When I think about where artificial intelligence is headed in healthcare, I keep coming back to a simple tension I live with every day: the same technology that can save a patient’s life can also harm one if it’s deployed carelessly. As a Senior R&D Manager leading a software portfolio focused on patient safety, I’ve spent years building systems where that tension isn’t abstract—it’s the thing that determines whether a product ships or stays on the bench. AI raises the stakes on both sides of that equation, and learning to navigate it well is, in my view, one of the defining challenges of the next decade in medical devices.
The Promise Is Real—And So Is the Burden of Proof
It’s easy to get swept up in the capabilities. AI models can now detect patterns in surgical and clinical data that no human reviewer would catch in real time. In my own work on surgical safety platforms, the value of pulling signal out of noise—tracking what’s happening in an operating room, flagging risk before it becomes an event—is exactly the kind of problem machine learning is suited to.
But a demo is not a device. In consumer software, you ship, measure, and iterate. In medical devices, a false negative can mean a retained surgical item or a missed hemorrhage. That asymmetry changes everything about how you build. Every model has to be explainable enough to defend, validated against the messy reality of clinical environments, and monitored long after it leaves your hands. The hardest engineering question isn’t “can the model do this?” It’s “can we prove it does this safely, repeatedly, for the patients who will actually depend on it?”
Regulation Is Not the Enemy of Innovation
There’s a popular narrative that regulation slows progress. I’ve come to see it differently. The FDA’s evolving frameworks for AI and machine learning—particularly the move toward predetermined change control plans for adaptive algorithms—are an attempt to solve a genuinely novel problem: how do you regulate a product that’s designed to change after approval?
Traditional device regulation assumes a fixed product. AI breaks that assumption. A model that learns and updates is a moving target, and the regulatory community is working through how to allow beneficial improvement without reopening the door to uncontrolled risk. Teams that treat this as a checkbox exercise will struggle. Teams that engage with it early—building traceability, risk management, and cybersecurity controls into the development lifecycle rather than bolting them on at the end—will move faster, not slower, because they won’t be rebuilding under audit pressure.
Cybersecurity deserves special mention here. A connected, data-hungry AI system is also an attack surface. Frameworks like RMF exist because a compromised medical device isn’t just a data breach—it’s a patient safety event. The discipline of threat modeling belongs in the same conversation as clinical validation, not in a separate silo owned by a different team.
Data Is the Foundation, and It’s Usually the Problem
Most failed AI initiatives in this space don’t fail because of the algorithm. They fail because of the data. Clinical data is fragmented, inconsistently labeled, and full of edge cases that matter enormously. A model trained on clean, representative data from one hospital may quietly underperform in another with different equipment, workflows, or patient populations.
Getting this right requires investment that doesn’t always show up in a roadmap: data governance, careful curation, and honesty about where the model’s confidence is justified and where it isn’t. The teams I trust most are the ones willing to say “we don’t have enough data to claim that yet.” That restraint is a feature, not a weakness.
Leadership at the Intersection
What I’ve learned managing teams across QA and cloud engineering is that the technical and the regulatory can’t be owned separately. The most expensive mistakes happen at the seams—when engineering optimizes for capability, regulatory optimizes for documentation, and no one owns the integrated risk picture.
The leader’s job at this intersection is to hold both truths at once: push hard on what AI can unlock for patients, while never letting the organization confuse “it works in testing” with “it’s safe to deploy.” That means building cultures where engineers feel responsible for clinical outcomes, not just feature delivery, and where raising a safety concern is rewarded rather than treated as friction.
Where This Goes Next
The medical device companies that win the AI era won’t be the ones with the flashiest models. They’ll be the ones that earn trust—from regulators, from clinicians, and from patients—by being rigorous when it would be easier to be fast. AI is a powerful tool, but in this field, credibility is the actual product. Everything else follows from it.
That’s the work I’ve committed my career to, and it’s the lens I’d bring to any board navigating these decisions: an insistence that innovation and safety aren’t competing priorities, but two halves of the same responsibility.
