Mapping the LLM-Powered Clinical Workflow to the Clinical Trial Patient Workflow: An Analysis of Trially and Margo AI
The technological shift of using Large Language Models (LLMs) into clinical workflows demonstrates the potential of Artificial Intelligence (AI) to enhance both patient care and clinical research efficiency. The recently published study, “A comparison of the performance of Chinese large language models and ChatGPT throughout the entire clinical workflow”, evaluated several LLMs, including ChatGPT-4 and the Chinese Doubao and ERNIE Bot 3.5. These LLM models were tested across the entire clinical workflow, which includes differential diagnosis, diagnostic workup, diagnosis, and management. The results show that modern LLMs can deliver performance comparable to, and sometimes superior to, human emergency medicine fellow physicians. While this research focuses on direct patient care scenarios using simulated cases, the methodologies and successes align directly with the functionality of the Trially platform and its core conversion engine, Margo AI, particularly in accelerating the critical patient workflow necessary for clinical trial enrollment.
High-Precision Diagnosis and Patient Identification
The full clinical workflow requires LLMs to successfully navigate complex tasks, starting with differential diagnosis (diff) questions, which ask which conditions cannot be excluded from an initial assessment, and diagnostic questions (diag), which determine appropriate next steps. The LLM comparison study found that the performance of models like Doubao and ChatGPT-4 was similar across these dimensions, although differential diagnosis represented the area of poorest overall performance for the LLMs tested. This stage requires synthesizing information to narrow possibilities and form hypotheses.
Trially’s approach mirrors this data synthesis requirement, but instead of forming a general clinical differential, Trially Match uses its LLM-powered engine to synthesize 100% of structured and unstructured EHR data in real-time to match patients against complex trial protocols. This deep analysis provides the clarity needed to manage data overload, a key function noted for LLMs in healthcare workflows, enabling doctors to make "quicker and better decisions". Trially achieves high precision matching with approximately 95% accuracy against eligibility criteria. This capability serves as an automated, high-speed initial "diagnostic workup" in the clinical trial context, allowing sites to rapidly prioritize and stack rank ten times the number of candidates.
In the LLM clinical comparison study following the workup, the clinical workflow progresses to diagnosis questions (dx), where the LLMs excelled, achieving an average correct proportion above 97% for all three models (ChatGPT-4, Doubao, ERNIE Bot 3.5). In the research setting, Trially translates this diagnostic accuracy into clinical utility by ensuring patient time and resources are not wasted on ineligible candidates. The high precision of Trially Match results in a 73% reduction in screen failure rates. This outcome demonstrates that LLMs can deliver a highly accurate "diagnosis" of trial eligibility when sufficient clinical data is provided.
Automating Management and Addressing Staff Burnout
One key piece of the LLM clinical comparison study is the management stage (mang), which involves recommending appropriate clinical interventions. ChatGPT-4 was statistically superior to human emergency fellow physicians in management questions, suggesting LLMs are approaching practical clinical applications in this domain. However, the study notes that LLMs currently lack the ability to actively solicit information or perform practical operations.
This is precisely where Trially Connect, powered by the Margo AI Agent, intervenes to bridge the gap between AI recommendation and practical execution, particularly regarding the patient workflow. The LLM clinical comparison highlighted that AI’s most appropriate role is as a medical decision support tool to supplement gaps in physician knowledge and aid in complex processes.
Trially uses Margo AI to act as a specialized clinical research assistant, directly addressing the systemic crisis of healthcare staff shortages and burnout caused by excessive administrative overload. Clinicians often spend about 40% of their time on paperwork. Margo replaces the manual and time-intensive 'Patient-Physician Handoff' by automating the entire conversion funnel, including pre-screening, scheduling appointments, and sending reminders. This automation translates directly to efficiency gains for clinical research coordinators (CRCs), achieving a 90% reduction in manual EHR chart review hours per month.
Enhancing the Patient Workflow through Conversational AI
The LLM clinical comparison study reveals a critical transition point in the full clinical workflow at the management (mang) stage, which involves recommending appropriate clinical interventions. While LLMs such as ChatGPT-4 demonstrated superiority over human emergency fellow physicians in management questions, suggesting a strong approach toward practical clinical application, the study highlighted a significant limitation: LLMs currently lack the capacity to actively solicit information or perform the practical operational steps necessary for a complete clinical encounter.
This specific challenge is addressed by Trially Connect, powered by the Margo AI Agent, which bridges the gap between AI recommendation and practical execution in the patient enrollment workflow. Margo serves as a specialized clinical research assistant, automating follow-up steps and eliminating the manual and time-intensive "Patient-Physician Handoff". By handling the conversion funnel—automating pre-screening, scheduling appointments, and sending reminders—Margo fulfills the operational aspect of the management phase in the trial context.
This decision support role aligns directly with the LLM study's conclusion that the most appropriate application for LLMs is as a tool to supplement clinician knowledge. This automation effectively targets administrative overload, a leading cause of physician burnout, and achieves a 90% reduction in manual EHR chart review hours per month for clinical research coordinators (CRCs).
The LLM clinical comparison study focused on assessing performance across the four dimensions of medical reasoning (diff, diag, dx, mang); however, operationalizing these reasoning steps for the patient journey requires strong conversational AI capabilities. Margo AI acts as this essential conversational agent, ensuring patient engagement and adherence through timely outreach, scheduling appointments, and reminders via SMS and Voice. This functionality allows Margo to actively re-engage "dormant patient populations" and maximize conversion, aligning with the necessity to transform "complex medical guidance into clear, personalized interactions".
By combining Trially Match and Margo AI, the platform effectively translates the diagnostic accuracy and superior management capabilities observed in models like ChatGPT-4 and Doubao directly into the clinical trial environment. The platform automates the core trial steps that correspond to the four clinical workflow stages: Differential Diagnosis/Diagnostic Workup (via high-accuracy patient identification using LLM-powered prescreening), Management (via Margo's automated outreach and scheduling), and Diagnosis (resulting in a high-quality candidate pool and a reported 73% reduction in screen failure rates). This full-cycle automation across the patient workflow enables sites to multiply enrollment by 2-6X in complex studies.
