Looking for my fertility education site? Visit angelathyermd.com
Physician · Behavioral scientist · Twenty years in practice
I help clinical AI teams build evidence that survives contact with a real clinic: which workflow to test, what to measure beyond model accuracy, where clinicians will override you, and when to stop. I do the same work for health systems standing up AI governance.
The problem I’m interested in
Helping a physician get through the day faster is useful. But it leaves the same physician responsible for an ever-growing number of patients. Staffing shortages, long waits, and rising costs call for a closer look at how care is organized.
I’m interested in where AI can take on defined clinical tasks, with evidence that it works and a clear route back to a person when it doesn’t. That would give clinicians more room for the decisions and relationships that need them.
Why most clinical AI is tested on the wrong questionAI care model design
I help teams decide which tasks AI can handle, which decisions need a clinician, and when the system should ask for help.
This can include mapping a patient’s care journey, assigning responsibility at each step, and setting clear rules for when to hand back to a person.
Testing and validation
A model can perform well on a test and still fail in practice. I help teams design pilots that look at what happens to patients, clinicians, and the cost of care.
That starts with choosing a workflow worth testing, given your patients, staff, and technology. Then: what to measure, what gets missed, when clinicians override the system, who benefits and who is left out, and when to pause or stop.
Adoption and trust
Will clinicians trust the system? Will patients follow its advice? I bring behavioral science to the practical problem of knowing when to rely on AI and when to question it.
This means looking at resistance, over-reliance, and the small decisions in a workflow or interface that change how people behave.
Who acts. Who reviews. Who is responsible.
AI gathers and summarizes information. Clinicians decide.
AI suggests an action. A clinician approves it first.
AI handles routine tasks. A clinician checks afterward.
AI runs a defined part of care. People audit it and handle the exceptions.
AI manages a tested workflow on its own, with clear rules for handing back to a person and someone accountable for the result.
Each step to the right needs evidence, safeguards, monitoring, and someone clearly accountable.
“The question isn’t whether AI can replace doctors. It’s which clinical decisions still need a doctor, and what evidence we should demand before responsibility can safely move to AI.”
An illustrative workflow
A connected cuff can send a reading. The harder work is deciding what happens next: what the patient hears, what the team sees, and who follows through. Explore one possible model below.
The readings and responses are fictional. This is a design example, not a working service or medical advice, and nothing you select here is collected.
Routine check-in: Continue the agreed plan
01 / The next action
In this fictional case, the readings meet the care team’s agreed monitoring rules. The system records them and adds them to the next summary.
02 / Back to the clinical team
Trend summary with both source readings, measurement times, reported symptoms, and any missing information. No extra alert for this check-in.
03 / Responsibility
The assigned team reviews summaries on the schedule agreed before enrollment.
Agree the care plan before enrolling anyone: consent, eligibility, a validated cuff, measurement training, a named clinical team, and a supported alternative for people who can’t use the app. Write down what the AI may do (record readings, send approved reminders, summarize trends, route cases by clinician-approved rules) and what always goes back to a person (diagnosis, medication changes, conflicting information, anything outside the agreed protocol).
Store the source reading, its time and quality flags, the rule or model version, the message sent, who acknowledged it, and what happened next. Silence and stale data are treated as problems, not as good news.
Start with synthetic failure cases. Run in shadow mode without changing care. Then run a small, supervised pilot with success measures and stop rules agreed in advance. Compare with current care on missed and unnecessary escalations, time to action, patient effort, clinician overrides, results across patient groups, and the full cost of the pathway, not just time saved.
Measurement guidance: American Heart Association, monitoring your blood pressure at home.
The Future of Care
I’m developing a series of conversations with people who see healthcare from different angles: clinicians, technologists, employers, payers, and colleagues in behavioral science, policy, cybersecurity, and finance. I want to put difficult questions on the table before the answers are treated as settled.
Join the ConversationOn the agenda
Which of our clinical workflows could AI manage safely, and which need a human in the loop?
How do we test AI-enabled clinical workflows?
Can AI make healthcare cheaper and care more reliable?
Conversations coming soon.About Angela
Angela Thyer, MD, MSc
Physician · 20+ years of clinical experience
After more than two decades practicing medicine, I became increasingly interested in a problem medicine alone couldn’t solve: knowing what patients should do is very different from helping people actually do it.
That led me to behavioral science. AI now raises a second question: if machines can increasingly perform parts of clinical reasoning and care delivery, how should healthcare be redesigned around that capability?
These are the questions I now work on with organizations planning how to use AI in care.
Let’s Talk
Thirty minutes. Most useful if you’re designing an AI-enabled clinical workflow, planning a pilot, or working out what evidence your buyers will ask for before they say yes.