AI in radiology often looks impressive in a demo. It scores well in validation studies. It clears regulatory review. Then it lands in a clinic with different protocols, different workflows, and different staffing patterns. Suddenly the AI does not perform as expected. That drift was the spark behind HOPPR.
In a recent conversation with Dr. Khan Siddiqui, Founder and CEO of HOPPR, we explored what happens when AI tools are built to adapt to specific radiology teams instead of the market at large.
What This Conversation Revealed
- AI performance is local, not universal. Most organizations overlook this.
- Locked-down, one-size-fits-all models struggle where imaging technique and clinical behavior vary.
- The strongest AI use cases are operational and financial, not diagnostic.
“Does It Work Here?” Is the Only Question That Matters
Health IT teams are used to vendor claims that an AI model “works everywhere.” The assumption is that performance travels with the algorithm. Dr. Siddiqui challenged that assumption directly: “Instead of asking whether a model works in general for healthcare, we should ask, does it work here?”
A model tuned on one population, one acquisition style, and one workflow may degrade in another. Scanner configurations differ. Protocols differ. AI models will demonstrate different results at different organizations.
Frozen Models Don’t Travel Well
Regulatory clearance creates confidence. It does not eliminate environmental variability.
Dr. Siddiqui put it plainly: “The models that are trained and frozen don’t perform the same when they go from site to site. If I was trained at Hopkins, I would learn to scan patients this way. If I was trained at Stanford, I would learn to scan patients a different way.”
Differences in image acquisition cascade into differences in pixel distribution and ultimately model behavior. A static AI model assumes healthcare practices are uniform. That is simply not the case.
That reality is pushing some organizations to look for AI that can be tuned to their operating context.
HOPPR and the “Market of One”
Rather than shipping another fixed-model AI application, HOPPR built what Dr. Siddiqui calls an AI Foundry. The premise is straightforward: give organizations the components to fine-tune models against their own data, protocols, and risk profiles.
In practice, that means a radiology team can address a problem that only it has. A true “market of one.”
Dr. Siddiqui shared one example. A radiology practice faced a shortage of specialists and outsourced certain studies overnight. The cost of that teleradiology service exceeded reimbursement for many of those exams.
Together with HOPPR, they built a narrowly focused model to identify only the critical findings that required immediate reads at night. The result was a significant reduction in operating costs.
It may not be a broadly marketable feature. It does not need to be. It solved their problem.
The question is no longer “Does it work?”
It is “Does it work HERE?”
Learn more about HOPPR at https://www.hoppr.ai/
Listen and subscribe to the Healthcare IT Today Interviews Podcast to hear all the latest insights from experts in healthcare IT.
And for an exclusive look at our top stories, subscribe to our newsletter and YouTube.
Tell us what you think. Contact us here or on Twitter at @hcitoday. And if you’re interested in advertising with us, check out our various advertising packages and request our Media Kit.
No comments:
Post a Comment