SiteWorks demonstration · November 2026
AI prescreening that reads a volunteer's own health record
We built a prescreening assistant into SiteWorks (siteworks.health) for Harbor Marsh Research, an invented research site near Charleston. The demonstration volunteer is Camila Lopez, Epic's public test patient, and everything runs in Epic's developer sandbox. The volunteer asks about a chronic pain study, connects a health record through Epic's MyChart, and the assistant checks that record against studies recruiting nearby, cites the chart for every decision and books a screening call.
Signing in to MyChart and approving access
The chat opens with one line from the volunteer: a doctor said the volunteer might qualify for a study for chronic pain. The assistant offers to check recruiting studies against the volunteer's own record and puts a Connect button beside the chat.
Connecting uses SMART on FHIR, the standard way a patient app asks a health system for a record. The volunteer signs in on Epic's own MyChart page, and Epic's consent screen lists exactly what the app wants: demographics, conditions, medicines, lab results, visits and vitals. Access is read-only and lasts one hour. The assistant runs on Microsoft Azure, gpt-5-mini on Azure AI Foundry signed in with Microsoft Entra, and Azure Prompt Shields screens every message the volunteer types.
The volunteer's chart and a live ClinicalTrials.gov search
The chart opens on the canvas beside the chat: four conditions (chronic pain made worse by daily activity, ischemic chest pain, polycystic ovaries and a stomach ache), one medicine, two lab results and six vital signs. Each entry gets a short reference, like C4 or M1, so anything the assistant says can point back to it.
The assistant then searches ClinicalTrials.gov live for studies recruiting within 50 miles of Charleston for those conditions. On the day of filming it found six, and it read the full eligibility criteria of the three closest matches.
A checklist for each study, with every decision cited
Each study gets a checklist. Every criterion is marked met, not met or unknown, and every met or not met carries the chart reference that decides it. The best fit was a Charleston study of neuromodulation plus prolonged exposure therapy for chronic pain with PTSD: three criteria met, four unknown, none failed, so it's rated a possible fit.
Age and sex are met from the record, and the pain criterion is met because the chart has listed pain made worse by daily activity since May 2023, well past the six months the study asks for. The chart doesn't mention PTSD, so that criterion stays unknown and goes on the list for the screening call, with the study's other open questions such as pregnancy and any history of seizures. Before any checklist reaches the screen, a check on the server confirms every cited reference exists in the chart.
How the assistant rates the other studies
The third study tests stem cell therapy for pain from chronic pancreatitis. The chart lists a stomach ache but no pancreatitis diagnosis, so the assistant rates the study an unlikely fit and shows the reason. Where the chart is silent, the checklist says so: the study wants a stable opioid dose, the medicine list has no opioids, and that row reads unknown, since a missing entry is not proof either way.
The assistant recommends the study with no failed criteria and asks whether the volunteer would like a research coordinator to call. A second study, for lower back and leg nerve pain, also came back with nothing failed, and it stays on the canvas for later questions.
Referral, booked call, and what a real site needs
The volunteer says yes. Referral R-1041 lands in the recruiter's queue, filled in with the checklist, a plain summary of the chart and the open questions for the call. The coordinator, Renee Alston (an invented person), books a 90-minute screening for Monday at 1:30 PM, and the booking appears in the volunteer's chat. The recruiter can take over the conversation at any point and hand it back.
The whole run took 24 engine steps, 46.6 seconds of agent time and $0.018 of model time. The chart stays in the session: nothing is stored and nothing is indexed. A real deployment needs a business associate agreement with the covered entity, which Azure OpenAI is covered by under Microsoft's terms, and the chat script is recruitment material, so it needs IRB approval before volunteers see it.
Want this working against your own systems? Call (844) 422-7000.
See the prescreening demo →