AACE 2026: AI systems may improve diabetes trial recruitment and enrollment 

Artificial intelligence-assisted screening showed strong evidence of improving the productivity and accuracy of clinical trial recruitment among people with type 2 diabetes mellitus, according to a research abstract presented at the 2026 AACE Annual Meeting in Las Vegas. 

Researchers evaluated an artificial intelligence (AI)–enabled system with four integrated components:

1. Trained large language models (LLMs): record extraction pipeline, clinical intelligence layer, and patient eligibility algorithms. 

2. Clinical intelligence layer: therapeutic-area-specific knowledge graph synthesizing clinical guidelines, molecular profiling, and patient-reported outcomes for contextual patient matching.

3. Precise record retrieval, ingestion, and matching engine: automatic trial protocol and guideline ingestion to create tailored patient matching algorithms. 

4. AI coordinator: specially trained AI Agent for patient engagement and nurturing. 

The researchers examined 1,311 patients with diabetes, of whom 377 were deemed eligible for a clinical trial; 60 were eventually enrolled. The AI system achieved 91% overall accuracy in identifying eligible patients compared with 82% for manual review. For trial matching, accuracy was 91% with AI vs 63% with manual processes. The AI system reduced the time from initial screening to final trial recommendations to less than one week compared with approximately 18 weeks for conventional prescreening. 

Performance metrics for the AI approach included sensitivity of 0.95, specificity of 0.95, precision of 0.87, and an F1 score of 0.91.  

Future improvements will include enhanced predictive modeling of patient attitudes, such as willingness to participate and propensity to complete all visits. 

The study was conducted by investigators affiliated with Areti Health. No additional funding sources or conflicts of interest were reported. 

(Editor's note: According to the lead researcher, "final analysis showed better results than initial findings at time of abstract submission.")

Expert Insight

AACE Endocrine AI invited lead researcher Joshua Ransom, PhD, chief strategy officer and data protection officer at Areti Health in Hillsborough, CA, to elaborate on the findings. 

Why does this study matter?

Clinical trial recruitment is one of the biggest bottlenecks in bringing new diabetes therapies to patients. The manual process of sifting through medical records, checking eligibility criteria, and reaching out to patients is incredibly time-intensive and often inconsistent. Coordinators are doing their best, but when they’re juggling dozens of inclusion and exclusion criteria across complex diabetes profiles with multiple comorbidities and medication histories, things can be missed. Our study shows there is a better way.  

What are the biggest obstacles to recruiting patients for diabetes studies, and how do the AI systems collectively address those problems?  

One of the main issues is that many diabetes trials are overly restrictive and exclude otherwise eligible participants. This is due to the complexity of diabetes research, with different subtypes and comorbidity profiles (eg, cardiovascular disease, chronic kidney disease, retinopathy), as well as the diversity of concomitant medication regimens. Together, these factors make manual eligibility screening very labor-intensive and error-prone.  

AI systems help streamline the process and reduce friction points, which can decrease the number of potential participants who initially express interest but later drop off or lose interest.  For example, with the AI system, engagement with potential candidates is immediate. Within 60 seconds of a patient expressing interest—such as clicking on a trial website or advertisement—the system initiates a conversation, answers questions, and guides them through the process, in a compassionate, compliant, and patient-centered manner. If a patient is not eligible for a specific study, we will keep them engaged as potential candidates for future trials.

 The systems also help with scheduling screening visits without handing patients off to another system, which can increase the drop-off rate. If they need a four-week washout period before a visit, such as after they recovered from a cold, the AI systems help them reschedule and support them through the process, including reminders about when to fast, how to prepare for their visit, and
where to park.  

Was there a finding, or perhaps more than one, that surprised you? 

The gap in matching accuracy really stood out. We demonstrated that an AI platform can screen more than 1,300 patients and match them to trials with 91% accuracy, compared with 63% for manual matching. That 28-point difference represents real patients who might otherwise have been overlooked, people who deserved a chance to participate in research that could benefit them and others like them. We were also encouraged by the 0.95 sensitivity. Knowing the system identifies 95% of truly eligible patients gives us confidence that we're not leaving enrollment opportunities on the table. 

How might the findings influence clinical practice for diabetes?

 In the near term, the most immediate impact is operational. If sites and sponsors adopt AI-assisted recruitment, enrollment timelines could realistically be compressed from months to weeks, allowing patients to access therapies more quickly.  There is also the broader implication for equity. One of the underappreciated problems with manual screening is that it tends to favor patients who are already visible—those who visit frequently, have well-organized charts, or see physicians aware of open trials. AI reduces this bias by systematically reviewing every patient population, ensuring that individuals who might otherwise be overlooked receive fair consideration.  More broadly in endocrinology, the knowledge graph approach we used—which synthesizes clinical guidelines, molecular profiles, and patient-reported outcomes—is readily adaptable to other conditions like obesity, thyroid disorders, or adrenal insufficiency.  

Is there anything else you'd like to say about this work? 

This isn’t about replacing the clinical research team, but about helping them focus on what matters most: human-to-human interaction.  AI handles high-volume repetitive tasks such as record extraction and eligibility screening, freeing coordinators to build trust with patients, guide them through informed consent, and support them throughout the trial. We estimate that this approach can save hundreds of hours of manual work per trial. That's not just an efficiency gain, it’s a meaningful step toward reducing burnout among clinical and research staff who are already stretched thin.