Transcript: How physicians review and sign AI-drafted medical documents
An AI-generated audio overview (Google Gemini), transcribed by machine on Blaine's Mac (Whisper). The audio was checked against primary sources; what it gets wrong is listed at https://blainomd.com/learn#audio-doctors-verify-ai-decisions

So welcome to this deep dive. You know, usually hospital software is, well, it's like a massive
immovable concrete bunker. Oh, absolutely. Yeah. Like if you want to change the plumbing,
you basically have to tear down the whole building. But today we are looking at a
completely new blueprint. We're examining some architectural evaluations of platforms like
MedPlum and ClinicalSwipe. Right. Which are really changing the game.
Exactly. Our mission here is to uncover how modern healthcare is basically tearing up that
concrete. They're decoupling where your medical data is stored from who actually legally signs off on
your AI generated health decisions. Yeah, it's a huge shift. Okay. Let's unpack this. Traditional
electronic health records are monolithic, right? It's all or nothing. But these new tools offer an
entirely modular API first approach. So looking at our sources, they describe MedPlum as a,
and this sounds dense, headless FHIR R4 backend. Yeah, I know. It sounds super technical. Right.
So what are they actually doing differently to avoid locking hospitals into one of those giant
ecosystems? Well, think of MedPlum as basically a highly standardized digital filing cabinet. But the
front end user interface is completely detached. That's the headless part. Got it. So it operates
purely as a system of record. It holds the data, but it doesn't force you into total ecosystem lock-in
just to access it. Okay, wait. So is MedPlum basically the stripe of healthcare? Oh, that's a
good way to put it. Like you just plug in the specific infrastructure layers you need. Say you
want to use external Okta Identity Federation, or you need to use the MedPlum agent to tunnel those
really old legacy 1987 HL7 hospital device feeds. Exactly. Yeah. You only consume what you need.
And because it's built like that, engineering teams can use these MedPlum bots to easily automate AI
workflows. Right. Which is amazing, but it creates a massive regulatory bottleneck. I mean,
an AI can draft clinical notes all day long, but laws like California's AB 1979 actually explicitly
ban an autonomous AI from legally signing those decisions. Right. Because obviously an AI cannot
legally practice medicine. Exactly. So if AI can't sign off, how do we verify these AI drafts
without just reverting to, you know, really inefficient paper pushing? Well, that's where clinical
swipe comes in as the system of attestation. Okay. It acts as a conduit, not a cloud. It doesn't
store the long-term patient records. It just routes those AI drafts like, say, a letter of medical
necessity directly to a specialty matched state licensed physician for review. Okay. Here's where
it gets really interesting. The economics of this are wild. Clinical swipe uses a flat fee model.
Yeah. It's super clever. Right. The doctor gets paid the exact same flat fee for an honest decline
as they do for an approve. It totally eliminates that incentive to just rubber stamp things, which is
usually why these platforms fail audits. What's fascinating here is how they enforce that.
Doctors use this really simple three-gesture mobile interface. Oh yeah. Just sign, revise,
or decline, right? Right. And it forces a mandatory 20-character clinical rationale if they don't just sign
it. Plus, the system tracks the AI text changes line by line. Oh, but wait. If they're acting like a
conduit, not storing the records, how do they prove to an auditor like the IRS or Medicare that the
doctor actually read the chart? Without exposing private patient data, you mean? Exactly.
How do you do that? So they solve this through a cryptographic handshake using the hash care
ledger. Okay. When the doctor makes their decision, it gets hashed using SHA-256, and that hash is
countersigned daily into the Bitcoin blockchain. Oh, wow. So they're just putting the mathematical
proof of the ledger, not the actual private medical chart. Exactly. And then MedTlum pulls that log back
in via a FHIR provenance resource to securely lock the chart. That is wild. So what does this all mean
for you? It ensures that when AI inevitably drafts your family's medical paperwork, a legally accountable
human is mathematically proven to have reviewed it. Right. It's real accountability without the bulk.
But think about this. If this platform captures every single line-by-line correction a human doctor
makes to an AI draft, could this exact audit trail eventually be the dataset used to legally prove when
AI finally becomes more accurate than the humans correcting it? Oh, that is a huge question.
Definitely something to mull over. Catch you on the next Deep Dive.