The headache
AI in medicine
AI won’t replace doctors.
But doctors who understand AI
will replace those who don’t.
This isn’t a talk about replacing medicine with AI.
It’s about learning where AI helps—and where it can hurt.
CLINICAL AI / SESSION 001Human supervised
Patient context
History
Examination
Medication list
Evidence
Clinical judgment
Your guide
About me
- DevRel at Docker
- Founder @Collabnix
- Docker Community Leader
- Distinguished Arm Ambassador
- Worked at Dell EMC, VMware, Redis
Join at
slido.com
#3474 854
Active poll
Which AI tool have you used in the past?
The good
AI is not coming to healthcare.
It’s already here.
AI + Pharmacology
From molecule to bedside support.
Drug discovery
↓Molecular screening
↓Drug repurposing
↓Interaction detection
↓Adverse-event analysis
↓Literature synthesis
↓Clinical support
AI can process more information.
It does not automatically mean AI understands the patient.
It does not automatically mean AI understands the patient.
Clinical reality
Your colleagues are already asking AI.
ChatGPT
Gemini
OpenEvidence
Copilot
Medical AI tools
Relative visual emphasis only · not a statistical ranking
What are doctors asking AI?
“Summarize this paper.”
“Explain this mechanism.”
“What are the differential diagnoses?”
“Check these drug interactions.”
“Explain this lab result.”
“Draft patient instructions.”
“What dose should I prescribe?”
That last question changes everything.
Audience pulse check
Have you used a general-purpose AI tool for a medical question?
The bad · the confidently wrong
AI tool
horror stories.
When a plausible answer becomes a clinical risk.
The wrong cream
The drug interaction
The hallucinated paper
The perfect dose
Under the hood
AI doesn’t think like your professor.
Question
↓Tokens
↓Probability
↓Next token
↓Next token
↓Answer
FluentProfessionalDetailedConfident
Fluency ≠ truth
Confidence ≠ certainty
AI output ≠ clinical decision
Confidence ≠ certainty
AI output ≠ clinical decision
Designing safer behavior
What if AI asked questions
before giving the answer?
Bad AI
VSPatient question
↓Immediate answer
Safer clinical assistant
Patient question
↓Clarify
↓Collect context
↓Check red flags
↓Identify missing information
↓Retrieve evidence
↓Express uncertainty
↓Escalate when appropriate
↓Answer
The safest AI is sometimes the AI
that refuses to answer yet.
that refuses to answer yet.
Major reveal
So I built my own.
Private medical AI assistant.
What if the medical AI ran locally instead of sending everything to a public chatbot?
Doctor
Running locally
Medical chat interface
Safety / context layer
Local medical model
Medical knowledge / references
Local execution improves control. Security still depends on the surrounding system.
Model ecosystem
Think of Hugging Face as a model ecosystem.
A large ecosystem includes models targeting medical Q&A, radiology, pathology, dermatology, biomedical language, imaging, clinical text, drug discovery, protein science, and embeddings.
Google
MedGemma
medical textmedical imagesmultimodal
Medical model family from GoogleMicrosoft Research
BiomedCLIP
biomedical visionimage-text
Microsoft
BioGPT
biomedical NLPgeneration
Community
ClinicalBERT
clinical textembeddings
Model ecosystem
Protein language models
protein sciencedrug discovery
Results reflect this curated teaching demo—not a live ecosystem count.
Featured family
Meet MedGemma.
MedGemma
A family of open models designed for medical text and medical image comprehension.
Medical text
Clinical questions
Radiology
Dermatology
Pathology
Ophthalmology
Medical images
Medical model ≠ medical deviceModel performance ≠ clinical validation
A healthcare-focused model still needs validation for its intended use.
Hands-on concept
Your laptop can become
a private AI laboratory.
$ docker model run hf.co/google/medgemma-4b-ptDownloading model layers…Starting local inference…✓ Interactive session ready
Hardware needs vary. Larger models require substantially more RAM or VRAM. Quantized variants can reduce requirements when available and appropriate.
Model downloads locally
↓Inference runs locally
↓Your application talks to the local model
↓Build your own medical assistant
Real-world proof
From prompt to working demo
in about ten minutes.
A real project: an AI coding agent built a live CPR coaching tool—pose detection, compression rate, and arm-angle feedback—end to end.
~10 minutes from prompt to working demo
An AI coding agent built this real-time CPR coaching tool—pose tracking, compression-rate feedback, and arm-angle correction—end to end, in about ten minutes.
- 0:00Described the goal in plain English: coach real CPR from a phone camera.
- 0:02Agent scaffolded a MediaPipe / YOLO11 pose-detection pipeline.
- 0:05Added compression counting, rate targeting, and arm-angle checks.
- 0:08Wired the live video overlay with OK / ADJUST feedback.
- 0:10Running demo — coaching real chest compressions in the browser.
Interactive centerpiece
A safer medical AI conversation.
Questioning improves context. It does not turn a language model into a clinician.
Private Medical AI AssistantEducational demonstration
Dedicated teaching mode
Pharmacology lab.
Mechanism
Beta blockers
A receptor-level explanation. Select a lens to inspect the same concept.
β-adrenoceptor blockade
↓ cAMP signaling
↓ heart rate + contractility
↓ myocardial oxygen demand
Patient-specific dose questions require indication, age, route, formulation, organ function, medicines, and other clinical context.
Architecture choices
Public AI vs private AI.
DimensionPublic general-purpose AIPrivate / controlled medical AI
Data control
Provider-dependent
Higher control; system-dependent
Model choice
Provider catalog
Chosen by your team
Customization
Usually limited
Potentially extensive
Knowledge sources
Often opaque or mixed
Can be curated
Auditability
Variable
Can be designed in
Offline operation
Usually no
Possible
Latency
Network-dependent
Hardware-dependent
Cost
Subscription / usage
Hardware + maintenance
Governance
Shared with provider
Your organization owns it
Security responsibility
Shared
Primarily yours
Local ≠ safeMedical model ≠ doctorPrivate ≠ validated
The right equation
The doctor + AI model.
AI alone
Fast, but can miss context.
Doctor alone
Clinical judgment, but limited time and information bandwidth.
Doctor + well-designed AI
Human judgment + computational assistance.
Doctor + AI + evidence + context + governance
= Augmented medicine
= Augmented medicine
Keep these
Seven rules for medical AI.
Never confuse confidence with correctness.
Ask for context before answering.
Verify clinical claims and citations.
Protect patient information.
Know which model you are using.
Understand the model’s limitations.
Keep a qualified human responsible for clinical decisions.
The final prescription
AI won’t replace doctors.
But doctors who understand AI
will replace those who don’t.
Don’t fear AI.Understand it.Question it.Verify it.Control it.Use it well.