The Hook
For years, the holy grail of medical device regulation has been using “real-world evidence” (RWE)—data from everyday patient care—to speed up approvals. But that promise has consistently hit a wall: the data itself is often a messy, untraceable black box. Regulators don’t just want to know what the data says; they are increasingly demanding to verify how it was found, cleaned, and processed. A new launch this week suggests the industry is finally building the plumbing to meet those demands.
The News
On January 23, 2026, healthcare AI company John Snow Labs launched the Patient Journey Intelligence (PJI) platform. The company claims it is the first secondary-use data platform specifically architected to meet the FDA’s final guidance on Real-World Evidence, which was issued just last month in December 2025.
The new platform distinguishes itself by tackling the hardest part of medical data: the unstructured text. While traditional systems rely on structured fields like billing codes, PJI uses medical large language models (LLMs) and vision-language models to extract facts from clinical notes, pathology reports, and imaging. Most importantly, it addresses the FDA’s new “provenance” requirement by creating a deterministic audit trail. Every extracted fact can be traced back to its original source document, with a record of exactly which model and prompt was used to extract it.
Why It Matters
This release signals a maturation in health AI from experimental tools to critical infrastructure. The FDA’s December 2025 guidance was a wake-up call: it explicitly stated that structured EHR data alone is often insufficient for regulatory decision-making. By requiring “unstructured clinical narratives,” the FDA effectively mandated the use of advanced NLP (Natural Language Processing).
For health systems and device manufacturers, this platform attempts to solve the “data engineering fatigue.” Currently, organizations spend years building custom pipelines to convert FHIR data to research standards like OMOP, de-identify it, and normalize terminology. By unifying these steps into an audit-ready product, it could drastically lower the barrier to generating regulatory-grade evidence. This doesn’t just save money; it could allow a heart valve manufacturer, for example, to detect long-term safety signals years earlier than a traditional post-market study would allow.
The Skeptic’s View
However, “FDA-aligned” is a marketing term, not a regulatory status. The platform itself is not a medical device; it is a tool to generate evidence for them. The burden of proof still rests on the device manufacturer to show that the AI didn’t hallucinate.
There is a valid concern about the “black box” of the LLMs themselves. Even if the platform tracks lineage, the models (medical LLMs) are probabilistic. If a model infers a diagnosis from a vague physician note, does that count as a fact? John Snow Labs mentions “human-in-the-loop validation workflows,” but scaling that human review is expensive. If the AI is doing 99% of the work, the risk of automation bias—where humans just rubber-stamp the AI’s output—remains a significant hurdle for regulatory trust.
Looking Ahead
The real test will not be the software launch, but the first regulatory win. Watch for the first medical device application in late 2026 that explicitly cites this platform (or one like it) as the primary source of its clinical evidence. Until the FDA accepts AI-curated unstructured data for a high-risk device approval, this remains a promising technology in search of a precedent.