STAT AI: An Emergency Medicine Podcast · Ethan Abbott and DJ Apakama

EP14 Fran Riley, MD, MSE Stanford Emergency Medicine

·48 min·3 clips

DeepSeek and other LLMs can produce plausible hallucinations that are factually wrong.

Stat AI episode 14 puts the hosts in conversation with a guest about where AI might actually help in emergency medicine. The discussion starts from the idea that the field is full of promising models, but only a few will prove useful on shift. The guest describes work on an image-recognition algorithm focused on regional wall motion abnormality. The goal is not novelty for its own sake. It is a tool that could affect disposition and help identify patients who may benefit from cardiology involvement. The episode frames that goal carefully. The hosts and guest keep returning to the difference between a flashy prediction and a result that changes care. The guest explains that the target is especially relevant for higher-grade NSTEMIs rather than only the obvious STEMI pathway. The conversation acknowledges that this is not something emergency medicine does alone. Cardiology partnership is part of the work from the beginning. A major theme is technical difficulty. Ultrasound images are noisy. Image quality depends on the person scanning and on the patient and clinical situation. That variability makes medical imaging a harder target than casual consumer image filters or face-tracking systems, even if the underlying computer vision ideas feel familiar. The show uses that comparison to make the challenge easier to grasp. It is a useful reminder that the gap between a clever model and a dependable clinical tool is large. The hosts then press on adoption. Even if the model validates, it still has to overcome the habits of clinicians who are used to eyeballing echo and relying on bedside intuition. That is where the episode gets most useful. It does not treat skepticism as resistance to progress. It treats it as part of implementation. The guest discusses the next step as moving from a visualization to something more quantitative and precise. The emphasis is on showing more than a hopeful result. The point is to make the output trustworthy enough to matter. The episode keeps that practical lens throughout. It is interested in workflow, not applause. It is interested in whether the tool can survive the noise and uncertainty of actual emergency care. The tone stays measured, curious, and grounded in real clinical use. The result is a conversation about AI that feels closest to the bedside question the hosts care about most: what would this change on a real shift, for a real patient, in a real department?

As heard by us

A grounded AI-in-emergency-medicine piece on ultrasound noise, validation, and bedside trust.

Stat AI asks a practical question: can machine learning help emergency medicine in a way that actually matters at the bedside? The discussion stays with one image-recognition effort for regional wall motion abnormality, and it keeps circling the same hard constraint.

Read the full review in PlayNext →

Why you'd press play

You want AI that could actually change an ED disposition.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from STAT AI: An Emergency Medicine Podcast