Cardiac AI: Who Owns Post-Clearance Evidence?

Putting AI into cardiac care is a double-edged sword. The promise is incredible, but the responsibility we take on is just as big. Unlike a simple wellness app, cardiac AI often kicks off a long monitoring period, weeks or months. An initial, even subtle, finding from the algorithm can set off a whole chain reaction of downstream tests, procedures, and interventions. Because it has such a long-lasting influence, we have to be absolutely rigorous about assessing its performance continuously, not just at the moment it gets cleared by regulators, but for the entire time it’s in use.

Post-Market Surveillance

Getting a device cleared through a pathway like the FDA 510(k) just means it’s considered safe and effective for its intended use, or at least equivalent to something already on the market. For an AI device, that clearance is the starting pistol, not the finish line. Once these tools hit real-world clinics, an obligation called Post-Market Surveillance kicks in. The FDA’s Center for Devices and Radiological Health (CDRH) oversees this whole process, which is there to watch how the device actually performs, spot risks nobody saw coming, and make sure it stays safe. AI’s tendency for ‘algorithmic drift’ is precisely why this is so important. A model trained on one set of data will inevitably run into new patient types, different comorbidities, and weird data inputs out in the wild that it never saw during training. Its performance can start to degrade over time, leading to anything from bad drug interaction advice and missed diagnoses to undertriaging a heart attack or delaying stroke identification. If you’re not constantly watching, these performance dips will fly right under the radar, slowly making the device less useful and even dangerous. The whole game shifts from that one-time initial validation to constantly checking that the model is still behaving itself across the messy, real-world patient groups we actually treat.

Patient Harm Isn’t Just Device Failure

Poor AI performance causes problems that go way beyond the device simply not working. Harlan Krumholz’s work really drives this home: we need to measure harm from the diagnostic cascades AI can trigger, not just from the device failing. Think about it. A single false positive from a long-term cardiac monitor can kick off a whole series of invasive, expensive, and potentially harmful follow-ups, unnecessary angiograms, EP studies, you name it. That means huge patient anxiety, financial pain, and the risk of complications from procedures that were never needed in the first place. On the flip side, a false negative can mean a critical intervention gets delayed, with obvious and severe consequences. An AI ‘failure’ isn’t some spectacular system crash. It’s usually much quieter, a pattern of subtle inaccuracies that, once you scale it across thousands of patients in our already-complex workflows, adds up to a mountain of cumulative harm. As clinical informaticists and cardiac monitoring leads, we have to get our heads around this bigger picture of harm. The proof that an AI is safe has to include what happens downstream. This means we have to actively hunt for and shut down these diagnostic cascades, making sure the AI is actually helping, not just making patient care more complicated.

The Never-Ending Evidence Duty: A Look at the Cardiac Players

Just look at the big players in cardiac AI: iRhythm Technologies, HeartFlow, and AliveCor. They are living proof of this constant need to provide evidence long after getting regulatory clearance. iRhythm is a good example, working with long-duration monitoring where their Zio XT and Zio AT devices generate a firehose of data over weeks. The usefulness and safety of these devices aren’t fixed on day one. They need constant checking to make sure their diagnostic accuracy is still solid across all the different kinds of patients we see and that their reports aren’t sending people for needless procedures or, worse, missing a major cardiac event. iRhythm Technologies regulatory filings and clinical studies Then there’s HeartFlow. Its FFRCT Analysis gives a non-invasive look at coronary artery disease from a CT scan, but that clearance comes with a very long tail of required evidence. Doctors use its fractional flow reserve calculations to decide on revascularization, so the accuracy of those numbers has to be constantly checked in the real world to keep us from over-treating some patients and under-treating others. AliveCor and its KardiaMobile personal ECGs are in the same boat. While they offer incredibly convenient monitoring, the AI doing the interpretation needs nonstop validation to make sure it’s catching arrhythmias without spreading AI health misinformation or scaring patients with false positives. AliveCor clinical validation studies Every one of these companies, even with their 510(k) clearance in hand, has the same post-market job to do. The FDA’s 510(k) pathway determines who gets to sell a device, but the real question is who has to keep proving the model still works? This goes way beyond hitting some technical spec in a lab. It’s about whether the tool remains clinically useful and safe when used by different people on different patients. The job of gathering evidence never ends. It demands constant watchfulness, generating real-world data, and maybe even adapting the algorithms under frameworks like the FDA’s Predetermined Change Control Plan (PCCP) if the device is built to learn on the fly. FDA guidance on AI/ML-based SaMD Action Plan

So, Who Owns the Evidence After Clearance?

So who’s on the hook for this evidence after an algorithm is cleared? For those of us in clinical informatics or running cardiac monitoring programs, this is the most important question. Getting clearance is a snapshot in time, but as the big cardiac AI companies show, the duty to provide evidence never stops. The manufacturer can’t be expected to do it all, and the hospitals using the tech can’t just wash their hands of it. It has to be a shared responsibility. Any health system using cardiac AI needs its own solid, internal processes for watching device performance, tracking patient outcomes, and reporting adverse events when they happen. This means having clear protocols for collecting and analyzing data, and then feeding it back to the manufacturers and regulators. Clinical informaticists are in the perfect position to lead this, baking AI performance monitoring right into existing QI programs and the EHR. At the same time, cardiac monitoring program leads have to make sure their teams are trained on how to use the tools properly and (just as important) how to recognize when an algorithm might be drifting or giving weird results. Safe AI in cardiac care boils down to a continuous loop of generating evidence, evaluating it, and adapting. That initial clearance is just a license to get started. It’s the sustained clinical validation and serious post-market surveillance that actually protect patients. This shared commitment to evidence is the only way we’ll build real trust in these AI tools.