A Framework for Large Language Models (LLMs) in Clinical Trials: Trially and Margo as a Case Study in Application
Two recent studies, "Large Language Models in Clinical Trials: applications, technical advances, and future directions" and "LLM Auditing of Clinical Trial Reporting Quality" explain how using Large Language Models (LLMs) in clinical research solves critical operational bottlenecks while also improving the integrity of the research itself. These studies also provide a framework that can be used to benchmark new technologies serving the clinical research ecosystem. This makes them the perfect tool for evaluating the AI platform Trially and its agent, Margo, which are built to fix the exact inefficiencies in clinical trial design, patient recruitment, and data quality that the research highlights.
How LLMs Address Foundational Challenges in Clinical Trial Design and Protocol Rigor
The Large Language Models study highlights the complexity of trial design, noting the need to extract and synthesize essential components, such as PICO elements, to construct scientifically grounded protocols. This is further supported with research showing LLMs are critical for accelerating the clinical design phase by synthesizing metadata into protocol documents. Trially directly matches this need for rigorous, efficient information processing through its proprietary AI, which performs rapid protocol parsing of complex inclusion/exclusion (I/E) criteria into manageable criteria cards, often in under 5 to 10 minutes. This ability relies on the LLM’s technical advantage in contextual understanding, allowing it to capture semantic connections across complex text.
Implementing strong, precise planning early in a clinical trial—commonly referred to as operational rigor in the design phase—is necessary to combat the integrity problems identified by the LLM Auditing study. That study found that when researchers publish their results, incomplete reporting in randomized clinical trials (RCTs) obscures bias and limits reproducibility. A major example of this persistent failure is the low reporting rate for critical steps like the allocation-concealment mechanism (only 16.1% of audited trials reported this methodological detail). By deploying LLMs to optimize and clearly define the trial’s operational criteria upfront, Trially helps ensure the fundamental methodological integrity necessary for comprehensive and complete reporting later on. This LLM capability allows for rapid protocol parsing of complex criteria, and the system's ability to analyze criteria complexity and their association with trial termination risks further aligns with Trially’s planned feature, the Protocol Refiner, which is designed to optimize I/E criteria before study initiation, thereby mitigating initial risks that could lead to subsequent reporting difficulties or trial failure.
Overcoming the Patient Recruitment Bottleneck and Unstructured Data
The most prominent operational challenge facing clinical trials is inefficient participant recruitment, which results in up to 68% of sites failing to meet enrollment targets. This inefficiency is fundamentally linked to the reliance on manual screening, which can take 250+ hours per month per site. The Large Language Models study stresses that manual methods are hampered by the discrepancy between unstructured data in electronic health records (EHRs) and structured eligibility criteria.
Trially Match is engineered as a direct LLM response to this specific bottleneck. Trially’s LLM-powered platform is designed to synthesize 100% of EHR data (both structured and unstructured information) in real-time to generate clinical insights. This capability provides a key advantage over traditional NLP approaches that struggle with free-text EHRs. This LLM application achieves superior performance, demonstrated by its approximate 95% accuracy when screening patients against eligibility criteria and resulting in a reported 73% reduction in screen failure rates. This efficacy is empirically supported by findings showing LLM-assisted prescreening yields a higher eligibility rate (20.4%) compared to manual methods (12.7%).
LLMs in Patient Conversion and Communication during Trial Conduct
Beyond patient identification, the Large Language Models study notes that trial operations are challenged by lead conversion issues and the manual "Patient-Physician Handoff", where qualified candidates often "slip away due to insufficient follow-up". LLMs offer a critical technical advancement here: dynamic text generation.
Margo AI, Trially’s LLM-powered agent, leverages this technical advantage to solve the conversion gap. Margo is designed for outreach, automated prescreening, scheduling, and follow-up. The LLM enables Margo to coherently generate real-time patient qualification summaries, schedule appointments and reminders (via SMS and voice), and re-engage dormant patients. Margo’s success is rooted in the LLM's capability to adapt its narrative style based on context. For example, the LLM can use precise, objective terminology for data analysis but employ more accessible language for patient outreach, improving comprehensibility and conversion. This automation allows sites to multiply enrollment by 2–6X in complex studies, directly tackling the staffing burden and low enrollment rates mentioned in the operational challenge sources.
Navigating LLM Implementation Issues: Security, Privacy, and Trustworthiness
The integration of LLMs in highly regulated clinical settings faces significant implementation hurdles, including concerns over data quality, privacy, blinding control, and navigating legal and regulatory concerns such as HIPAA and GDPR. Furthermore, a fundamental technical limitation noted in the Operational Source is output hallucination, which undermines credibility.
Trially directly aligns with the solutions required to mitigate these risks by integrating a comprehensive suite of security and compliance protocols. The Trially platform is explicitly certified as HIPAA-compliant, SOC 2, FDA Part 11, and ISO 27001 compliant. To safeguard patient private health information (PHI) and control data quality, Trially employs AES-256 encryption for data at rest and TLS encryption for data in transit. Regarding blinding and access control—a key concern when deploying LLMs—Trially implements Role-Based Access Control (RBAC) and automated audit logging of all PHI access.
To counter the inherent "black box" lack of transparency and mitigate the risk of hallucinated misguidance, Trially ensures decision traceability. The platform allows users to triple click into I/E criteria to see the synthesis of physician notes and understand exactly why a patient qualifies. This transparency aligns with the emphasis on rationale and confidence levels utilized by the auditing LLM framework, where the model provided step-by-step reasoning for its CONSORT compliance decisions to enhance explainability and trust. By providing these robust technical and compliance solutions, Trially addresses the critical necessity for trustworthy and secure LLM deployment required for broader and safer integration into clinical trials.
