🤖 The stethoscope becomes a co-pilot

Doctors & AI — co-pilot, not autopilot

In 2024 the FDA had already cleared over 950 AI/ML-enabled medical devices. Most are radiology. The rest are crawling into pathology, cardiology, ophthalmology, and the EHR sidebar. Doctors are not being replaced. They are being merged. The choice of how is yours.

950+ FDA-cleared

🟢 Tier 1 — FDA cumulative AI/ML-enabled devices, 2024 updated: 2024

📡 Where AI is already in the clinic

Not science fiction. Active, billed, FDA-cleared.

🩻 Radiology

Stroke triage (Viz.ai), lung nodule detection (Aidoc, Lunit), mammography reading (DeepHealth, iCAD). Now reviewing >75% of FDA AI clearances.

🔬 Pathology

Paige.AI — first FDA-cleared AI for prostate cancer detection (2021). Whole-slide analysis enters academic centers.

👁️ Ophthalmology

IDx-DR (now Digital Diagnostics) — first autonomous AI cleared by FDA (2018). Diagnoses diabetic retinopathy without a doctor in the loop.

❤️ Cardiology

AliveCor + Apple Watch ECG — atrial fibrillation alerts. Caption Health — AI-guided cardiac ultrasound for non-experts.

📋 Documentation

Ambient scribes (Abridge, Nuance DAX, Suki) listen to the visit and draft the note. Now used by Kaiser, Mass General Brigham, Stanford.

🧬 Drug discovery

DeepMind AlphaFold — 200M+ protein structures. Insilico Medicine pushed an AI-designed drug to phase II. Cuts early discovery from years to months.

Source: FDA AI/ML-Enabled Medical Devices list 🟢 Tier 1 — Public Domain

🏗️ Hospitals & institutions building with AI

Not vendor decks. These are real flagship programs.

Institution	Program	Focus
Mayo Clinic	Mayo Clinic Platform	Distributed-data AI partner network (Solv.Health, Avive)
Cleveland Clinic	IBM-Cleveland AI partnership	Discovery accelerator, quantum + AI
Mass General Brigham	CCDS	Center for Clinical Data Science · radiology AI
Stanford Medicine	AIMI Center	Stanford AI in Medicine & Imaging
Google Health / DeepMind	Med-Gemini, AlphaFold	Multimodal clinical reasoning, protein structure
Microsoft + OpenAI + Epic	GPT-4 in EHR	Inbox draft replies, note summarization at UC San Diego, Stanford
NIH	Bridge2AI	130M $ federal program for ethical biomedical AI datasets
WHO	Ethics & governance of AI for health	Global guidance on multi-modal clinical AI (2024)

📊 What the evidence actually shows

Headline claims have a way of cooling down once peer review arrives. Here's what the literature has settled on.

✅ Solid wins (peer-reviewed, multiple trials)

Diabetic retinopathy screening — non-inferior to ophthalmologists in primary care
Stroke large-vessel-occlusion alerting — faster door-to-needle time
Sepsis early warning (e.g., Epic / TREWS) — earlier antibiotic time
Ambient scribing — measurable reduction in physician note time and after-hours work

⚠️ Mixed (real but fragile, hospital-dependent)

Generative AI for clinical question answering — high accuracy in benchmarks, hallucinations in real cases
Mammography AI — equivalent on average, worse on under-represented populations
Pathology AI — strong on common cancers, weaker on rare entities

❌ Famously failed

IBM Watson for Oncology (MD Anderson 2017–2018) — recommendations called "unsafe and incorrect" in internal memos
Several COVID-19 imaging AIs (2020–2021) — Nature Machine Intelligence review found "none of the 232 papers reviewed were suitable for clinical use"

References: NEJM · Nature Machine Intelligence · JAMA 🟡 Tier 2 — peer-reviewed

🧭 What doctors will not (yet) hand over

The replaceable parts of medicine and the sacred parts are not the same. The first is bigger than most people think. The second is smaller than most people fear.

Likely automated by 2030

First-pass radiology screen
Routine pathology grading
Clinical note drafting
Inbox triage
Insurance pre-authorization replies

Mixed (human + AI)

Differential diagnosis
Treatment planning in oncology
Discharge planning
Surgical pre-op planning

Stays human (for now)

Breaking bad news
End-of-life conversations
Pediatric / vulnerable adult exam
Open-ended clinical interview
Final accountability for any decision

"AI will not replace radiologists. Radiologists who use AI will replace radiologists who don't."
— Curtis Langlotz, Stanford AIMI (paraphrased; widely attributed)

⚠️ Warnings worth keeping

Bias amplification. An AI trained on Boston EHRs is not a doctor for rural Mississippi. Performance gaps by race, sex, age are documented and persistent.

Liability is unsettled. If AI says benign and the lesion was malignant, who is sued — the doctor, the hospital, or the vendor? US courts have not answered.

Hallucination is not a synonym for "rare error". Generative models confabulate citations and dosages with full confidence. The cost in medicine is bigger than in code.

Quiet de-skilling. A junior radiologist who never reads without AI may not learn to read without AI. The next generation's ceiling is the previous generation's floor minus AI assist.