#317 - What It Takes to Build Medical AI Doctors Will Trust | Dr. Suchi Saria & Mike Haney
LEVELS – A Whole New LevelShow Notes
Medical AI is full of impressive models that work on retrospective datasets. The much harder problem is turning one into a tool that can operate continuously inside a hospital, earn clinicians’ trust, fit into an overloaded workflow, stay reliable across different health systems, and ultimately improve patient outcomes. Mike Haney talks with Johns Hopkins researcher and Bayesian Health founder Dr. Suchi Saria about the decade-long journey from an academic sepsis-prediction algorithm to AI that clinicians are using in hospitals today. Her experience is a reality check on what it actually takes to make medical AI useful, and an argument for aiming higher than documentation and billing: toward systems that continuously watch patients, catch deterioration earlier, and help prevent avoidable harm.
Free course: Improve your metabolic health Get our free email course on how glucose, nutrition, exercise, sleep, and measurement can help you build habits that support better energy and long-term health: https://levels.link/wnl
What We Cover: Why sepsis fits continuous AI monitoring How causal reasoning separates signal from a noisy alert Why real-time access to the full record matters When a working algorithm still failed to get attention How adoption and human-machine teaming became engineering problems What happened when the system went live How the same infrastructure extends beyond sepsis Why healthcare is too conservative about medical AI
🎙️ About the Guest: Dr. Suchi Saria is the John C. Malone Associate Professor at Johns Hopkins University, where she directs the Machine Learning and Healthcare Lab, and is the founder and CEO of Bayesian Health. Her work focuses on building and deploying machine-learning systems that use real-world clinical data to improve diagnosis, treatment, and patient safety.
📍What Dr. Suchi Saria & Mike Haney discussed: 04:30 - Why the real AI risk is what it is allowed to do 08:32 - Why sepsis is a race against organ failure 11:57 - From checklists to reading the full medical record 16:30 - How causal reasoning finds a patient’s baseline 18:50 - Using uncertainty to avoid another noisy alert 22:11 - Why an hour-late warning is already too late 25:06 - Reading Epic in real time without slowing it down 32:35 - A working algorithm still couldn’t get clinicians’ attention 35:53 - What happened when the system went live 45:32 - Human-machine teaming, and who still makes the call 54:17 - The same infrastructure, beyond sepsis 58:35 - Why billing and coding is too small a target
🔗 Helpful Links: Suchi Saria — Johns Hopkins Research, publications, and academic background. Bayesian Health: Saria’s company commercializing real-time clinical AI. TREWScore — Science Translational Medicine (2015): The original real-time early warning model for septic shock. TREWS patient outcomes — Nature Medicine (2022): Prospective multisite evaluation after deploying the sepsis warning system in clinical care. TREWS clinician adoption — Nature Medicine (2022): Study of provider adoption and effects on sepsis treatment timing. Human-machine teaming — npj Digital Medicine (2022): Clinician interviews on what made the deployed AI system trustworthy and usable. Bayesian Health Sepsis Flagging Device — FDA (2026): FDA 510(k) clearance for the sepsis prediction software discussed in the episode.
Watch the conversation: https://youtu.be/uVEKgSvei4s Find us on YouTube: https://youtube.com/levelshealth?sub_confirmation=1
👋 Who we are: Levels helps you understand your metabolic health with personalized data, expert guidance, and tools that connect your daily choices to measurable changes in your body. Our goal is to help you make better decisions about food, exercise, sleep, and long-term health.
Look for new shows every month on A Whole New Level, where we have in-depth conversations with thought leaders about metabolic health.
More from LEVELS – A Whole New Level
Heart disease is the leading cause of death, but it’s also one of the most preventable. Decades of research have established a strong causal role for LDL and other ApoB-containing particles in atherosclerosis, and we now have multiple proven ways to lower them. The harder question is how to turn that evidence into a decision for an individual person. Mike Haney talks with Johns Hopkins cardiologist Dr. Seth Martin, a co-author of the new 2026 dyslipidemia guidelines, about what the latest evidence says we should actually measure and act on. The throughline is pragmatic: lower exposure to atherogenic particles earlier and for longer, use additional tests when they can change a decision, and distinguish interesting new measurements from ones that have actually been shown to improve outcomes.
Most of NextLevel starts with a measurement problem: we can see more biology than ever, but we still have to figure out what to do with it. Bioelectronic medicine flips that sequence. We already know how to send electrical signals into the nervous system to change biology, including an FDA-approved vagus nerve treatment for rheumatoid arthritis, even though we still can’t map or read most of the individual nerve fibers carrying those signals.
Most health decisions happen before the science is settled. A study points one way, another points somewhere else, experts disagree about what matters, and even strong population-level evidence may not tell you what will happen to you.
Most medical imaging happens after a symptom, injury, or abnormal test gives a clinician a reason to look. But what changes if imaging becomes a more routine measurement, creating a record of the body before something goes wrong? In this episode of Next Level, Mike Haney talks with Dr. Daniel Durand, chief medical officer at Prenuvo, about whole-body MRI as a test case for a broader shift from episodic medicine toward longitudinal health data. Whole-body MRI is controversial, and Durand is clearly an advocate. The conversation takes the skepticism seriously, then asks a larger question: as measurement gets cheaper, more frequent, and increasingly visible to patients themselves, how should medicine decide what is worth knowing?
Alcohol has spent the last 40 years bouncing between two health narratives: a little is good for your heart, or any amount is dangerous. The science doesn’t really support either slogan. Dr. Tim Naimi has spent decades studying the health effects of alcohol and, just as importantly, why the research is so difficult to interpret. His read of the evidence is less dramatic but more useful: there’s no good reason to start drinking for your health, very low intake may carry little measurable added risk, and risk rises as consumption does. Understanding why earlier studies made moderate drinking look protective helps explain a lot of the whiplashing headlines.