DR GPT™ · Weekly Medical-AI Intelligence

Future Health Intelligence

The benefit is real. The proof is thinner than the headline.

Every week I read what the medical-AI headlines actually prove, check the claim against its primary source, and show you the one rung the breakthrough is stuck on. I am Harvey Castro, MD, an emergency physician who has watched a lot of confident technology meet a crowded department at 3 a.m.

Get the weekly brief Follow on LinkedIn Watch my TEDx talks

Verified to primary sources A new signal every week Updated November 9, 2026

01 · The framework

The Ladder

Every breakthrough is a ladder, and you read it from the bottom up. Each rung has to hold before the one above it can, so a press release usually celebrates the bottom rung while the rung above it sits empty. I mark the rung a story is actually stuck on, because that rung decides whether any of it reaches a patient.

Done

Proven, published and in place. Nothing left to argue about.

At risk

The one stuck rung. Everything above it waits on this.

Not yet

Still empty. Nobody has the evidence to fill it.

The goal

A patient is measurably better off. That is the whole point.

How to read itRead bottom to top ↑

A study is publishedDone

The result holds up on a second lookDone

It works in the messy place where care happensAt risk

Someone is responsible for acting on itNot yet

It is paid forNot yet

A patient is better offThe goal

Why this matters. Most coverage stops at the second rung and calls it a breakthrough. The stuck rung is the story. It is also why two tools can carry the same clearance and change completely different numbers of lives.

02 · This month's signal

Four weeks, read straight

Only cleared devices and published studies make this list. Each card names the evidence class, gives you the honest chart where a clean comparison exists, and ends with the rung the story is stuck on.

LeadNov 2 to 8Established

Cleared to read, with nobody reading it

PulseAI won FDA clearance for a wearable AFib rhythm model that cuts the inconclusive ECG reading from about one in five toward one in a hundred. The watch does not diagnose anything. It triages a waveform and routes it to a human, and that human is the part nobody has paid for yet.

Inconclusive ECG readings

Source: FDA 510(k) clearance, November 5, 2026. Rates as reported by the company and in Heart Rhythm (n=296). Availability and coverage vary by country and plan.

Stuck at integrationA cleared cloud model is not a model wired into the watch with a clinician paid to read the result. Until that last step exists, the clearance changes a label, not a life.

LeadOct 26 to Nov 1Promising

Found 475 days early, nobody lives longer

Two groups caught pancreatic cancer on CT scans people had already had. Mayo's model, REDMOD, flagged it a median of 475 days early, and PANORAMA beat 68 radiologists on the same images. All of it retrospective, which means the model can see something a reader misses, not that a patient walks out healthier.

The same week, the federal AI model-card disclosure label was deleted. Nobody announced that part.

Cancers caught, same scans, two readers

Source: REDMOD, Mayo Clinic, published in Gut, 2026. Comparison figures from the PANORAMA study. Both from retrospective test sets, not cleared for this use.

Stuck at prospective proofNot one living patient has been shown to live longer. Finding a cancer earlier is a measurement. Surviving it is the endpoint, and that study has not been run.

LeadOct 19 to 25Established

The evidence file became the legal record

The FDA closed its comment period on how to regulate generative AI devices, and eight state attorneys general sued over AI medical triage. The paperwork that proves a device works is now also the paperwork a court will read.

I have signed charts that ended up in a deposition. Once a file is discoverable, the quality of the documentation stops being an academic question.

Stuck at the competency benchmarkThe FDA itself calls competency assessment a concept rather than a validated method. You cannot regulate a device against a benchmark nobody has agreed on.

LeadOct 12 to 18Company claim

The pathway moved, not the algorithm

SimBioSys cleared the first MRI-derived breast cancer recurrence score, a genuine first for imaging. In the same week, a Tanzania trial showed the referral pathway, not the model, nearly doubled specialist follow-up. Read those two together and the lesson gets uncomfortable for anyone selling software.

Stuck at head-to-head proofOncologists already order Oncotype DX and MammaPrint. A new score has to beat the assay they trust on recurrence-free survival, not just agree with it.

03 · Where each one is really stuck

The ladder behind every story

Same four stories, this time rung by rung. The amber rung holds everything above it, and the line underneath says what it means in plain English.

PulseAI, wearable rhythm modelRead bottom to top ↑

A model can classify a rhythm stripDone

Clearance as a medical deviceDone

Runs inside the watch on your wristAt risk

A clinician is paid to read the resultNot yet

A reimbursed review pathway existsNot yet

A patient whose stroke was preventedThe goal

What the stuck rung means. The clearance covers a cloud model. Getting it onto the device and attaching a paid reviewer to the alert is engineering and reimbursement work, not a better algorithm. That is why I said 55 percent, not 90.

REDMOD and PANORAMA, pancreatic imagingRead bottom to top ↑

AI finds the cancer on scans already takenDone

Peer reviewed and beating readers on test setsDone

Prospective trial in a screening populationAt risk

Survival benefit shownNot yet

Cleared for early detectionNot yet

A patient treated earlier and living longerThe goal

What the stuck rung means. A retrospective result tells you the model can spot a pattern. It cannot tell you how many false alarms you get when a hundred thousand healthy people walk through the scanner. That number is the whole argument.

Generative AI devices, the regulatory fileRead bottom to top ↑

AI devices are regulated as devicesDone

Public comment closedDone

A validated competency benchmarkAt risk

Final binding guidanceNot yet

Courts settle triage liabilityNot yet

Rules a clinician can rely on at the bedsideThe goal

What the stuck rung means. You cannot write a rule against a benchmark that has not been validated. Until competency assessment stops being a concept, guidance stays provisional and every early adopter carries the risk personally.

SimBioSys, MRI recurrence scoreRead bottom to top ↑

A recurrence score derived from MRIDone

Clearance to sell itDone

Prospective head to head against the genomic assaysAt risk

Recurrence-free survival benefitNot yet

Oncologists drop the assay they order todayNot yet

Fewer women treated who did not need itThe goal

What the stuck rung means. Agreement with an existing test is a nice start. Replacing it requires a prospective trial on the endpoint oncologists already trust. The Tanzania result says the pathway might matter more than the model anyway.

04 · The signal log

Every week, newest first

One line per week, with the tier, the evidence class and the rung it is stuck on. This table grows, and older weeks stay put so you can check whether I was right.

Four weeks of medical AI, scored against primary sources.
Week	The lead	Tier	Evidence class	Stuck rung
Nov 2 to 8	Wearable AFib model clears, cutting inconclusive readings toward one in a hundred	Lead	Established	Integration
Oct 26 to Nov 1	Two groups catch pancreatic cancer on existing CT, a median of 475 days early	Lead	Promising	Prospective proof
Oct 19 to 25	FDA closes comments on generative AI devices, eight attorneys general sue over triage	Lead	Established	Competency benchmark
Oct 12 to 18	First MRI-derived breast recurrence score clears, Tanzania shows the pathway moved	Lead	Company claim	Head to head proof

Tier reflects how much of the week's attention a story deserves, not how excited I am about it. Evidence class reflects what kind of proof exists right now.

05 · The prediction tracker

Public, dated and falsifiable

A forecast without a falsifier is just an opinion, so every call below names the date and the thing that would prove me wrong. All four are open. I score them in public when they resolve.

A cleared wearable arrhythmia model ships inside a smartwatch with a reimbursed clinician review pathway

Open

Target November 2027Logged November 9, 2026

55% confidence

Wrong if: there is no live payment pathway for the review step at twelve months.

No pre-diagnostic pancreatic-cancer AI holds FDA clearance for early detection on standard CT

Open

Target November 2027Logged November 9, 2026

75% confidence

Wrong if: the FDA clears any such device for that indication within twelve months.

The FDA publishes draft guidance, not a binding final rule, on competency assessment for generative AI devices

Open

Target July 2027Logged November 9, 2026

60% confidence

Wrong if: a binding final rule lands, or nothing at all lands, by the target date.

No prospective study shows an MRI recurrence score beats genomic assays on recurrence-free survival

Open

Target December 2027Logged November 9, 2026

70% confidence

Wrong if: a prospective trial beats Oncotype DX or MammaPrint on that endpoint.

06 · From the DR GPT desk

Where else this work shows up

Same analysis, different rooms. Broadcast, stage and print.

On camera

Telemundo, AI in health

My segment on what AI is actually doing in a hospital right now, and the clip library that grew out of it.

Watch on YouTube → On stage

Speaker reels

Keynotes on healthcare AI for clinical, executive and innovation audiences. Reels and booking details live on my main site.

harveycastromd.com → On the page

The weekly brief

Where each week's signal gets the full treatment, including the charts and the sources I did not have room for here.

Read on Substack →

Briefing archive

November 2026

The Ladder applied to the cleared watch and the pancreatic imaging results, with the deck, the pocket brief and the audio briefing.

Open the November briefing → Briefing archive

October 2026

Health alerts beyond the screen, the trial that started this series, and the questions that shaped November.

Open the October briefing → The series

Future Health with DR GPT™

The monthly patient-facing briefing, with every deck, one-pager and recording kept on one page.

See all editions →

Six talks, six ideas

See my six TEDx talks

TEDx is where the ideas behind this brief get the stage treatment. If you want the argument rather than the weekly data, start there, then come back here for what changed this week.

Watch the TEDx talks

07 · Follow DR GPT™

Pick the room you like

Same signal, different format. The brief goes out weekly on Substack, and the rest is where I answer questions.