Canada recently became the first country to require clinical trial protocols in the ICH M11 format on the fourth. M11 is a rigid, standardized template that almost nobody is set up to produce, and we found very little tooling for it. We expect the US and the EU to follow as regulators modernize their systems. The end goal is a protocol that the FDA and EMA can audit by machine through FHIR. ADAPT already outputs FHIR, even though no one requires it yet. When the exact specifications for trial review are defined, we expect to need only minor tweaks.
The hard part is the conversion. Documents come in wildly varied forms, and M11 must be produced exactly. People are error-prone at following technical specifications, and that mismatch is why direct translation hasn't really been done. Today, big pharma companies hand this work to consulting firms to convert the protocols their teams already know how to write into the new required format. Very clearly there is a better way.
ADAPT automates that conversion. We started by learning how real protocols are actually written, including their quirks, such as tracked changes in Word, which most attempts miss. We use that knowledge to extract the content into an intermediate protocol schema we designed. For a full protocol, the schema often reaches over 100,000 lines of JSON. From there we rebuild the document in M11, including the title page, standardized section text, and tables laid out correctly. A validity check then tests the result against the standard to ensure today we are M11 compliant.
Generative AI does part of the work, but under a strict chain of authorship where generated values are marked as such in the JSON the whole time, displayed differently in the M11 format and given a Provenance in FHIR. That matters in a regulated setting, where a model that invents or alters protocol content is worse than no tool at all.
The result is that ADAPT converts an old-format protocol to the new format in a few minutes. The same job otherwise takes a team of medical writers multiple work weeks. Repurposing existing medical resources was critical in testing if our output was valid M11, and hours fine tuning the small mistakes in these hundred page documents is what makes our output unique.
