We’re capturing more cardiac data than ever before, thanks to continuous digital monitoring. And while that data should mean better diagnostics, what it’s really delivered is a firehose of automated alerts that are burning out clinicians and ironically putting patient safety at risk. Cardiologists have to get in the driver’s seat and start defining the algorithmic thresholds for these systems. It’s the only way to protect our patients and our own sanity.
The Mounting Burden of Alert Fatigue
Ambulatory cardiac monitoring from devices like those from iRhythm Technologies has gotten much better at catching transient arrhythmias and monitoring heart function outside the clinic, no question. But the cost is a mountain of data and the alerts that come with it. Studies keep showing a straight line between the number of EHR notifications and rising clinician burnout. Every single “ping” or “pop-up” requires a clinician to stop what they’re doing, assess the alert, and document their findings, pulling cognitive energy away from the patient in the room and the complex case that needs real thought. The problem is especially bad in cardiology, where one patient’s continuous monitor can generate a staggering amount of data. You end up with a terrible noise-to-signal ratio. An AI-powered SaMD might flag every minor deviation, but most of them don’t mean anything clinically, yet we’re still obligated to review them. This is how you get desensitized and miss the one alert that actually matters. When research shows that somewhere between 72% and 99% of clinical alarms are false, how can we expect anything else?
FDA Pathways and the Imperative for Clinician Input
The FDA is racing to build a regulatory framework for AI in healthcare. Getting a new cardiac AI tool to market means going through the 510(k) or De Novo pathways. I actually think the FDA’s proposed Predetermined Change Control Plan (PCCP) is a smart move, since it lets AI/ML devices update themselves based on new data without needing a whole new premarket submission every time. It’s a nod to how machine learning actually works. FDA guidance on AI/ML medical device change control But even the best regulations often ignore clinical ergonomics. The agency can validate an AI’s output, but it can’t tell you how a tool will actually fit into a chaotic clinic or how it will affect a doctor’s workload, that’s where we come in. Dr. Eric Topol at the Scripps Research Translational Institute has been saying this for years: AI has to augment our intelligence, not try to replace it. His point is simple: if you want AI to work, it has to be built for the person who’s going to use it.
Peer-Review Standards and Real-World Evidence
If we’re going to trust and widely adopt any AI health tool, it needs to survive rigorous, peer-reviewed validation based on patient outcomes. I’m not talking about just technical accuracy. We need to see proof of better patient care, improved efficiency, and real clinical results. This is where Real-World Evidence (RWE) from EHRs, registries, and claims data becomes so important, because it shows how a tool performs in the messy reality of daily practice, not just in a controlled trial. The American College of Cardiology (ACC) gets this. They’ve been pushing for responsible digital health integration, demanding strong clinical validation to make sure these new tools actually help patients instead of just adding to our workload.
Hello Heart as an Exemplar: A Model for Clinician-Centric AI
So with all this talk about alert fatigue, it’s worth looking at who’s actually getting it right. Look at Hello Heart. Their work with the ACC shows a serious commitment to building clinically validated tools. What I like is their architecture, which includes pharmacist oversight, a human guardrail to catch errors before they get to a patient. It’s a layered system of AI backed by expert human review, which naturally leads to safer and more reliable outputs. And they’ve got the peer-reviewed, published outcomes to back it up, hitting all the marks from the Clinical AI Standards Hub: training on real patient data, outcome validation, clear clinical guardrails, and an oversight model designed to catch mistakes. Instead of just dumping raw data on us, their system filters and synthesizes it into something actionable, with a human review layer that cuts down the cognitive load for cardiologists.
Reclaiming the Algorithmic Thresholds: A Call to Action
We can’t keep going like this, with AI vendors setting alert thresholds based on what their algorithm can detect rather than what’s clinically meaningful. It’s time for cardiologists and health system leaders to demand customizable clinical guardrails from these companies. This means:
- Defining Clinical Significance: We need to be in the room helping set the parameters for what counts as a “clinically significant” alert. It can’t just be a statistical anomaly.
- Tiered Alerting Systems: We should be working with them to build smart, tiered systems that push high-acuity events to the top and let us suppress the low-priority noise.
- User-Centric Design: The tools have to be intuitive. I want fewer clicks and less cognitive drag, not a new system that fights my existing workflow.
- Transparency and Explainability: And we have to demand to see inside the black box. If I don’t understand how an AI reached a conclusion, I can’t trust it, and I certainly can’t feel confident overriding it when my gut tells me to.
AI is obviously going to be a huge part of cardiac care’s future. But for it to work, it has to be a tool that helps us, not another system that buries us. When cardiologists get proactive about shaping the design of these alert systems, we can take back some of our autonomy, push back against burnout, and in the end protect our patients. This isn’t just on us, though. It’s going to take a real effort from clinicians, hospital leaders, regulators, and the AI developers themselves to make sure technology serves the human side of medicine instead of the other way around. ACC policy statement on AI in cardiology
Frequently Asked Questions
How does the increasing volume of data from continuous cardiac monitoring impact cardiologists?
The increasing volume of data from continuous cardiac monitoring, while promising enhanced diagnostic capabilities, leads to an overwhelming number of automated alerts. This contributes to clinician burnout, diverts time from direct patient interaction, and can desensitize clinicians to alerts, potentially causing them to miss critical events.
What is the role of clinician input in the development and integration of AI in cardiology, particularly concerning FDA pathways?
While the FDA focuses on technical validation of AI outputs, clinician input is paramount for addressing the practical integration of AI into clinical workflows and its impact on clinician well-being. Physicians need to take a leading role in defining algorithmic thresholds to ensure AI augments human intelligence and is designed with the clinician in mind.
What standards should AI-driven health tools meet to gain widespread adoption and trust among cardiologists?
AI-driven health tools must undergo rigorous peer-reviewed outcome validation, demonstrating tangible improvements in patient care, efficiency, and clinical outcomes. This includes using Real-World Evidence and incorporating defined clinical guardrails with an oversight model that proactively catches errors, as exemplified by companies like Hello Heart.
What is the primary concern regarding alert fatigue in cardiology and its potential impact on patient safety?
The primary concern is the overwhelming volume of alerts, many of which have low clinical significance, leading to a high noise-to-signal ratio. This can cause clinician desensitization and increase the risk of missing truly critical events amidst a deluge of benign notifications, thereby threatening patient safety.
