Harvey Castro, MD, MBA.

I am Advisor to the CEO of Phantom Space. I also still work clinical shifts in the emergency department. Most weeks those two jobs feel like they belong to different people.

In one, the question is where computation should physically sit. Orbit, edge, ground. Latency budgets. Power. What happens to a model when the link drops.

The other job asks whether the 78-year-old in bed 4, with a vague story and a soft blood pressure, is about to crash.

For a long time I treated these as separate problems. I don't anymore, and the reason is worth explaining, because I think most healthcare AI strategy is being written at the wrong altitude, aimed at a layer where the interesting engineering happens rather than the layer where the failures actually occur.

Here is what I mean. Almost every serious conversation about clinical AI right now happens at the level of the model. Which foundation model. What context window. How it scores on a benchmark built from board questions. Those are real questions. They are also the wrong place to be spending most of the attention, because the model is almost never what fails.

What fails is the last three feet.

The last three feet is the distance between where the inference happens and where the decision gets made. It is the nurse who cannot see the alert because it fired into a tab nobody has open. It is the recommendation that arrives four minutes after the patient left. It is the rural hospital where the connection drops during a transfer and the tool that was supposed to help is simply gone.

I have watched all three of those happen. None of them are model problems. All of them are infrastructure problems, and infrastructure is exactly what the space work taught me to think about properly.

When you design for orbit you can't assume the link. You can't assume power. You cannot assume that the clever thing you built on the ground will behave the same way once it is somewhere hard to reach. So you design for degradation. You decide in advance what the system does when it is cut off, and you make sure the degraded state is still useful rather than merely safe.

Healthcare hasn't done that work. It has built for the demo environment instead, which is a conference room with good wifi, a curated case, a patient whose record is complete, and an audience that already wants the thing to succeed.

A rural emergency department at two in the morning during a storm is closer to a spacecraft than it is to a conference room. Intermittent connectivity. Limited power redundancy. One physician, no specialist on site, and a decision that cannot wait for the link to come back.

If your clinical AI only works when everything works, you haven't built a clinical tool. You have built a demo that happens to run in a hospital.

So the question I now ask vendors is not which model they are using. It is: what does this do when the network drops mid-case? Where does the inference actually run? What is the degraded mode, and has anyone here watched a clinician use it?

Most can't answer. The ones who can are building something different from the rest.

There is a second thing the orbit work taught me, and it is less technical.

Distance is not the enemy. Distance is a design constraint. Satellites made distance tractable by refusing to pretend it was not there. Healthcare keeps pretending distance is not there. It builds for the academic medical center and then acts surprised when the tool does not survive contact with a critical access hospital 200 miles from the nearest neurologist.

Those hospitals are where the need is largest and where the deployments are thinnest. That is not a coincidence. It is a consequence of designing for the easy case.

I don't think the fix is more capable models. I think the fix is deciding, before anything is built, that the hardest environment is the design target and the easy one is the special case.

That is how you build for orbit. It is also, as far as I can tell, how you build something that still works at three in the morning when the person in bed 4 is getting worse and the link just dropped.

The model is not the hard part. It has not been the hard part for a while now.

TAGS: 2,91,25

On a recent shift, an ambient scribe listened to everything said in one of my rooms. It listened to the patient, not just to me. It heard her mention something about her husband that had nothing to do with her chief complaint and everything to do with why she was actually in my department at two in the morning.

That audio exists somewhere right now. I could not tell you where it lives, who can replay it, how long it is kept, or what happens to it in five years.

I am a board-certified emergency physician and I still take shifts. Ambient documentation is the best thing to happen to my charting in twenty years. I am not arguing against it. I am arguing that the capability arrived before the rules did, which is the normal order of operations in medicine and not a scandal by itself. What concerns me is how long it has been that way. Clinical settings have been recording routinely for years now, and three questions are still open.

The first question is consent

The form a patient signs at registration covers treatment and billing. It was drafted before a microphone sat in the room. A patient who nods at a clipboard while in pain is not making an informed decision about a voice recording, and anyone who tells you otherwise has never watched triage.

Then there is everyone else in the room. The daughter in the corner chair. The paramedic finishing a handoff. The patient behind the curtain who can hear my entire conversation, which means my microphone can hear his. In an emergency department a curtain is not a wall. It never was, but it did not used to matter this much, because human memory is lossy and a recording is not.

The second question is ownership

Whose voice is it, legally and practically?

The patient produced the sound. The hospital owns the chart. The vendor owns the infrastructure, and depending on the contract, may own something derived from the audio. Three parties, one artifact, and in most agreements I have read the patient is the only one without a seat.

A voice is biometric data. It identifies the speaker about as reliably as a fingerprint does, and unlike a fingerprint it carries payload. Accent, emotional state, respiratory effort, neurological signal. Research groups have built models that detect disease from speech alone. Set aside whether those models are ready for the clinic. The point is that a voice recording is not a transcript with extra steps. It is a specimen. Medicine already has rules about specimens.

The third question is the one almost nobody is asking

Can you prove the recording is real?

Voice conversion is cheap now and getting cheaper. Convincing synthetic speech runs on consumer hardware. If audio becomes part of the medical record, and audio can be altered or fabricated, then the record can be forged, and there is currently no routine way to prove a given file is what it claims to be.

Picture the deposition. A malpractice case turns on what was said about a risk. Both sides have audio. The files differ. What now? Picture a disciplinary proceeding where a physician insists the recording was edited. Picture a consent dispute where the patient says the same thing. None of this is exotic. It is the predictable consequence of putting unsigned audio into a legal record.

Where I stand, and my conflict

I have spent the last year filing on exactly this problem. Six US patent applications, all pending, all naming me as sole inventor. Four of them make one argument: authentication of clinical audio recordings, selective audio privacy using phoneme-based masking, a wearable voice privacy device, and detection of voice conversion and post-recording alteration in audio data.

So I am not a disinterested party and I will not pretend to be. I will say what I think anyway, because the alternative is staying quiet about something I have spent a year studying.

Three things should change, and none of them require new technology.

Consent should name the microphone. A separate line, plain language, a real opt-out, no penalty for taking it. If a patient declines, I type. I have been typing for twenty years.

The patient should get a copy. If a recording is about you, you should be able to hear it. The infrastructure to release it already exists, because patients already have a right to their chart. Audio is just another part of the chart nobody has gotten around to including.

Audio entering a record should be signed at capture. Cryptographic provenance at the device, not reconstructed later from logs. Prove the file has not moved since the moment it was made, or do not treat it as evidence.

What happens if the field skips this

The sound of a clinical encounter belongs to the person who made it and the person it is about. Everything else in the chain is infrastructure. Build the infrastructure to serve that principle and ambient AI becomes the best documentation tool medicine has had in a generation, which I think it will be.

Skip it, and hospitals will learn what an unprovable recording is worth the first time one is entered into evidence. That lesson is going to be expensive, and it is going to be taught by a plaintiff attorney rather than by a standards body.

Location: Dallas-Fort Worth, TX. Travels worl    Fees: 25000

Service Type: Service Offered