression confirmed that elevated textual uniformity and diminished ethical diversity within consent documents were associated with greater odds of trial termination. These findings suggest that over-standardisation and informational sparsity in ICFs may negatively affect participant engagement and trial viability. Ethics committees and trial sponsors are urged to foster participant-centred, context-specific consent documentation to support both ethical obligations and operational success.

Informed consent forms (ICFs) play a crucial ethical role in safeguarding participant autonomy and ensuring voluntary enrolment in clinical trials (Ssali et al., 2017). Despite the centrality of informed consent, persistent issues undermine its effectiveness, particularly in the context of complex biomedical research where linguistic complexity, over-standardisation, and poor communication are pervasive (Grant, 2021; Kadam, 2017). Multiple studies have noted that ICFs often exceed recommended readability levels, employ technical jargon, and prioritise legal protection over participant comprehension, ultimately challenging the core ethical mandate of informed consent (Licari and Manti, 2018; Rebers et al., 2016).

In developing and developed research environments alike, participants may formally consent without fully understanding the study procedures, risks, and benefits, thereby compromising the integrity of the process (Ssali et al., 2017; Bhupathi and Ravi, 2017). Notably, the focus on procedural adherence rather than communicative clarity has resulted in consent documents that may be signed without genuine informed engagement. Furthermore, while ethical guidelines such as the Declaration of Helsinki and the Belmont Report emphasise participant understanding as a cornerstone of research ethics, practical application remains inconsistent and often unverified (Rebers et al., 2016).

While existing literature has addressed the qualitative shortcomings of ICFs, there remains a critical gap: no comprehensive quantitative analysis has linked the structural and linguistic characteristics of ICFs to tangible clinical trial outcomes, such as trial termination. Previous efforts have described challenges related to participant literacy, cultural differences, and ethical oversight (Ssali et al., 2017; Bhupathi and Ravi, 2017; Fons-Martínez et al., 2022), but none has systematically assessed whether language structure in consent forms predicts the operational fate of clinical trials.

This study addresses this gap by applying advanced natural language processing (NLP) techniques to a curated set of ICFs from clinical trials that have either been completed or terminated. The research objectives are threefold: first, to assess textual similarity across trial outcomes using semantic embedding measures; second, to quantify ethical clause richness as an indicator of depth and breadth of ethical disclosures; and third, to test the predictive association between linguistic features of ICFs and trial termination through logistic regression modelling.

By systematically interrogating the linguistic fabric of consent documentation, this study advances the ethical discourse on informed consent from descriptive critique to quantitative evaluation, providing actionable insights for improving both ethical standards and operational trial outcomes.

Study Design and Data Source
A retrospective cross-sectional study design was employed. Data were extracted from ClinicalTrials.gov, a publicly accessible registry maintained by the United States National Library of Medicine. Eligible trials were identified based on predefined criteria, and full-text informed consent forms (ICFs) were sourced either directly from trial records or through linked external repositories. All documents were screened to ensure completeness and relevance to study objectives.

Trial Selection Criteria
Trials were included if they met the following criteria: (i) industry sponsorship, (ii) Phase II, III, or IV designation, (iii) recruitment status listed as "Terminated" or "Completed", and (iv) availability of a full-text ICF. Phase I studies, withdrawn or suspended trials, and trials lacking a complete ICF were excluded. The focus on later-phase, industry-sponsored trials ensured uniform regulatory context and maximised the relevance of informed consent content to mature clinical research practices.

Text Processing
Informed consent forms (ICFs) were extracted from PDF documents, followed by cleaning and text normalisation. Sentences were embedded using the Sentence-BERT model (all-MiniLM-L6-v2). Textual similarity between documents was computed using three metrics: cosine similarity (semantic closeness), Jaccard similarity (binary overlap of clause presence), and edit distance (token-level structural variation).

Ethical Clause Scoring
Ethical clause presence was assessed using predefined regular expression patterns (Appendix B), covering domains such as purpose, risks, confidentiality, and voluntary participation. Each clause was checked for presence either manually or using automated pattern matching to ensure detection of key ethical disclosures.

Statistical Analysis
Descriptive statistics were calculated for word counts, similarity scores, and ethical clause counts. Mann-Whitney U tests compared terminated and completed trials across linguistic features. Logistic regression models were constructed to predict trial termination, with model evaluation based on Akaike Information Criterion (AIC), McFadden’s R², and the Hosmer-Lemeshow goodness-of-fit test.

Software and Tools
Data extraction, text processing, and document similarity analyses were performed using Python (version 3.12.1). Statistical analyses, including Mann-Whitney U tests and logistic regression modelling, were conducted using R (version 4.3.2). Full scripts, software environment specifications, and reproducibility materials are available at the GitHub repository here.

Sample Characteristics
A total of 85 clinical trials were included in the final analysis, comprising 34 terminated trials and 51 completed trials. Trials were drawn from four major therapeutic areas: oncology, infectious diseases, cardiology, and other therapeutics. Oncology was the most represented area, comprising approximately 51% of the analysed sample. Trials were excluded if the informed consent forms (ICFs) were defective or scanned images unsuitable for text extraction. Table 1 summarises the distribution of the analysed trials by status and therapeutic area.

Linguistic Characteristics and Similarity Analysis
The linguistic characteristics of informed consent forms (ICFs) were compared between terminated and completed trials using mean cosine similarity scores, mean ethical clause counts, and mean readability scores.

Semantic Similarity (Cosine Similarity)
The mean cosine similarity for completed trials was higher (Mean = 0.607) than for terminated trials (Mean = 0.539), suggesting greater internal linguistic coherence among ICFs from successfully completed studies. Boxplots illustrated an upward shift in similarity distribution for completed trials relative to terminated trials. A Mann-Whitney U test confirmed that the difference was statistically significant (p = 0.0000173).

Ethical Clause Coverage
Completed trials exhibited a higher mean number of ethical clauses identified within their ICFs compared to terminated trials. Visualisations indicated greater clause inclusion frequency among completed trials, supporting the hypothesis that richer ethical disclosures may enhance trial success. This difference was statistically significant (p = 0.0000000598).

Readability Analysis
Preliminary readability analysis based on Flesch-Kincaid scores revealed no substantial difference between groups. Both terminated and completed trials exhibited readability levels above the recommended eighth-grade threshold, suggesting a general need for simplified language across all ICFs.

Regression Results
A logistic regression analysis was conducted to examine the association between document linguistic features and clinical trial outcomes (completed versus terminated), specifically to test Hypotheses 1–3. The predictors included group mean cosine similarity (H1: linguistic consistency), group mean Jaccard similarity (H2: ethical clause consistency), group mean edit distance (H3: structural consistency), and clause count (document richness).

The model demonstrated excellent overall fit, with a McFadden’s pseudo-R² of 0.484, indicating that approximately 48.4% of the variance in trial outcome was explained by the predictors. The Akaike Information Criterion (AIC) was 68.55, suggesting strong model parsimony. The Hosmer-Lemeshow goodness-of-fit test was non-significant (χ²(8) = 5.16, p = 0.741), indicating no evidence of model miscalibration.

As shown in Table 3, group mean cosine similarity was a significant positive predictor of trial completion, supporting Hypothesis 1 (H1). Trials with higher internal linguistic coherence were substantially more likely to be completed. Similarly, group mean Jaccard similarity emerged as a significant positive predictor (H2), indicating that greater consistency in ethical clause coverage increased the likelihood of trial success.

In contrast, group mean edit distance did not significantly predict trial outcome (p = 0.113), providing no support for Hypothesis 3 (H3). Finally, clause count was significantly negatively associated with trial completion. Higher document complexity, reflected in more clauses, was linked to a reduced probability of successful trial completion.

Predicted Probability of Trial Completion Across Key Document Similarity Metrics
Logistic regression model predictions were plotted to examine the relationship between key document similarity measures and the probability of clinical trial completion, corresponding to Hypotheses 1–3.

H1: Group Mean Cosine Similarity (Linguistic Consistency)
The predicted probability of trial completion increased sharply with rising group mean cosine similarity, following a sigmoidal logistic pattern with an inflection around 0.55–0.60.

Higher internal linguistic coherence across informed consent forms (ICFs) was associated with substantially greater trial completion probabilities, supporting Hypothesis 1.

H2: Group Mean Jaccard Similarity (Ethical Clause Consistency)
Similarly, the probability of trial completion rose markedly with increasing group mean Jaccard similarity, with the steepest increase observed between values of 0.70 and 0.80.

This pattern visually supports Hypothesis 2, indicating that greater ethical clause consistency is associated with improved trial success.

H3: Group Mean Edit Distance (Structural Consistency)
In contrast, group mean edit distance demonstrated only a modest, near-linear relationship with trial outcome, with minimal impact on completion probability.

This finding corroborates the earlier statistical result, providing no strong support for Hypothesis 3.

Odds Ratios and Confidence Intervals for Predicting Trial Completion
Odds ratios (ORs) and corresponding 95% confidence intervals (CIs) were computed from the logistic regression model to assess the strength and direction of association between document similarity metrics, clause count, and trial outcome (completed versus terminated).

Group mean cosine similarity demonstrated an extremely large odds ratio (OR ≈ 9.67 × 10¹?), with a confidence interval entirely above 1, confirming a strong and statistically significant association with trial completion (H1). Similarly, group mean Jaccard similarity showed a very large odds ratio (OR ≈ 1.16 × 10¹?), also supporting a significant positive effect (H2). In contrast, group mean edit distance, despite yielding a large odds ratio (OR ≈ 1.50 × 10²²), exhibited a wide and unstable confidence interval crossing 1, indicating no statistically reliable association with trial outcome (H3).

Principal Findings
This study demonstrated that internal linguistic consistency and ethical clause richness within informed consent forms (ICFs) are significantly associated with clinical trial success. Trials with higher semantic similarity and greater ethical disclosure were more likely to reach completion, while structural formatting consistency alone showed no predictive value. Additionally, greater document complexity, reflected in increased clause counts, was associated with reduced likelihood of trial completion.

Interpretation
These findings suggest that excessive standardisation and structural uniformity may undermine the communicative purpose of informed consent, potentially disengaging participants. High clause counts and rigid formatting do not guarantee better understanding and may instead contribute to cognitive overload. By contrast, meaningful ethical communication, emphasising clarity, semantic coherence, and accessible ethical disclosures appears critical for participant engagement and trial retention. Optimising ICFs for linguistic consistency and ethical clarity, rather than mere procedural standardisation, could thus enhance both ethical standards and operational outcomes in clinical research.

Comparison with Prior Work
Prior research has extensively examined issues surrounding informed consent, including readability, ethical transparency, and participant engagement. Table 5 compares major previous findings with the present study, highlighting key differences and methodological advancements.

Strengths and Limitations
Strengths
This study represents the first large-scale, systematic application of natural language processing (NLP) techniques to informed consent forms (ICFs), quantitatively linking linguistic characteristics to clinical trial outcomes. By combining advanced semantic embedding models (Sentence-BERT) with manual ethical clause detection, it offers a novel integration of machine learning and bioethical analysis, contributing uniquely to both clinical trial operations and research ethics literature. The use of a real-world, publicly available dataset enhances transparency and reproducibility. Furthermore, rigorous statistical methodologies, including Mann-Whitney U tests, logistic regression modelling, and model calibration checks substantially strengthen the internal validity of the findings.

Limitations
However, several limitations must be acknowledged. First, reliance on publicly available ICFs introduces potential selection bias, as not all trials disclose full consent documentation, and available forms may differ systematically from the broader trial population.

Second, the focus exclusively on English-language documents may limit the generalisability of findings to non-English-speaking contexts; cross-cultural variation in consent practices remains unexplored.

Third, site-specific modifications and verbal consent processes were not captured, which may affect linguistic consistency in practice.

Fourth, while statistically powered, the sample size remains modest relative to complex multivariate analysis, limiting the exploration of interaction effects across therapeutic areas or sponsor types.

Finally, the observational nature of the study precludes firm causal inferences; associations between ICF characteristics and trial outcomes must be interpreted cautiously.

Future Directions
Future research should expand datasets to include non-public and non-English ICFs, enabling broader cross-cultural validation. Prospective trials embedding semantic evaluation tools into the consent design process are recommended to enhance document quality systematically. Further, participant-centred metrics such as comprehension scores and satisfaction surveys should be incorporated to investigate mediating pathways between document characteristics and trial outcomes. Stratified analyses by therapeutic area are warranted, as differing disease contexts may moderate the relationship between consent quality and trial success. Lastly, examining the impact of culturally tailored ethical clauses on participant engagement and retention would provide valuable insights for global clinical research practices.

Implications for Practice
The findings highlight the need for sponsors and investigators to prioritise semantic clarity and ethical transparency in ICF development, beyond mere compliance with regulatory templates. Ethics committees should evaluate not only content completeness but also linguistic coherence. Regulators could consider guidance revisions to emphasise participant-centred readability, recognising its potential operational impact on trial success.

This study demonstrates that data-driven, language-based evaluation of trial documentation is both feasible and necessary to enhance clinical research quality and participant protection. Future improvements in informed consent practices will be critical for rebuilding participant trust, ensuring ethical engagement, and advancing the transparency and integrity of global clinical research.

Data and Code Availability
The datasets analysed during the current study and the full natural language processing (NLP) codebase, including text pre-processing scripts, Sentence-BERT model specifications, and similarity computation methods, are publicly available at here.

Access is unrestricted and provided to promote transparency, reproducibility, and further research development.

Appendix A: Full Trial Dataset
The full list of clinical trials included in the analysis, including NCT ID, Phase, status, and sponsor information, is available via the study's public GitHub repository: [here]

Appendix B: Ethical Clause Definitions and Detection Patterns
Ethical clauses were defined based on standard requirements for informed consent disclosures. Each clause was operationalised through targeted regular expressions (regex) to automate presence detection within the consent documents. Table B1 lists the clause categories and their corresponding search patterns.

Each ICF was scanned for the presence of these clauses using case-insensitive regex matching. Detection was considered binary (presence/absence) for each clause per document.