Dr. Vivek Singh talks with Artem Trotsyuk, a Stanford-trained engineer who built a wireless, AI-powered smart bandage that could monitor wounds and trigger healing without human intervention. The work was published in Nature Biotechnology and covered by the NIH. Then Trotsyuk tried to move it from the lab to an actual patient, and he learned that deployment is a completely different problem than the one he had just solved. That experience redirected his career: instead of building the next tool, he started studying why the last one never made it into practice. He now works with a governance group at Stanford Health Care that evaluates AI tools before they move from pilot to enterprise use. Here are three takeaways from the conversation.
Benchmark performance and sensitivity get most of the attention, but Trotsyuk argues they are rarely the reason a new tool never reaches patients. The harder problems are workflow friction, integration failures, and unintended consequences that no one catches in the lab. A tool can perform well in testing and still stall out entirely once it meets a hospital's infrastructure, staffing, and existing processes.
When Trotsyuk's team evaluates a new tool, they look at financial impact, ethics, and whether the tool does what its developers claim. Patients, clinicians, and developers rarely define success the same way. A patient may care most about consent. A developer may care most about data quality. A clinician wants to know whether the tool adds to their workload or genuinely lightens it and helps their patients. Reconciling those views, not just proving accuracy, is what determines whether a tool actually gets adopted.
Imaging volume keeps climbing, and someone still has to review it, so the idea that AI will eliminate radiologists or other clinical roles does not hold up. The bigger concern is over-reliance, where clinicians trust an output enough that they stop examining the source images or questioning what the tool might have missed. Training programs are starting to guard against this by holding AI tools back from residents until they have built the underlying skill to catch an error themselves, since trust in a tool has to be earned and continually verified, not assumed.
For Trotsyuk, the throughline is that healthcare AI does not succeed in a lab or a benchmark. It succeeds only when it survives contact with the people, workflows, and incentives inside a real hospital, and when the humans using it stay engaged enough to catch what it gets wrong.
Most healthcare AI does not fail because of poor accuracy. It fails during deployment, when workflow friction, integration problems, and unintended consequences surface that no one caught during lab testing. A tool can perform well in a benchmark and still stall out once it meets a hospital's actual infrastructure and staffing.
Many hospitals evaluate AI tools across several dimensions, including financial impact, ethics, and whether the tool performs as developers claim. Patients, clinicians, and developers often define success differently, so evaluation teams work to reconcile those competing priorities before a tool moves from pilot to enterprise use.
No. AI helps enable and support radiologists and clinicians but should never be used to replace the medical expertise of human medical professionals.
Want to watch the full conversation? Season 2, Episode 4 of Radiology Rewired with Artem Trotsyuk, PhD, available now on YouTube, Apple, or Spotify