Real-timeclinical trials.

A real-time clinical trial is one where protocol compliance, data quality, and safety signals are evaluated as the trial happens rather than reconstructed afterwards. In a conventional trial the protocol lives in a PDF, the data lives in an EDC, and the two are only compared during monitoring visits and data cleaning — which is why a systematic error at one site can run for months before anyone sees it. In an RTCT the protocol itself is compiled into a machine-readable model, so every visit, form, and field is checked against it at the moment of capture. The practical difference is when you learn: minutes instead of the interval between monitoring visits.

A live dashboard shows you data faster. It still cannot tell you whether that data complies with your protocol, because a dashboard has never read your protocol. RTCT is about real-time judgment, not real-time display: the system holds the protocol’s actual requirements — day windows, the schedule-of-activities matrix, eligibility bounds, consent-before-procedure, safety reporting windows, prohibited medications, visit sequencing — and evaluates them continuously. A dashboard tells you a visit happened on Day 22. RTCT tells you the protocol allowed Day 18 plus or minus 2, quotes the sentence that says so, and records the deviation with its category and severity.

The protocol intelligence layer runs a staged pipeline over the source documents. It parses the document, including scanned pages and the schedule-of-activities tables; harvests the requirements the protocol states in prose; plans the visits and forms; builds each form in depth against a CDASH-aligned field library; reconciles the harvested rules against the forms they constrain; compiles a deterministic study bundle with provenance attached to every element; and finally walks a synthetic patient through the result to catch structural faults before a human ever sees it. What comes out is a portable study definition — visits, forms, fields, branching logic, validation rules, arms and cohorts, task definitions, and a protocol-deviation specification.

No. Every generated study lands in quarantine — a study that exists so you can review it section by section, but that is hidden from normal study lists and cannot enroll anyone until a person explicitly activates it. The import reports what it did with every single element it received: what was written into the study, what was kept verbatim but not acted on, what was deliberately dropped and why, and what it did not recognize. Where two source statements disagree about the same field, both readings are carried through unresolved for a human to settle, because the pipeline cannot know which one is right.

An amendment is treated as a diff against the study you are already running, not as a fresh import. The system works out which incoming form replaces which existing one, then — and this is the part only the EDC can do — counts how much collected data sits behind every artifact the amendment touches. “This field was retired” is not a decision; “this field was retired and 412 subjects have values recorded against it” is. Anything the crosswalk could not confidently place is held in its own bucket at the top of the review rather than folded into a safer-looking category. You approve a matrix, and the matrix is stored, so your approval refers to something fixed.

Data does not only arrive when a coordinator types it. AccuraTrials accepts patient-reported outcomes through a patient portal between visits, versioned eConsent with a signing workflow, vitals and labs over FHIR from EHR systems such as Epic and Cerner, randomization and trial supply events from the built-in RTSM, and paper source documents through OCR that pre-fills the real eCRF with every value traceable to its position on the scanned page. All of them land in the same validated forms, under the same validation rules, and inside the same 21 CFR Part 11 audit trail as data typed at a workstation.

It moves the work earlier, which is where it is cheapest. A field-level check costs a site seconds at entry; the same error found in data cleaning costs a query, a site response, a monitor review, and a re-entry, months later, against a memory nobody trusts. The escalation levels exist precisely so that speed is preserved: only genuinely unsafe values block a save, most issues raise a query or a warning, and a visit-timing concern asks for a confirmation rather than a refusal. Nothing in the design makes a coordinator wait on a model.

It is the direction both point in. ICH E6(R3) asks sponsors to manage quality proactively and proportionately to risk rather than to inspect it in at the end, and risk-based quality management is much easier to evidence when your protocol requirements are executable and your deviations are detected and categorized as they occur. Part 11 is not relaxed by automation — it is what makes automation defensible. Every automated evaluation writes to the same immutable audit trail as a human edit, with the clause it enforced, the protocol version it came from, and the person who accepted it.